physical layer security for multi-user ... - wordpress.coma special thanks also goes to my...
TRANSCRIPT
Physical Layer Security forMulti-User MIMO Systems
Giovanni GeraciB.Sc. M.Sc.
Supervisor: Prof. Jinhong Yuan
A DISSERTATION SUBMITTED FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY
March 2014
School of Electrical Engineering and Telecommunications
The University of New South Wales, Australia
When I was seventeen, I was working as an apprentice mechanic in myvillage, in Sicily. On a summer’s day, Mr L.G. walked in the workshopto get his car battery changed. While I was replacing the battery, heasked me how long I was going to work there. I said “until September,when school reopens”. Mr L.G. replied that I would have been betteroff quitting school, because I was wasting my time.
Mr L.G., this dissertation is dedicated to you.
iii
AcknowledgmentsBoarding a plane to Australia three years ago to start a Ph.D. was a choice I did not regret. That
flight marked the beginning of a journey whose destination is this thesis. For me, it has been
all about the journey, not the destination. Difficulties and challenges have made me stronger and
mature. I have learnt how to bounce back from failures. Travelling and coming across different
cultures has made me more open minded, flexible and creative. Since that long flight, I feel I have
grown as a researcher, but most importantly as a man. Therefore, even if one day I will become
Dr. Giovanni, I will always be Giovanni first, Doctor second.
Many people have helped me complete this Ph.D. successfully. None of them need to be named,
for if they know me, then they already know how grateful I am to them all, even if all they did was
just smile at me once.
Most of the things I did in the last three years would not have been possible without my supervisor.
He is a great person, and he always treated me as an adult, not just as a student. He probably
thinks I am smarter than I really am. A special thanks also goes to my co-supervisor at CSIRO
for providing many insightful suggestions and remarks throughout my Ph.D. Thanks to all the
sources of funding that have supported my studies, and allowed me to present my work around the
globe. Thanks to my advisors in France and United States, for offering me internships that greatly
contributed to my professional and personal growth. Thanks to all my research collaborators, not
just for their help in making my publications list longer, but more importantly for all I have learnt
from them. I also thank all the colleagues and staff in Sydney, Paris, and Austin. You know I
enjoy much more interacting with humans than with computers. Thanks to all the friends I am
lucky to have in every single continent of this amazing world. Thanks to my family for giving me
opportunities that not everyone was privileged to have.
v
Abstract
T his thesis studies physical layer security for multi-user multiple-input multiple-output (MIMO)
wireless systems. Due to the broadcast nature of the physical medium, wireless multi-user
communications are very susceptible to eavesdropping, and it is critical to protect the transmit-
ted information. Security of wireless communications has traditionally been achieved at the net-
work layer with cryptographic schemes. However, classical cryptography might not be suitable
in large dynamic networks, since it requires key distribution and management, and complex en-
cryption/decryption algorithms. A method that exploits the characteristics of wireless channels,
known as physical layer security, was proposed as an alternative to achieve perfect secrecy without
requiring encryption keys. The rates at which messages can be reliably transmitted to an intended
user while no information is leaked at the eavesdroppers, denoted as the secrecy rates, have been
studied for several network topologies. However, the secrecy rates achievable in generic multi-
user networks are still unknown. Hence in this thesis we study the secrecy rates achievable in
multi-user systems with practical transmission schemes.
We propose a linear precoder based on regularized channel inversion (RCI) for the broadcast
channel with confidential messages (BCC). In the BCC, a multi-antenna base station (BS) simul-
taneously transmits independent confidential messages to several spatially dispersed malicious
users that can eavesdrop on each other. We carry out a large-system analysis and obtain closed
form expressions for the achievable secrecy rates under Rayleigh fading, as well as the optimal
regularization parameter and the optimal network load. Simulations confirm that our analysis is
accurate even for finite systems. We compare the secrecy rate of the proposed precoder to two
upper bounds obtained without secrecy requirements and without interference, respectively, and
show that it has the same scaling factor as the two bounds in the high signal-to-noise ratio (SNR)
regime. We further extend our analysis to more practical scenarios where only imperfect chan-
nel state information is available at the BS, and where channel correlation is present among the
transmit antenna elements.
We then introduce the broadcast channel with confidential messages and external eavesdroppers
(BCCE). Unlike the BCC, in the BCCE not just malicious users, but also randomly located ex-
ternal nodes can act as eavesdroppers. We obtain the probability of secrecy outage and the mean
secrecy rate for the RCI precoder in the BCCE, for the two cases of non-colluding and colluding
eavesdroppers. We show that, irrespective of the collusion strategy at the external eavesdroppers,
a large number of transmit antennas drives both the probability of secrecy outage and the rate loss
due to the presence of external eavesdroppers to zero. Increasing the density of eavesdroppers by
a factor n, requires n2 as many antennas to meet a given probability of secrecy outage and a given
vii
mean secrecy rate. Our analysis demonstrates that the number of transmit antennas at the BS is a
key resource to secure communications against malicious users and external eavesdropping nodes.
We finally turn our attention to cellular networks where, unlike the case of isolated cells, multiple
BSs generate inter-cell interference, and malicious users of neighboring cells can cooperate to
eavesdrop. We characterize the probability of secrecy outage and the mean secrecy rate with
RCI precoding, accounting for the spatial distribution of BSs and users and the fluctuations of
their channels. We find that RCI precoding can achieve a non-zero secrecy rate with probability
of outage smaller than one. However we also find that unlike isolated cells, the secrecy rate in
a cellular network does not grow monotonically with the SNR, and the network tends to be in
secrecy outage if the transmit power grows unbounded. We further show that there is an optimal
value for the density of BSs that maximizes the secrecy rate, and this value is a decreasing function
of the SNR. Using the developed analysis, we clearly establish the importance of designing the
transmit power and the BS deployment density to make communications robust against malicious
users in other cells.
x
Contents
Acknowledgments v
Abstract vii
List of Acronyms xxi
Basic Notations xxiii
1 Introduction 11.1 Background and Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Thesis Overview and Contributions . . . . . . . . . . . . . . . . . . . . . . . . . 41.3 Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2 Overview of Physical Layer Security 132.1 Introduction to Physical Layer Security . . . . . . . . . . . . . . . . . . . . . . 142.2 The Wiretap Channel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142.3 Fading Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162.4 Multi-antenna Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172.5 Multi-user Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3 Physical Layer Security in Isolated Cells: Achievable Rates 213.1 The Broadcast Channel with Confidential Messages . . . . . . . . . . . . . . . . 223.2 Achievable Secrecy Rates in the BCC . . . . . . . . . . . . . . . . . . . . . . . 23
3.2.1 Linear Precoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243.2.2 Achievable Secrecy Rates with Linear Precoding . . . . . . . . . . . . . 253.2.3 Regularized Channel Inversion Precoding . . . . . . . . . . . . . . . . . 273.2.4 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.3 Optimal Regularization Parameter and Power Allocation . . . . . . . . . . . . . 303.3.1 Regularized Channel Inversion Precoding with Power Allocation . . . . . 313.3.2 Optimal Power Allocation . . . . . . . . . . . . . . . . . . . . . . . . . 323.3.3 Joint Optimal Power Allocation and Regularization Parameter . . . . . . 34
3.4 Achievable Secrecy Rates in Practical Channels . . . . . . . . . . . . . . . . . . 363.4.1 Secrecy Rates in the Presence of Imperfect Channel State Information . . 36
xi
3.4.2 Secrecy Rates Under No CSI or Poisoned CSI . . . . . . . . . . . . . . . 393.4.3 Secrecy Rates under Transmit Channel Correlation . . . . . . . . . . . . 40
3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4 Physical Layer Security in Isolated Cells: A Large-System Analysis 434.1 Large-System Analysis of the Secrecy Rates . . . . . . . . . . . . . . . . . . . . 44
4.1.1 Large-System Secrecy Rates with RCI Precoding . . . . . . . . . . . . . 444.1.2 Secrecy Sum-Rate Maximizing Regularization Parameter . . . . . . . . . 454.1.3 Optimal Secrecy Sum-Rate . . . . . . . . . . . . . . . . . . . . . . . . . 474.1.4 Optimal Network Load . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.2 RCI Precoder with Power Reduction . . . . . . . . . . . . . . . . . . . . . . . . 524.2.1 Optimal Transmit SNR . . . . . . . . . . . . . . . . . . . . . . . . . . . 524.2.2 Power Reduction Strategy . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.3 Performance Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 554.3.1 Secrecy Loss . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 554.3.2 Multi-User Loss . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 574.3.3 Comparison to Other Linear Schemes . . . . . . . . . . . . . . . . . . . 58
4.4 Imperfect CSI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 604.4.1 Minimum Required CSI . . . . . . . . . . . . . . . . . . . . . . . . . . 604.4.2 Channel Feedback in FDD Systems . . . . . . . . . . . . . . . . . . . . 634.4.3 Channel Training in TDD Systems . . . . . . . . . . . . . . . . . . . . . 63
4.5 Channel Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 654.5.1 Large-System Secrecy Sum-Rates . . . . . . . . . . . . . . . . . . . . . 664.5.2 Selection of the Optimal Regularization Parameter . . . . . . . . . . . . 674.5.3 Comparison to Other Linear Schemes . . . . . . . . . . . . . . . . . . . 684.5.4 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5 Physical Layer Security in a Random Field of Eavesdroppers 735.1 Introduction to the BCCE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 745.2 Probability of Secrecy Outage . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.2.1 Non-colluding Eavesdroppers . . . . . . . . . . . . . . . . . . . . . . . 775.2.2 Colluding Eavesdroppers . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.3 Mean Secrecy Rates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 805.3.1 Mean Secrecy Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 815.3.2 Secrecy Rate Loss due to the External Eavesdroppers . . . . . . . . . . . 825.3.3 Optimal Regularization Parameter . . . . . . . . . . . . . . . . . . . . . 83
5.4 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 845.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
6 Physical Layer Security in Cellular Networks 916.1 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
6.1.1 Network Topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 926.1.2 RCI Precoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 936.1.3 Malicious Users . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.2 Achievable Secrecy Rates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 946.2.1 SINR at a Typical User . . . . . . . . . . . . . . . . . . . . . . . . . . . 956.2.2 SINR at the Malicious Users . . . . . . . . . . . . . . . . . . . . . . . . 956.2.3 Achievable Secrecy Rates . . . . . . . . . . . . . . . . . . . . . . . . . 97
6.3 Large-system Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 996.3.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 996.3.2 Characterization of Interference and Leakage . . . . . . . . . . . . . . . 1016.3.3 Probability of Secrecy Outage . . . . . . . . . . . . . . . . . . . . . . . 1036.3.4 Mean Secrecy Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
6.4 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1076.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
7 Conclusions and Future Work 113
Appendices 117
A Appendix for Chapter 4 117A.1 Proof of Theorem 4.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117A.2 Proof of Theorem 4.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
B Appendix for Chapter 5 121B.1 Proof of Lemma 5.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121B.2 Proof of Lemma 5.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122B.3 Proof of Lemma 5.6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123B.4 Proof of Lemma 5.9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124B.5 Proof of Corollary 5.11 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
C Appendix for Chapter 6 127C.1 Proof of Proposition 6.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127C.2 Proof of Lemma 6.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128C.3 Proof of Proposition 6.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129C.4 Proof of Theorem 6.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130C.5 Proof of Corollary 6.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
Bibliography 132
xiv
List of Figures
2.1 The wiretap channel in WYNER [1975], where the eavesdropper’s channel is adegraded version of the main channel. . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 The general MIMOME wiretap channel. . . . . . . . . . . . . . . . . . . . . . . 17
3.1 The MISO broadcast channel with confidential messages (BCC). . . . . . . . . . 233.2 Simulated secrecy sum-rate SBCC achievable by RCI precoding in the BCC versus
SNR ρ, for various values of N = K. . . . . . . . . . . . . . . . . . . . . . . . 303.3 Comparison between the RCI precoder and the plain CI precoder, for various val-
ues of N = K. The secrecy loss is also shown as the gap between dashed andsolid lines. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.4 Per-user secrecy rate vs. SNR ρ for K = 4 users: with separate (dashed) and joint(circle) optimization of ξ and p, and with equal power allocation (solid). The rateof the RCI-EP precoder without secrecy requirements (square) and the secrecycapacity of the MISOME channel (diamond) are also plotted. . . . . . . . . . . . 37
4.1 Comparison between the secrecy sum-rate with RCI precoding in the large-systemregime (4.8) and the simulated ergodic secrecy sum-rate for finite N . Three setsof curves are shown, each one corresponds to a different value of β. . . . . . . . 49
4.2 Mean normalized secrecy sum-rate difference between S?BCC (obtained using theoptimal ξ?BCC,N ) and SBCC(ξBCC) (obtained with ξBC from large-system analy-sis), for β = 0.8 and various values of the SNR. . . . . . . . . . . . . . . . . . . 50
4.3 Large-system secrecy sum-rate per transmit antenna as a function of β for RCIprecoding. The value of β? is indicated next to each curve. . . . . . . . . . . . . 51
4.4 Comparison betweenK? (obtained via simulations),K? (obtained via exhaustivesearch and large-system analysis), and the analytical approximation K? (obtainedvia large-system and large-SNR analysis). . . . . . . . . . . . . . . . . . . . . . 51
4.5 Comparison between the ergodic secrecy sum-rates SBCC and SrBCC achieved bythe RCI precoder and by the proposed RCI-PR precoder, respectively, for N = 10
transmit antennas. Three values of β are considered: 1.2, 1.4, and 1.6, correspond-ing to K = 12, 14, and 16 users. . . . . . . . . . . . . . . . . . . . . . . . . . . 54
xv
4.6 Comparison between the simulated ergodic per-user secrecy rate with RCI-PR(solid) and the two upper bounds: (i) per-user rate without secrecy requirements(dashed) and (ii) MISOME secrecy capacity (dotted), for K = 12 users. Threevalues of β are considered: 0.8, 1, and 1.2, corresponding to N = 15, 12, and 10
antennas. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 584.7 Comparison between the per-antenna secrecy sum-rate SiBCC/N with RCI pre-
coding in the large-system regime and the simulated ergodic secrecy sum-rateSiBCC/N , for finite N , and in the presence of a channel estimation error τ = 0.1.Three sets of curves are shown, each one corresponds to a different value of β. . . 62
4.8 Comparison between the ergodic per-user secrecy rates SBCC/K and SiBCC/K
with RCI-PR precoder, for N = 10, in the presence of perfect CSI and in thepresence of a channel estimation error τ2 = C
ρ , with C obtained from (4.37) forlog2 b = 1 bit. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.9 Optimal relative amount of training Tt/T vs high-SNR approximation, for N =
K = 10 and rρ = ρ/ρul = 10. . . . . . . . . . . . . . . . . . . . . . . . . . . . 654.10 Comparison between the simulated ergodic per-antenna secrecy sum-rate ScBCC/N
from (3.48) and the large-system approximation ScBCC/N from (4.50), for ν = 0.5
and various values of β. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 704.11 Relative secrecy rate loss (RBCC −RcBCC)/RBCC as a function of the correlation
coefficient ν, for β = 0.8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 714.12 CCDF of the normalized secrecy rate difference between using: (i) ξcBCC obtained
from (4.51) and (ii) ξc?BCC obtained by bi-sectional search for every channel real-ization, for β = 1, and ν = 0.5. . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.1 Example of a BCCE with K = 5 malicious users and a density of external eaves-droppers λe = 0.2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.2 Comparison between the simulated probability of outage OBCCE,k and the large-system results OBCCE provided in Theorem 5.2 and Theorem 5.7, for a networkload β = 1, an SNR ρ = 10dB, and various values of λe. . . . . . . . . . . . . . 85
5.3 Comparison between the simulated ergodic per-antenna secrecy sum-rate undernon-colluding and colluding eavesdroppers, and the large-system results fromTheorem 5.10, for λe = 0.1, N = 10 transmit antennas, and various values ofthe network load β. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.4 Comparison between the simulated ergodic per-user secrecy rate E[Rk] under non-colluding and colluding eavesdroppers, and the large-system results R from The-orem 5.10, for a network load β = 1, an SNR ρ = 10dB, and various values ofλe. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.5 Comparison between the simulated ergodic per-user secrecy rates of: (i) the BCCEwith non-colluding eavesdroppers, (ii) the BCCE with colluding eavesdroppers,and (iii) the BCC without external eavesdroppers, for a network load β = 1, anSNR ρ = 10dB, and various values of λe. . . . . . . . . . . . . . . . . . . . . . 87
5.6 Comparison between the large-system regularization parameter ξBCCE in (5.29)and the value ξBCCE that maximizes the average simulated secrecy sum-rate SBCCE
for a finite system with N = 10 transmit antennas, a network load β = 1, and anSNR ρ = 10dB. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.7 Normalized mean secrecy sum-rate difference between using ξ?BCCE, that maxi-mizes each realization of the secrecy sum-rate SBCCE, and ξBCCE, obtained fromlarge-system analysis in (5.29), under colluding eavesdroppers, for a network loadβ = 1, various values of the density of eavesdroppers λe, and various values ofthe SNR ρ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
6.1 Illustration of a cellular network. The star denotes a typical user. The circles,squares, and triangles denote BSs, out-of-cell users, and in-cell users, respectively,as discussed in Subsection 6.1.3. . . . . . . . . . . . . . . . . . . . . . . . . . . 92
6.2 Example of percolation in a random plane network. Dots represent malicioususers, and discs represents the cooperation range of malicious users. Two mali-cious users can cooperate when their respective discs overlap. . . . . . . . . . . . 96
6.3 Comparison between the simulated ergodic secrecy rate RCELL in (6.11) and theapproximation RCELL in (6.23) versus the SNR, for N = 20 transmit antennas,an average of K = 20 users per BS, and η = 4. . . . . . . . . . . . . . . . . . . 101
6.4 Comparison between the simulated cumulative distribution functions (CDFs) of Iand L and the lognormal approximations in (6.32) and (6.33), for an SNR ρ =
10dB, N = 20 transmit antennas, K = 20 users per BS, ‖c‖ = r, and η = 4. . . 1036.5 Comparison between the simulated probability of secrecy outage OCELL and the
analytical result from Theorem 6.3, for N = 20 transmit antennas, K = 20 usersper BS, and three values of the density of BSs λb. . . . . . . . . . . . . . . . . . 107
6.6 Comparison between the simulated mean secrecy rate RCELL and the analyticalresult from Theorem 6.4, for N = 20 transmit antennas, K = 20 users per BS,and two values of the density of BSs λb. . . . . . . . . . . . . . . . . . . . . . . 108
6.7 Simulated probability of secrecy outage versus transmit SNR, for K = 10 usersper BS and various values of the number of transmit antennas N and density ofBSs λb. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6.8 Simulated ergodic secrecy rate versus transmit SNR, forK = 10 users per BS andvarious values of the number of transmit antennas N and density of BSs λb. . . . 110
6.9 Simulated ergodic secrecy rate versus density of BSs, for N = 20 transmit anten-nas, K = 20 users per BS, and various values of the transmit SNR ρ. . . . . . . . 110
xviii
List of Tables
3.1 Algorithm for optimal power allocation. . . . . . . . . . . . . . . . . . . . . . . 343.2 Algorithm for joint optimal power allocation and regularization parameter. . . . . 35
6.1 Notation Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
xix
List of Acronymsa.s. Almost Surely.BC Broadcast Channel.BCC Broadcast Channel with Confidential messages.BCCE Broadcast Channel with Confidential Messages and External Eavesdroppers.BS Base Station.CCDF Complementary Cumulative Distribution Function.CDF Cumulative Distribution Function.CI Channel Inversion.CI-PR Channel Inversion with Power Reduction.CSI Channel State Information.DPC Dirty Paper Coding.FDD Frequency Division Duplex.GSVD Generalized Single Value Decomposition.i.i.d. Independent and Identically Distributed.MIMO Multiple-Input Multiple-Output.MISO Multiple-Input Single-Output.MISOME Multiple-Input Single-Output Multiple-Eavesdropper.MRC Maximal Ratio Combining.pdf Probability Density Function.PGFL Probability Generating Functional.PPP Poisson Point Process.RCI Regularized Channel Inversion.RCI-EP Regularized Channel Inversion with Equal Power Allocation.RCI-PR Regularized Channel Inversion with Power Reduction.RMT Random Matrix Theory.r.v. Random Variable.RVQ Random Vector Quantization.SG Stochastic Geometry.SINR Signal-to-Interference-Plus-Noise Ratio.SNR Signal-to-Noise Ratio.SUB Single User Beamforming.TDD Time Division Duplex.w.p. With Probability.
xxi
Basic NotationsR Real numbers.C Complex numbers.x Scalar in R or C.x Column vector.xk The kth element of x.X Matrix.I Identity matrix.xk The kth column of X.(·)∗ Complex conjugate.(·)T Transpose.(·)† Conjugate transpose.(·)−1 Inverse.tr · Trace.‖x‖ Euclidean norm.log Natural logarithm.log2 Logarithm in base 2.lim Limit.arg Argument.max Maximum.min Minimum.|·| Set cardinality.B Complement of a set B., Defined as.∀ For all.Ex [·] Mean with respect to random variable x.x ∼ CN (µ, σ2) Complex Gaussian random variable with mean µ and variance σ2.x ∼ Γ(k, θ) Gamma-distributed random variable with shape k and scale θ.x ∼ exp(λ) Exponential random variable with mean 1/λ.P Probability.1· Indicator function.[·]+ Maximum between the quantity in the brackets and zero.Lf Laplace transform of f .sgn (·) Sign function.erfc (·) Complementary error function.δ (·) Dirac’s delta.
xxiii
Chapter 1
Introduction
1.1 Background and Motivation
O ver the last two decades, wireless communications have advanced tremendously, becom-
ing an indispensable part of our lives. Wireless networks have become more and more
pervasive to guarantee global digital connectivity, and wireless devices have quickly evolved into
multimedia smartphones, which run applications that demand high-speed data connections. Multi-
user multiple-input multiple-output (MIMO) wireless techniques have received much attention as
a way to meet such demand by achieving high spectral efficiency LIM et al. [2013]. In a multi-user
MIMO wireless system, a central multi-antenna base station (BS) simultaneously communicates
to several users over the same frequency band. While it is known that the sum-capacity of multi-
user MIMO systems is achieved by using dirty paper coding (DPC) CAIRE AND SHAMAI [2003],
DPC requires high-complexity coding schemes that make it too complex to be implemented LI
et al. [2010].
Suboptimal precoding schemes have proven to be practical and effective in controlling inter-user
interference for the downlink of multi-user MIMO networks HOCHWALD et al. [2005]. Among
those, linear precoding schemes were proposed as a low-complexity alternative to DPC for multi-
user MIMO downlink implementations SPENCER et al. [2004a]. A popular and practical linear
precoding scheme to control inter-user interference is channel inversion (CI) precoding, some-
times known as zero forcing precoding YOO AND GOLDSMITH [2006]. To increase the sum-rate
1
2 Background and Motivation 1.1
performance of the CI precoder, the regularized channel inversion (RCI) precoder was proposed
to tradeoff the inter-user interference and the desired signal through a regularization parameter
PEEL et al. [2005]. The performance of these linear precoding schemes has been studied by a
large-system approach that employs random matrix theory (RMT) tools WAGNER et al. [2012].
As well as spectral efficiency, security is regarded as a critical concern in wireless multiuser net-
works, since users rely on these networks to transmit sensitive data. Due to the broadcast nature of
the physical medium, wireless multiuser communications are very susceptible to eavesdropping,
and it is essential to protect the transmitted information. Security of wireless communications has
traditionally been ensured by network layer key-based cryptography. However, classical cryptog-
raphy may not be suitable in the case of large dynamic wireless networks, since it raises issues like
key distribution and management (for symmetric cryptosystems) and high computational complex-
ity (for asymmetric cryptosystems). Moreover, classical cryptography is potentially vulnerable,
because it relies on the unproven assumption that certain mathematical functions are hard to invert
LIANG et al. [2009b]; LIU AND TRAPPE [2010]. To provide an additional level of protection and
to achieve perfect secrecy without requiring encryption keys, methods exploiting the randomness
inherent in noisy channels, known as physical layer security, have been proposed WYNER [1975];
CSISZAR AND KORNER [1978].
In the past few years, physical layer security has become a very active area of research, and has
witnessed significant growth MUKHERJEE et al. [2013]. The maximum rate at which a mes-
sage can be reliably transmitted to an intended user while the rate of information leakage at the
eavesdroppers vanishes asymptotically with the code length, denoted the secrecy capacity, has
been studied for several network topologies. The secrecy capacity of the wiretap channel, a three-
terminal network consisting of a transmitter, an intended user and an eavesdropper, was derived for
the MIMO case when all terminals have full channel state information (CSI) KHISTI AND WOR-
NELL [2010] . These results were then extended to the case of multiple users with a single external
eavesdropper, known as the multi-receiver wiretap channel EKREM AND ULUKUS [2011]. When
the eavesdropper’s channel is not known by the transmitter, it was shown that the transmission
of artificial noise is an effective method to reduce the eavesdropper’s signal-to-noise ratio GOEL
AND NEGI [2008]. Recently, the study of physical layer security was also extended to networks
where the intended users can act maliciously as eavesdroppers, but only two-user systems have
been considered in the literature LIU et al. [2013]. Determining the achievable secrecy rates for
1.1 Introduction 3
multi-user networks where any number of intended users are potentially eavesdropping remains
an open problem. Moreover, the performance of practical transmission schemes in such generic
multi-user networks is also unknown.
In this thesis, we study physical layer security for generic multi-user multi-antenna systems. We
propose a practical linear transmission scheme based on RCI precoding to simultaneously transmit
independent confidential messages to several spatially dispersed malicious users. We first consider
the case of an isolated cell, where all malicious users can eavesdrop on each other. We carry out a
large-system analysis and obtain closed form expressions for the achievable secrecy rates. These
expressions allow us to optimize several design parameters, like the regularization parameter of
the precoder, the power allocation, and the network load. We also extend our analysis to practical
scenarios where only imperfect CSI is available at the transmitter, and where channel correlation
is present among the transmit antenna elements.
We then consider confidential broadcasting in the presence of a random field of eavesdroppers.
In this case, not just malicious users, but also randomly located external nodes can act as eaves-
droppers. We study the performance of RCI precoding, and provide explicit expressions for the
probability of secrecy outage and the mean secrecy rate, for the two cases of non-colluding and
colluding eavesdroppers. We find that, as expected, the presence of external eavesdropping nodes
incurs a secrecy rate loss. However, irrespective of the collusion strategy at the external eaves-
droppers, a large number of transmit antennas drives this secrecy rate loss to zero. We also find
that increasing the density of eavesdroppers by a factor n, requires n2 as many antennas to meet
a given probability of secrecy outage. Our analysis demonstrates that the number of transmit an-
tennas at the BS is a key resource to secure communications against malicious users and external
eavesdropping nodes.
We finally turn our attention to cellular networks where, unlike the case of isolated cells, multiple
BSs simultaneously transmit confidential messages to several users, generating inter-cell inter-
ference. Moreover in cellular networks, malicious users of neighboring cells can cooperate to
eavesdrop. We obtain the probability of secrecy outage and the mean secrecy rate achievable with
RCI precoding, and we find that a non-zero secrecy rate can be achieved with probability of outage
smaller than one. Unlike isolated cells, we find that the secrecy rate in a cellular network does not
grow monotonically with the transmit power, and the network tends to be in secrecy outage if the
transmit power grows unbounded. Furthermore, we show that there is an optimal value for the
4 Thesis Overview and Contributions 1.2
density of BSs that maximizes the secrecy rate, and we thus establish the importance of designing
the BS deployment density to make communications robust against malicious users in other cells.
1.2 Thesis Overview and Contributions
In this section we outline the contributions of the thesis.
Chapter 3
Chapter 3 considers the multiple-input single-output (MISO) broadcast channel with confidential
messages (BCC), where a multi-antenna base station simultaneously transmits independent con-
fidential messages to several single-antenna users. First, we study the secrecy rates achievable
by a linear precoder based on regularized channel inversion (RCI). We then propose an algorithm
to jointly optimize the regularization parameter of the precoder and the power allocation vector.
Finally, we extend the secrecy rate analysis to the BCC with imperfect channel state information
(CSI) and transmit channel correlation.
The work in this chapter is based on the papers listed in Section 1.3 as J1 and C1.
New Contributions:
• Theorem 3.3: For a given channel matrix, we derive a secrecy sum-rate achievable by linear
precoding in the MISO BCC.
• Corollary 3.4: For a given channel matrix, we derive a secrecy sum-rate achievable by
regularized channel inversion precoding in the MISO BCC.
• Table 3.1: We propose an algorithm that optimizes the power allocation vector of the RCI
precoder, for a given regularization parameter.
• Table 3.2: We propose an algorithm that jointly optimizes the power allocation vector and
the regularization parameter of the RCI precoder.
• Subsection 3.4.1: We derive a secrecy sum-rate achievable by RCI precoding in the presence
of imperfect channel state information.
• Subsection 3.4.3: We derive a secrecy sum-rate achievable by RCI precoding in the presence
of transmit channel correlation.
1.2 Introduction 5
Chapter 4
In Chapter 4, we obtain deterministic approximations for the large-system secrecy rates achievable
by RCI precoding in the MISO BCC. We also derive the optimal regularization parameter and the
optimal network load. We then propose a precoder based on RCI and power reduction (RCI-
PR) that significantly increases the performance for large network loads and large signal-to-noise
ratio (SNR). Finally, we extend our large-system analysis to the case of imperfect CSI in both
frequency division duplex (FDD) and time division duplex (TDD) systems, and to the case of
transmit channel correlation.
The work in this chapter is based on the papers listed in Section 1.3 as J2, J3, C2, and C3.
New Contributions:
• Theorem 4.1: We obtain a deterministic large-system approximation for the secrecy rates
achievable by RCI precoding in the BCC.
• Theorem 4.2: We derive the optimal regularization parameter of the RCI precoder that max-
imizes the large-system secrecy sum-rate in the BCC.
• Theorem 4.3: We obtain a compact expression for the large-system optimal secrecy sum-
rate achievable by RCI precoding in the BCC under unitary network load.
• Proposition 4.1: We derive the optimal network load that maximizes the secrecy sum-rate
in the limit of large SNR.
• Proposition 4.2: We derive the value of the SNR that maximizes the secrecy sum-rate for
network loads larger than one, as well as the corresponding maximum secrecy sum-rate.
• Subsection 4.2.2: We propose a linear precoder based on RCI and power reduction (RCI-
PR), which increases the performance for large SNR and network loads larger than one.
• Theorem 4.4: We provide high-SNR approximations for the secrecy sum-rate achievable by
the proposed RCI-PR precoder.
• Section 4.3: We compare the performance of the proposed RCI-PR precoder to two upper
bounds obtained without secrecy requirements and without interference, respectively. We
also compare the RCI-PR precoder to other linear precoding schemes of similar complexity.
• Theorem 4.5: We provide a deterministic approximation for the secrecy rates achievable by
RCI precoding in the BCC under imperfect CSI.
6 Thesis Overview and Contributions 1.2
• Corollary 4.6: We obtain the minimum number of feedback bits required in an FDD system
in order to meet a given rate gap with the case of perfect CSI.
• Theorem 4.7: We study the optimal value of the training interval that maximizes the high-
SNR secrecy sum-rate in a TDD system.
• Theorem 4.8: We obtain a deterministic approximation for the secrecy rate achievable by
RCI precoding in the BCC under transmit channel correlation.
• Theorem 4.10: We study the optimal regularization parameter that maximizes the secrecy
sum-rate of RCI precoding in the BCC under transmit channel correlation.
Chapter 5
In Chapter 5 we introduce the MISO broadcast channel with confidential messages and external
eavesdroppers (BCCE), where a multi-antenna base station simultaneously communicates to mul-
tiple malicious users, in the presence of randomly located external eavesdroppers. We find that,
irrespective of the collusion strategy at the external eavesdroppers, a large number of transmit an-
tennas drives the probability of secrecy outage to zero. We finally show that increasing the density
of eavesdroppers by a factor n, requires n2 as many antennas to meet a given probability of secrecy
outage and a given mean secrecy rate.
The work in this chapter is based on the papers listed in Section 1.3 as J4 and C4.
New Contributions:
• Lemma 5.1: We derive the probability of secrecy outage for the RCI precoder in the BCCE
under non-colluding eavesdroppers.
• Theorem 5.2: We obtain a large-system approximation for the probability of secrecy outage
with RCI precoding in the BCCE under non-colluding eavesdroppers.
• Lemma 5.4: We derive the probability of secrecy outage in the BCCE caused by the external
eavesdropper nearest to the base station.
• Theorem 5.5: We obtain a large-system approximation for the probability of secrecy outage
in the BCCE caused by the external eavesdropper nearest to the base station.
• Lemma 5.6: We derive the probability of secrecy outage for the RCI precoder in the BCCE
under colluding eavesdroppers.
1.2 Introduction 7
• Theorem 5.7: We obtain a large-system approximation for the probability of secrecy outage
with RCI precoding in the BCCE under colluding eavesdroppers.
• Remark 5.1: We show that the collusion among eavesdroppers does not significantly affect
the number of transmit antennas required to meet a given probability of secrecy outage.
Moreover we show that increasing the density of eavesdroppers by a factor n requires n2 as
many antennas in order to meet a given probability of secrecy outage.
• Lemma 5.9: We derive the mean secrecy rate achievable by RCI precoding in the BCCE.
• Theorem 5.10: We derive a large-system approximation for the mean secrecy rate achievable
by RCI precoding in the BCCE.
• Remark 5.2: We show that, irrespective of the collusion strategy at the external eavesdrop-
pers, a large number of transmit antennas drives the probability of secrecy outage to zero.
• Lemma 5.12: We study the optimal regularization parameter of the RCI precoder in the
BCCE in two extreme cases.
Chapter 6
Chapter 6 considers physical layer security for the downlink of cellular networks. For this scenario,
we show that RCI precoding can achieve a non-zero secrecy rate with probability of outage smaller
than one. However we also show that unlike isolated cells, the secrecy rate in a cellular network
does not grow monotonically with the signal-to-noise ratio (SNR), and the network tends to be in
secrecy outage if the transmit power grows unbounded. We finally show that there is an optimal
value for the density of BSs that maximizes the secrecy rate, and this value is a decreasing function
of the transmit SNR.
The work in this chapter is based on the papers listed in Section 1.3 as J5 and C5.
New Contributions:
• Proposition 6.1: We obtain an expression for the secrecy rate achievable by RCI precoding
in the downlink of a cellular network.
• Lemma 6.2: We derive the Laplace transform of the inter-cell information leakage in a
cellular network.
• Theorem 6.3: We obtain an approximation for the probability of secrecy outage with RCI
precoding in a cellular network.
8 Thesis Overview and Contributions 1.2
• Remark 6.1: We find that in cellular networks, RCI precoding can achieve confidential
communication with probability of secrecy outage smaller than one. However unlike an
isolated cell, cellular networks tend to be in secrecy outage w.p. 1 if the transmit power
grows unbounded, irrespective of the number of transmit antennas.
• Theorem 6.4: We obtain an approximation for the mean secrecy rate achievable by RCI
precoding in a cellular network.
• Remark 6.2: We find that in cellular networks, RCI precoding can achieve a non-zero
secrecy rate. However unlike an isolated cell, the secrecy rate in a cellular network is
interference-and-leakage-limited, and it cannot grow unbounded with the SNR, irrespective
of the number of transmit antennas.
• Remark 6.3: We find that in a cellular network with a fixed load, there is an optimal value
for the deployment density of BSs that maximizes the mean secrecy rate, and this value is a
decreasing function of the SNR.
• Corollary 6.5: We obtain a lower bound on the secrecy rate in a cellular network, which can
be calculated without knowledge of the probability density functions of inter-cell interfer-
ence and information leakage.
1.3 Introduction 9
1.3 Publications
The following is a list of publications in books, refereed journals, and conference proceedings pro-
duced during this Ph.D. candidature. In some cases the journal papers contain materials partially
presented in the conference papers.
Book Chapter
GERACI, G. AND J. YUAN,. “Physical Layer Security for Multi-User MIMO Communica-
tions.” in Recent Trends in Multi-User MIMO Communications, InTech Publisher, 2013.
Journal Papers
J1 GERACI, G., M. EGAN, J. YUAN, A. RAZI, AND I. B. COLLINGS “Secrecy sum-rates for
multi-user MIMO regularized channel inversion precoding.” IEEE Trans. on Communica-
tions, vol. 60, no. 11, pp. 3472–3482, Nov. 2012.
J2 GERACI, G., A. Y. AL-NAHARI, J. YUAN, AND I. B. COLLINGS “Linear precoding
for broadcast channels with confidential messages under transmit-side channel correlation.”
IEEE Communications Letters, vol. 17, no. 6, pp. 1164–1167, June 2013.
J3 GERACI, G., R. COUILLET, J. YUAN, M. DEBBAH, AND I. B. COLLINGS “Large system
analysis of linear precoding in MISO broadcast channels with confidential messages.” IEEE
Journal on Selected Areas in Communications, vol. 31, no. 9, pp. 1660–1671, Sep. 2013.
(Second prize of the 2012-2013 IEEE Australia Council Student Paper Contest).
J4 GERACI, G., S. SINGH, J. G. ANDREWS, J. YUAN, AND I. B. COLLINGS “Secrecy rates
in the broadcast channel with confidential messages and external eavesdroppers.” IEEE
Trans. on Wireless Communications, accepted for publication.
J5 GERACI, G., H. S. DHILLON, J. G. ANDREWS, J. YUAN, AND I. B. COLLINGS “Physical
layer security in downlink multi-antenna cellular networks.” IEEE Trans. on Communica-
tions, under review.
10 Publications 1.3
Conference Papers
C1 GERACI, G., J. YUAN, A. RAZI, AND I. B. COLLINGS “Secrecy sum-rates for multi-
user MIMO linear precoding.” In Proc. of IEEE Int. Symp. on Wireless Commun. Systems
(ISWCS’11). Aachen, Germany, Nov. 2011.
C2 GERACI, G., J. YUAN, AND I. B. COLLINGS “Large system analysis of the secrecy sum-
rates with regularized channel inversion precoding.” In Proc. of IEEE Wireless Commun.
Networking Conference (WCNC’12). Paris, France, Apr. 2012.
C3 GERACI, G., R. COUILLET, J. YUAN, M. DEBBAH, AND I. B. COLLINGS “Secrecy sum-
rates with regularized channel inversion precoding under imperfect CSI at the transmitter.”
In Proc. of IEEE Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP’13). Van-
couver, Canada, May 2013.
C4 GERACI, G., S. SINGH, J. G. ANDREWS, J. YUAN, AND I. B. COLLINGS “MIMO Multi-
user secrecy rate analysis.” In Proc. of IEEE Int. Conf. on Commun. (ICC’14). Sydney,
Australia, June 2014, accepted for publication.
C5 GERACI, G., H. S. DHILLON, J. G. ANDREWS, J. YUAN, AND I. B. COLLINGS “A
new model for physical layer security in cellular networks.” In Proc. of IEEE Int. Conf. on
Commun. (ICC’14). Sydney, Australia, June 2014, accepted for publication.
During the course of these Ph.D. studies, several other joint contributory papers were produced.
They are not included in this thesis in order to maintain focus, however the respective details can
be found in the list below.
Journal Papers
J6 YANG, N., G. GERACI, J. YUAN, AND R. MALANEY “Confidential broadcasting via
linear precoding in non-homogeneous MIMO multiuser networks.” IEEE Trans. on Com-
munications, under review.
J7 YAN, S., G. GERACI, N. YANG, R. MALANEY, AND J. YUAN “Optimization of code rates
in SISOME wiretap channels.” IEEE Trans. on Wireless Communications, under review.
1.3 Introduction 11
Conference Papers
C6 KOBAYASHI, M., S. YANG, AND G. GERACI “Time-correlated MISO wiretap channel
with delayed CSIT.” In Proc. of IEEE Int. Workshop on Signal Processing Advances in
Wireless Commun. (SPAWC’13). Darmstadt, Germany, June 2013.
C7 LIU, C., G. GERACI, N. YANG, J. YUAN, AND R. MALANEY “Beamforming for MIMO
Gaussian wiretap channels with imperfect channel state information.” To appear in Proc. of
IEEE Global Communications Conf. (Globecom’13). Atlanta GA, USA, Dec. 2013.
C8 YAN, S., G. GERACI, N. YANG, R. MALANEY, AND J. YUAN “On the target secrecy rate
for SISOME wiretap channels.” In Proc. of IEEE Int. Conf. on Commun. (ICC’14). Sydney,
Australia, June 2014, accepted for publication.
C9 LIU, C., N. YANG, G. GERACI, J. YUAN, AND R. MALANEY “Secrecy in MIMOME
wiretap channels: Beamforming with imperfect CSI.” In Proc. of IEEE Int. Conf. on Com-
mun. (ICC’14). Sydney, Australia, June 2014, accepted for publication.
12
Chapter 2
Overview of Physical Layer Security
Summary This chapter provides a brief overview of physical layer security, which is a tech-
nique that enables the exchange of confidential messages over a wireless medium in the presence
of unauthorized receivers, without relying on key-based encryption. We first introduce the notion
of physical layer security for the simple wiretap channel, and then extend our discussion to fading
channels, multi-antenna channels, and multi-user channels. This chapter does not provide new
results, and it is intended to give some background information necessary for the understanding
of the rest of the thesis.
13
14 The Wiretap Channel 2.2
2.1 Introduction to Physical Layer Security
D ue to the broadcast nature of wireless communications, it is essential to protect the transmit-
ted signals from unintended receivers. Security of data transmission has traditionally been
achieved via key-based cryptographic techniques at the network layer MASSEY [1988]. However,
in dynamic wireless networks this raises issues such as key distribution (for symmetric cryptogra-
phy), and high computational complexity (for asymmetric cryptography). In addition, key-based
cryptography is based on the unproven assumption that it is computationally infeasible to decipher
a message without knowledge of the secret key SCHNEIER [1998]. These vulnerabilities motivated
a new information-theoretic approach to security at the physical layer. The fundamental principle
behind physical layer security is to exploit the features of wireless channels, i.e., fading and noise,
to control the amount of information that can be extracted by an unauthorized receiver. This is
possible by appropriately designing coding and transmit precoding schemes, and does not rely
on a secret key nor assumes any limitations for the eavesdropper’s computational power. While
the first studies on physical layer security for noisy channels date back to the ’70s, this field has
experienced a renewed interest only in the last decade.
This thesis deals with physical layer security for multi-user systems. Although this is meant as a
self-contained piece of work, we here assume that the reader is familiar with fading channels and
multiple-input multiple-output (MIMO) technologies FOSCHINI AND GANS [1998]; TELATAR
[1999]. Due to the advances in wireless communications, this knowledge is now readily available
in textbooks RAPPAPORT [1996]; GOLDSMITH [2005]; TSE AND VISWANATH [2005]; VUCETIC
AND YUAN [2003]. Rather than repeating this standard knowledge, the main focus of this chapter
is to review various aspects of physical layer security in modern wireless networks.
2.2 The Wiretap Channel
The foundations of information-theoretic security were laid in SHANNON [1949]. For a simple
three-terminal network comprising a transmitter, a legitimate receiver, and an eavesdropper, Shan-
non introduced the notion of perfect secrecy, by requiring that the a posteriori probability of the
secret message computed by the eavesdropper be equal to the a priori probability of the message.
Shannon showed that in order to achieve perfect secrecy with a non-reusable private key, the key
must be at least as long as the message.
2.2 Overview of Physical Layer Security 15
Source Encoder Main
channel Decoder
Wiretapchannel
𝑆𝑘 𝑋𝑛 𝑌𝑛
𝑍𝑛
𝑆 𝑘
Figure 2.1: The wiretap channel in WYNER [1975], where the eavesdropper’s channel is a de-
graded version of the main channel.
Physical layer security for noisy channels was first proposed by Wyner, who introduced the wiretap
channel WYNER [1975]. As shown in Fig. 2.1, in the wiretap channel a confidential message X
is transmitted to the legitimate receiver over a discrete memoryless channel (main channel). The
legitimate receiver observes Y , which then passes through an additional channel (wiretap channel)
before being received by the eavesdropper as Z. Unlike Shannon, whose notion of perfect secrecy
required that the mutual information between transmitter and eavesdropper be zero regardless of
the block length, Wyner considered the block-length-normalized mutual information, and defined
the equivocation rate of the eavesdropper as
Re ≤H(Sk|Zn
)n
, (2.1)
where the conditional entropyH(Sk|Zn
)represents the equivocation. If the equivocation rateRe
is arbitrarily close to the information rate R, then R is the secrecy capacity of the wiretap channel.
By assuming that the transmitter-eavesdropper link is a probabilistically degraded version of the
main channel, Wyner constructed a coding scheme to hide the information in the additional noise
impairing the eavesdropper, by mapping each confidential message to many codewords according
to an appropriate probability distribution. This way, one can maximize the transmission rate R in
the main channel while making negligible the amount of information leaked to the eavesdropper.
In LEUNG-YAN-CHEONG AND HELLMAN [1978], the authors considered the degraded wiretap
channel with additive Gaussian noise. They showed that the secrecy capacity Cs is given by
Cs = Cm − Cmw, (2.2)
16 Fading Channels 2.3
where Cm and Cmw are the capacities of the main channel and of the overall wiretap channel,
respectively. Therefore, a non-zero secrecy capacity can only be obtained if the eavesdropper’s
overall wiretap channel is worse than the legitimate receiver’s main channel.
A non-degraded version of Wyner’s wiretap channel was considered in CSISZAR AND KORNER
[1978], where the authors characterized the secrecy capacity as
Cs = maxV→X→Y Z
I(V ;Y )− I(V ;Z). (2.3)
The secrecy capacity is achieved by maximizing over all joint probability distributions such that a
Markov chain V , X , Y Z is formed, where V is an auxiliary input variable.
2.3 Fading Channels
Early work on physical layer security assumed non-fading channels and perfect channel state
information at the transmitter. More recently, the research community has considered channel
fading in wiretap channels as well as limited channel state information. It is usually assumed that at
least the statistics of the eavesdropper’s channel are known to the transmitter. This allows to define
secrecy rate outage metrics similarly to the conventional rate outage metrics. For instance, the
secrecy outage probability is the probability that the instantaneous secrecy rate falls below a given
threshold. The secrecy outage probability of slow fading channels was analyzed in BARROS AND
RODRIGUES [2006]; BLOCH et al. [2006], where the authors showed that secret communication is
possible even if the eavesdropper has a better average SNR than the one of the legitimate receiver.
In the special case when only the eavesdropper’s channel is affected by fading, and when this
fading is unknown, a transmission scheme was proposed in LI et al. [2007b] to achieve a positive
secrecy rate even when the main channel is arbitrarily worse than the eavesdropper’s average
channel. The case when no information at all is available about the eavesdropper’s channel was
studied, among others, in HE AND YENER [2010]; SWINDLEHURST [2009]; XIE AND ULUKUS
[2013]. For this case, a coding scheme that hides the secure message across different fading
states was proposed in GOPALA et al. [2008]. A different approach to this problem was taken
in LIANG et al. [2009a] by studying the compound wiretap channel, i.e., by characterizing the
eavesdropper’s channel with a finite set of states, and guaranteeing secure communication under
any state that may occur.
2.4 Overview of Physical Layer Security 17
𝑁𝑡
Eavesdropper
Transmitter
𝑁𝑟
𝑁𝑒
Legitimate receiver
Figure 2.2: The general MIMOME wiretap channel.
2.4 Multi-antenna Channels
When the transmitter is equipped with multiple antennas, it is possible to exploit the available
spatial dimensions to increase the secrecy rates of wireless channels. In a multiple-input multiple-
output (MIMO) system where the transmitter, legitimate receiver, and eavesdropper are equipped
with Nt, Nr, and Ne antennas respectively as in Fig. 2.2, the signals received by the legitimate
receiver and eavesdropper can be written as
yb = Hbxa + nb
ye = Hexa + ne
(2.4)
where xa ∈ CNt×1 is the transmit signal with covariance matrix E[xax†a] = Qx and average
power constraint trQx ≤ P , Hb ∈ CNt×Nr and He ∈ CNe×Nr are the Gaussian channel
matrices at the legitimate receiver and eavesdropper, respectively, and nb and ne are the respective
Gaussian noise vectors. Physical layer security in MIMO channels was first considered in HERO
[2003], which examined the utility of space-time block coding.
The secrecy capacity of the degraded single-input multiple-output (SIMO) wiretap channel was
obtained in PARADA AND BLAHUT [2005], whereas LI et al. [2007a]; SHAFIEE AND ULUKUS
[2007] studied the multiple-input single-output (MISO) case. The general MIMO case was consid-
ered in KHISTI et al. [2007]; OGGIER AND HASSIBI [2008] which introduced the multiple-input
multiple-output multiple-eavesdropper (MIMOME) wiretap channel. In KHISTI et al. [2007] the
18 Multi-user Channels 2.5
authors showed that when the eavesdropper’s CSI is known, a transmit precoder based upon the
generalized singular value decomposition (GSVD) of the pencil (Hb,He) is optimal in the high-
SNR regime. For the special case of Nr = 1 (MISOME), the optimal transmit beamformer was
shown to be the generalized eigenvector ψm corresponding to the largest generalized eigenvalue
λm of
h†bhbψm = λmH†eHeψm. (2.5)
A few contributions in the literature have also studied the MIMO wiretap channel under a more
general matrix power-covariance constraint LIU AND SHAMAI [2009]; BUSTIN et al. [2009];
FAKOORIAN AND SWINDLEHURST [2013].
The transmission of artificial noise was proposed as strategy to achieve secrecy in the case when
only the statistics of He are known to the transmitter NEGI AND GOEL [2005]; GOEL AND NEGI
[2008]. The artificial noise is transmitted with the information signal, and it is usually designed
to be orthogonal to the legitimate receiver, i.e., in the nullspace of Hb, such that only the eaves-
dropper’s channel is degraded. Although this is a common design choice made for the sake of
simplicity, in general it is not the optimal approach in terms of secrecy rate. An optimal power
allocation and beamforming method for the artificial noise strategy were presented in ZHOU AND
MCKAY [2009].
2.5 Multi-user Channels
The principles of physical layer security can be extended to multi-user networks with more than
two receivers, i.e., broadcast channels. We can distinguish the two following cases: (i) all the users
are trusted, and the confidential messages must be protected only from external eavesdroppers, and
(ii) each message must be kept confidential from all other unintended users, i.e., each user is seen
as an eavesdropper for messages not intended for it, and it is therefore denoted a malicious user.
While in the first case one can still apply most schemes proposed for the single-user scenario, in
the second case new transmission techniques must be designed.
The first case, where one transmitter wants to communicate with several legitimate users in the
presence of an external eavesdropper, was considered, among others, in KHISTI et al. [2008];
BAGHERIKARAM et al. [2013]; EKREM AND ULUKUS [2011]; LIU et al. [2010b]. The second
case, where two confidential messages are sent to two receivers, and each receiver acts as an
2.5 Overview of Physical Layer Security 19
eavesdropper for the other one, was studied in LIU et al. [2008]; LIU AND POOR [2009]. It
was shown in LIU et al. [2010a] that, under the matrix input power-covariance constraint, both
confidential messages can be simultaneously communicated at their respected maximum secrecy
rates by using dirty paper coding.
Although one can find several papers that study the MISOME channel with multiple eavesdrop-
pers, it should be noted that in the MISOME channel no messages are to be delivered to the
eavesdroppers, since they are not users. Other papers have studied the case of malicious users,
which are users acting as eavesdroppers. However, most of the prior art has focused on physical
layer security for systems with up to two malicious users only. Determining the achievable secrecy
rates for multi-user networks where any number of users are potentially eavesdropping remains
an open problem. Moreover, the performance of practical transmission schemes in such generic
multi-user networks is also unknown. Hence in this thesis we study physical layer security for
generic multi-user multi-antenna systems.
20
Chapter 3
Physical Layer Security in Isolated
Cells: Achievable Rates
Summary This chapter studies the multiple-input single-output (MISO) broadcast channel with
confidential messages (BCC). We consider the case where a multi-antenna base station simultane-
ously transmits independent confidential messages to several single-antenna users via regularized
channel inversion (RCI) precoding in the presence of Rayleigh fading. We derive expressions for
the secrecy rates achievable by RCI precoding, and we show that RCI precoding outperforms plain
channel inversion precoding. We then propose an algorithm to jointly optimize the regularization
parameter of the RCI precoder and the power allocation vector, and show that optimal power
allocation increases the secrecy sum-rate compared to equal power allocation. Finally, we derive
the achievable secrecy rates for more practical scenarios where only imperfect channel state in-
formation (CSI) is available at the transmitter, and where channel correlation is present among
the transmit antenna elements.
21
22 The Broadcast Channel with Confidential Messages 3.1
3.1 The Broadcast Channel with Confidential Messages
W e consider the downlink of a narrowband multi-user MISO system, consisting of a base
station (BS) with N antennas which simultaneously transmits K independent confidential
messages to K spatially dispersed single-antenna users, as depicted in Fig. 3.1. In this model,
denoted as the MISO broadcast channel with confidential messages (BCC), transmission takes
place over a block fading channel, and the transmitted signal is x = [x1, . . . , xN ]T ∈ CN×1. We
assume homogeneous users, i.e., each user experiences the same received signal power on average,
thus the model assumes that their distances from the transmitter are the same and unitary.
Note that despite this assumption, which is necessary to maintain tractability, the simplified model
captures all the key characteristics of the broadcast channels that affect physical layer security,
as discussed in the sequel. Our analysis can be extended to the scenario where the mobile users
have different distances from the serving BS. Results on this scenario can be found in YANG et al.
[2013].
The received signal at user k is given by
yk =N∑j=1
hk,jxj + nk (3.1)
where hk,j ∼ CN (0, 1) is the independent and identically distributed (i.i.d.) Rayleigh fading
channel between the jth transmit antenna element and the kth user, and nk ∼ CN (0, σ2) is the
noise seen at the kth receiver. The corresponding vector equation is
y = Hx + n (3.2)
where H = [h1, . . . ,hK ]† is the K ×N channel matrix. We assume E[nn†] = σ2IK , where IK
is the K × K identity matrix, define the transmit signal-to-noise ratio (SNR) as ρ , 1/σ2, and
impose the long-term power constraint E[‖x‖2] = 1.
It is required that the BS securely transmits each confidential message, ensuring that the unin-
tended users receive no information. This is performed at the secrecy rate R, defined as follows.
Let P(En) be the probability of error at the intended user, m be a confidential message, yne be the
vector of all signals received by the unintended users, and H(m|yne ) be the corresponding equivo-
cation. Then a (weak) secrecy rate R for the intended user is achievable if there exists a sequence
3.2 Physical Layer Security in Isolated Cells: Achievable Rates 23
𝑁
Transmitter
User 1
User 2
User K
Figure 3.1: The MISO broadcast channel with confidential messages (BCC).
of (2nR, n) codes such that P(En)→ 0 and 1nH(m|yne ) ≤ 1
nH(m)−εn with εn approaching zero
as n→∞ KHISTI AND WORNELL [2010].
For each user k, we denote byMk = 1, . . . , k − 1, k + 1, . . . ,K the set of remaining users. In
general, the behavior of the users cannot be determined by the BS. As a worst-case scenario, we
assume that for each user k, all users inMk can cooperate to jointly eavesdrop on the kth message.
This assumption is reasonable, because the confidentiality of the messages must be ensured in all
cases, including the worst case. Since the set of malicious usersMk can perform joint processing,
they can be seen as a single equivalent malicious user Mk with K − 1 receive antennas. Due
to the assumption of cooperating malicious users, interference cancellation can be performed at
Mk, which does not see any undesired signal term apart from the received noise. It will be shown
that despite this conservative but necessary assumption, a properly designed linear precoder can
achieve a per-user secrecy rate which is close to an upper bound on the secrecy capacity. Since
the average secrecy sum-rate is simply given by the average per-user secrecy rate multiplied by
number of users, our analysis is suitable for both individual secrecy rate and secrecy sum-rate, and
a properly designed linear precoder performs well also in terms of the secrecy sum-rate.
3.2 Achievable Secrecy Rates in the BCC
In this section, we derive achievable secrecy rates for the MISO BCC by using a linear precoder.
Although suboptimal, linear precoding schemes are of particular interest because of their low-
24 Achievable Secrecy Rates in the BCC 3.2
complexity implementations and because they can control the amount of crosstalk between the
users YOO AND GOLDSMITH [2006]; SPENCER et al. [2004b]; PEEL et al. [2005]; JOHAM et al.
[2005]. We then specialize and obtain the secrecy rates achievable by the regularized channel
inversion (RCI) precoder. RCI is a linear precoding scheme that was proposed to serve multiple
users in the MISO broadcast channel (BC). RCI precoding has better performance than plain
channel inversion, especially at low SNR PEEL et al. [2005].
3.2.1 Linear Precoding
In linear precoding, the transmitted vector x is derived from the vector containing the confidential
messages u = [u1, . . . , uK ]T through a deterministic linear transformation (precoding) YOO AND
GOLDSMITH [2006]; SPENCER et al. [2004b]; PEEL et al. [2005]; JOHAM et al. [2005]. We
assume that the entries of u are chosen independently, satisfying E[|uk|2] = 1, ∀k.
Let W = [w1, . . . ,wK ] be theN ×K precoding matrix, where wk is the kth column of W. Then
the transmitted signal and the power constraint are, respectively:
x = Wu =
K∑k=1
wkuk, (3.3)
E[‖x‖2
]= E
[‖Wu‖2
]=
K∑k=1
‖wk‖2 = 1. (3.4)
By employing linear precoding as in (3.3), the signals observed at the legitimate user k and at the
equivalent malicious user Mk are, respectively
yk = h†kwkuk +∑j 6=k
h†kwjuj + nk
yM,k =∑j
Hkwjuj + nk
(3.5)
where nk = [n1, . . . , nk−1, nk+1, . . . , nK ]T , h†k is the kth row of H, and Hk is a matrix obtained
from H by eliminating the kth row. The channel in (3.5) is a multi-input, single-output, multi-
eavesdropper (MISOME) wiretap channel KHISTI AND WORNELL [2010]. The transmitter, the
intended receiver and the eavesdropper of this MISOME wiretap channel are equipped with N , 1
andK−1 virtual antennas, respectively. Due to the simultaneous transmission of theK messages,
user k experiences noise and interference from all the uj , j 6= k.
3.2 Physical Layer Security in Isolated Cells: Achievable Rates 25
3.2.2 Achievable Secrecy Rates with Linear Precoding
In the following, we derive an achievable secrecy sum-rate SBCC for the MISO BCC. Although
the design of codes for the MISO BCC is not the focus of this thesis, we prove the achievability of
SBCC with a code construction based on independent codebooks and linear precoding.
Lemma 3.1. An achievable secrecy sum-rate SBCC for the MISO BCC is given by
SBCC4=
K∑k=1
RBCC,k, (3.6)
where RBCC,k is an achievable secrecy rate for the kth MISOME wiretap channel in (3.5), k =
1, . . . ,K.
Proof. Assume that the BS uses independent codebooks for each user, where each codebook is
a code for the scalar wiretap channel KHISTI AND WORNELL [2010]. The confidential message
uk is obtained as a codeword independently drawn from the code Ck, corresponding to the kth
user. The rate RBCC,k of the code Ck is chosen according to the secrecy rate achievable for user
k in the presence of the equivalent malicious user Mk, i.e., by the secrecy rate achievable for the
MISOME wiretap channel in (3.5). The existence of such code is guaranteed by the definition
of secrecy rate CSISZAR AND KORNER [1978]. To construct the vector codeword for the MISO
BCC, the scalar codewords for each MISOME wiretap channel are stacked according to u =
[u1, . . . , uK ]T , and no additional binning is required. The vector u is then linearly precoded as in
(3.3), which means that each message uk is transmitted by beamforming, i.e., by signaling with
rank one covariance, along the direction wk. The secrecy sum-rate SBCC is then by definition the
sum of the simultaneously achievable secrecy rates RBCC,k.
Lemma 3.2. An achievable secrecy rate for the MISOME wiretap channel in (3.5) is given by
RBCC,k =[log2
(1 + γk
)− log2
(1 + γM,k
)]+, (3.7)
where γk and γM,k are the signal-to-interference-plus-noise ratios (SINR) for the message uk at
the legitimate receiver k and at the equivalent malicious userMk, respectively, and where we have
used the notation [x]+ , max (x, 0).
26 Achievable Secrecy Rates in the BCC 3.2
Proof. By noting that the MISOME wiretap channel in (3.5) is a nondegraded broadcast channel
KHISTI AND WORNELL [2010], the secrecy capacity is given by CSISZAR AND KORNER [1978]:
Cs = maxuk→wkuk→yk,yM,k
I(uk; yk
)− I(uk; yM,k
)(3.8)
where I(x; y) denotes mutual information between two random variables x and y. The secrecy
capacity Cs is given by the difference of the mutual informations at the intended user and at the
eavesdropper, respectively. Cs is achieved by maximizing over all joint probability distributions
such that a Markov chain uk → wkuk → yk,yM,k is formed, where uk is an auxiliary input
variable. By evaluating (3.8) with uk ∼ CN (0, 1) and with the linearly precoded data wkuk, we
obtain an achievable secrecy rate for the MISOME wiretap channel (3.5) given by
RBCC,k =[I(uk; yk
)− I(uk; yM,k
)]+(3.9)
(a)=[I(wkuk; yk
)− I(wkuk; yM,k
)]+, (3.10)
where (a) follows from wkuk being a deterministic function of uk KHISTI AND WORNELL
[2010]. Equation (3.7) then follows from (3.10) and from the statistics of uk.
From equation (3.7) it is clearly observed that for high-performance linear precoder design an
efficient tradeoff between maximizing γk and minimizing γM,k is required.
Theorem 3.3. A secrecy sum-rate achievable by linear precoding in the MISO BCC is given by
SBCC =K∑k=1
log2
1 +
∣∣∣h†kwk∣∣∣2σ2+
∑j 6=k
∣∣∣h†kwj∣∣∣21 + ‖Hkwk‖2
σ2
+
. (3.11)
Proof. By using Lemma 3.1 and Lemma 3.2, we have that an achievable secrecy sum-rate is
obtained as the sum of the secrecy rates RBCC,k in (3.7). A lower bound on the quantities RBCC,k
can be obtained by considering a genie-aided equivalent malicious user Mk which observes not
only the signals yM,k received by its K − 1 antennas, but also all the confidential messages
uj , j 6= k. Such channel clearly has an achievable secrecy rate smaller than the original channel.
The genie-aided equivalent malicious user Mk can perform interference cancellation, and it does
not see any undesired signal term apart from the received noise nk.
3.2 Physical Layer Security in Isolated Cells: Achievable Rates 27
According to the previous considerations, the signals at the legitimate user and at the eavesdropper
of the kth equivalent MISOME wiretap channel become, respectively:
yk = h†kwkuk +∑j 6=k
h†kwjuj + nk,
yM,k = Hkwkuk + nk.
(3.12)
For the kth equivalent MISOME wiretap channel in (3.12), the SINRs at the legitimate user and
the eavesdropper are, respectively:
γk =
∣∣∣h†kwk
∣∣∣2σ2 +
∑j 6=k
∣∣∣h†kwj
∣∣∣2 , (3.13)
γM,k =‖Hkwk‖2
σ2. (3.14)
Since the noise in yM,k in (3.12) is spatially white, the optimal receive filter at Mk is the matched
filter (Hkwk)†. Equation (3.14) then follows. For a given channel H, substituting (3.13) and
(3.14) into (3.7) and then into (3.6) yields (3.11).
In the remainder of this chapter, we refer to equation (3.11) as the secrecy sum-rate. We note that
the secrecy sum-rate depends on the choice of the precoding matrix W, as well as on the channel
H and the noise variance σ2. A possible choice for W, based on regularized channel inversion, is
discussed in the following.
3.2.3 Regularized Channel Inversion Precoding
We now consider RCI precoding for the MISO BCC. Although plain channel inversion (CI) pre-
coding can achieve secrecy by canceling all signals leaked at the unintended users (provided that
N ≥ K), this comes at the cost of a poor sum-rate. The RCI precoder has better performance
than plain CI, particularly at low SNR PEEL et al. [2005]. For each message uk, RCI precod-
ing achieves a tradeoff between the signal power at the kth legitimate user and the crosstalk at the
other (K−1) unintended users for each signal. The crosstalk causes interference to the unintended
users. In the case when the unintended users are acting maliciously, the crosstalk also causes in-
formation leakage. Therefore, RCI achieves a tradeoff between signal power, interference, and
information leakage.
28 Achievable Secrecy Rates in the BCC 3.2
With RCI precoding, linear processing exploiting regularization is applied to the vector of mes-
sages u PEEL et al. [2005]. The RCI precoding matrix is given by
W =1√ζH†(HH† +NξIK
)−1=
1√ζ
(H†H +NξIN
)−1H† (3.15)
where
ζ = tr
H†H
(H†H +NξIK
)−2
(3.16)
is the power normalization constant. The transmitted signal x after RCI precoding can be written
as
x = Wu =1√ζH†(HH† +NξIK
)−1u (3.17)
=1√ζ
(H†H +NξIN
)−1H†u. (3.18)
The latter passes through the channel, producing the vector of received signals
y =1√ζH(H†H +NξIK
)−1H†u + n. (3.19)
The function of the regularization parameter ξ ∈ R is to improve the behavior of the inverse,
although it also produces non-zero crosstalk terms in (3.19). We note that when the regularization
parameter is zero, i.e., ξ = 0, the RCI precoder reduces to a CI precoder, which is therefore
a special case of RCI. As a result, an RCI precoder obtained by optimizing the regularization
parameter ξ will achieve a secrecy sum-rate greater or equal to the one achieved by the CI precoder,
i.e., in the special case of ξ = 0.
Using RCI precoding, the SINRs (3.13) and (3.14) at the legitimate user k and at the equivalent
malicious user Mk become, respectively
γk =
∣∣∣h†k (H†H +NξIK)−1
hk
∣∣∣2ζσ2 +
∑j 6=k
∣∣∣h†k (H†H +NξIK)−1
hj
∣∣∣2 , (3.20)
γM,k =
∥∥∥Hk
(H†H +NξIK
)−1hk
∥∥∥2
ζσ2. (3.21)
To rewrite (3.20) and (3.21) in a more compact way, we introduce the quantities
Ak = h†k
(H†kHk +NξIK
)−1hk (3.22)
and
Bk = h†k
(H†kHk +NξIK
)−1H†kHk
(H†kHk +NξIK
)−1hk. (3.23)
3.2 Physical Layer Security in Isolated Cells: Achievable Rates 29
It is then possible to express (3.20) as NGUYEN et al. [2009]
γk =A2k
Bk + ζσ2 (1 +Ak)2 . (3.24)
In a similar fashion, we rewrite (3.21) as
γM,k =Bk
ζσ2 (1 +Ak)2 . (3.25)
We then obtain the following expression for the secrecy sum-rates achievable with RCI precoding.
Corollary 3.4. A secrecy sum-rate achievable by RCI precoding in the MISO BCC is given by
SBCC =
K∑k=1
RBCC,k (3.26)
where RBCC,k is the achievable secrecy rate at user k, given by
RBCC,k =
log2
1 +A2k
Bk+ζσ2(1+Ak)2
1 + Bkζσ2(1+Ak)2
+
. (3.27)
Proof. Corollary 3.4 follows by substituting (3.24) and (3.25) into (3.7) and then into (3.6).
3.2.4 Numerical Results
We now show the performance of RCI precoding in the BCC via simulations. The precoding
matrix W is normalized by√ζ, as in (3.15), in order to meet the power constraint in (3.4). This
corresponds to a long-term power constraint, which does not require the receivers to know the
instantaneous value of ζ PEEL et al. [2005].
Figure 3.2 shows the simulated secrecy sum-rate SBCC of the RCI precoder in the BCC, versus
the SNR ρ, for various values of the number of transmit antennas N and users K, with N = K.
The value of SBCC was averaged over 103 channels, and it was calculated by using the optimal
regularization parameter ξ that maximizes the average secrecy sum-rate. This optimal value of ξ
was obtained by linear search.
In Fig. 3.3 we compare the simulated secrecy sum-rate of the RCI precoder in the MISO BCC
to the one of the plain CI precoder YOO AND GOLDSMITH [2006]. The sum-rate SBC of the
30 Optimal Regularization Parameter and Power Allocation 3.3
0 5 10 15 20 250
10
20
30
40
50
60
70
80
SNR, ρ [dB]
Secrecy
sum-rate
N = K = 4
N = K = 8
N = K = 16
N = K = 32
Figure 3.2: Simulated secrecy sum-rate SBCC achievable by RCI precoding in the BCC versus
SNR ρ, for various values of N = K.
optimal RCI precoder in the MISO BC without secrecy requirements is also plotted. The figure
shows plots for K = 4, 8, 16, 32. We observe that CI precoding exhibits a large performance
loss compared to the secrecy sum-rate of the optimal RCI precoder, especially for large values of
K. We note that although CI precoding achieves secrecy in a simple way by completely canceling
the information leakage, this comes at the cost of a poor sum-rate. Secrecy can be achieved with
a significantly larger sum-rate by using the RCI precoder. We also observe that the secrecy loss,
i.e., the gap between the sum-rate of the RCI precoder without secrecy and the secrecy sum-rate
of the RCI precoder, is almost constant at high SNR for large K. This means that the secrecy
requirements do not affect the multiplexing gain.
3.3 Optimal Regularization Parameter and Power Allocation
In this section, we consider power allocation for the RCI precoder. We first propose a new algo-
rithm to obtain the power allocation vector p which achieves the optimal secrecy sum-rate with a
fixed regularization parameter ξ. We then extend our algorithm to jointly optimize p and ξ.
3.3 Physical Layer Security in Isolated Cells: Achievable Rates 31
0 5 10 15 20 250
5
10
15
20
25
SNR, ρ [dB]
Secrecy
sum-rate
RCI without secrecy
RCI with secrecy
CI with secrecy
(a) K = 4
0 5 10 15 20 250
5
10
15
20
25
30
35
40
SNR, ρ [dB]
Secrecy
sum-rate
RCI without secrecy
RCI with secrecy
CI with secrecy
(b) K = 8
0 5 10 15 20 250
10
20
30
40
50
60
70
80
SNR, ρ [dB]
Secrecy
sum-rate
RCI without secrecy
RCI with secrecy
CI with secrecy
(c) K = 16
0 5 10 15 20 250
50
100
150
SNR, ρ [dB]
Secrecy
sum-rate
RCI without secrecy
RCI with secrecy
CI with secrecy
(d) K = 32
Figure 3.3: Comparison between the RCI precoder and the plain CI precoder, for various values
of N = K. The secrecy loss is also shown as the gap between dashed and solid lines.
3.3.1 Regularized Channel Inversion Precoding with Power Allocation
We consider the RCI precoding matrix with arbitrary power allocation given by
Wp = WD =1√ζH†(HH† +NξIK)−1D, (3.28)
where D = diag(√
p), and p = [p1, . . . , pK ]T is the power allocation vector. The vector p must
be chosen such that the power constraint tr
W†pWp
= 1 is met. Clearly, (3.28) generalizes the
RCI precoder W with equal power allocation (RCI-EP) in (3.15).
Clearly, (3.28) generalizes the RCI precoder W in (3.15). The columns of W will in general have
different norms, and hence each user will be allocated a different power level. However, in order
to make a distinction between the general case Wp, where the BS can choose the power allocation
32 Optimal Regularization Parameter and Power Allocation 3.3
matrix D, and the special case W, where it is implicitly assumed that D = I, in the following we
will denote W as the RCI precoder with equal power allocation (RCI-EP).
When the precoder Wp is used, the SINR at the kth legitimate user, given by (3.13), becomes
γk =pk|h†kwk|2∑
j 6=k pj |h†kwj |2 + σ2
, (3.29)
and the SINR at the equivalent malicious user Mk, given by (3.14), becomes
γM,k =pk‖Hkwk‖2
σ2=pk∑
j 6=k |h†jwk|2
σ2. (3.30)
From (3.29) and (3.30), we obtain the achievable secrecy sum-rate with power allocation
SpBCC =K∑k=1
[log2
(1 +
pk|h†kwk|2∑j 6=k pj |h
†kwj |2 + σ2
)− log2
(1 +
pk∑
j 6=k |h†jwk|2
σ2
)]+
. (3.31)
3.3.2 Optimal Power Allocation
To obtain the optimal power allocation vector p, we are required to solve the non-convex optimiz-
ation problem
maximizep
SpBCC(p)
subject to tr
W†pWp
≤ 1,
(3.32)
where SpBCC(p) is given by (3.31), Wp is given by (3.28), and the maximum total transmit power
over all antennas is one. In the following, we will ignore the notation [·]+ in (3.31) in the maxi-
mization problem. In fact, any negative term in the sum can be replaced by zero (thus increasing
the sum) by using pk = 0 which is always feasible.
We now reformulate the problem (3.32) by applying the transformation pk = log pk, k = 1 . . . ,K,
and obtain the optimization problem
maximizep
SpBCC(p)
subject to tr
W†pWp
≤ 1,
(3.33)
where p = [p1, . . . , pK ]T .
Lemma 3.5. The quantity l = − log2(1 + γM,k), with γM,k as in (3.30), is concave.
3.3 Physical Layer Security in Isolated Cells: Achievable Rates 33
Proof. The quantity l = − log2(1 + γM,k) and its first and second derivatives are
l = − log2 (1 + γM,k) = − log2
(1 +
epk∑
j 6=k |h†jwk|2
σ2
),
∂l
∂pk=−∂ log2 (1 + γM,k)
∂pk=−(log2 e)e
pk∑
j 6=k |h†jwk|2
σ2 + epk∑
j 6=k |h†jwk|2
,
∂2l
∂p2k
=−∂2 log2 (1 + γM,k)
∂p2k
=−(log2 e)e
pk∑
j 6=k |h†jwk|2σ2(
σ2 + epk∑
j 6=k |h†jwk|2
)2 ≤ 0.
(3.34)
Hence by the second order condition, l is concave.
In order to solve the problem (3.33), we consider a modified version of the method as in PAPAN-
DRIOPOULOS et al. [2008] and SUNG AND COLLINGS [2010] which is based on a reformulation
of (3.33). This approach guarantees an improvement in the performance over the standard high-
SNR approximation in fading channels SUNG AND COLLINGS [2010]. In order to obtain the
reformulation, we use the following bound obtained in PAPANDRIOPOULOS et al. [2008]
a log z + b ≤ log(1 + z),
a =z0
1 + z0and b = log(1 + z0)− z0
1 + z0log z0,
(3.35)
for some z0 ≥ 0, with equality when z = z0.
Lemma 3.6. With the change of variables pk = log pk, k = 1 . . . ,K, the lower bound
aklog 2
log
(epk |h†kwk|2∑
j 6=k epj |h†kwj |2 + σ2
)+
bklog 2
≤ log2
(1 +
pk|h†kwk|2∑j 6=k pj |h
†kwj |2 + σ2
),
(3.36)
is concave in pk, k = 1, . . . ,K.
Proof. The result follows immediately using the method in Lemma 3.5.
We showed in Lemma 3.5 that the second term of the sum in (3.31) is concave by the second order
condition. By using the lower bound in (3.36) for the first term of (3.31), we obtain a concave
objective function. Since the constraints are affine, the optimization problem arising from (3.33)
and the bound (3.36) is a convex optimization problem. This convex optimization problem is given
34 Optimal Regularization Parameter and Power Allocation 3.3
by
maximizep
K∑k=1
[ak
log 2log
(epk |h†kwk|2∑
j 6=k epj |h†kwj |2 + σ2
)
+bk
log 2− log2
(1 +
epk∑
j 6=k |h†jwk|2
σ2
)]subject to tr
W†
pWp
≤ 1
(3.37)
and it can be solved by bisection method.
The power allocation vector can then be obtained using Algorithm 1 in Table 3.1. To show that
Algorithm 1 converges monotonically to a local optimum, we note that the constraint is the same
for both the tth and (t + 1)th subproblems. Hence, the solution of the tth subproblem (3.37) is
also feasible for the (t + 1)th subproblem (3.37). Moreover, by the bound in (3.35), the objective
function is monotonically increasing and converges to a local optimum.
Table 3.1: Algorithm for optimal power allocation.
Algorithm 1
Initialize iteration counter t = 0
Initialize all a(t)k = 1, b(t)k = 0
repeat
Solve (3.37) to obtain p(t)
Update a(t)k , b
(t)k at z0 = γk(p
(t))
Increment t
until convergence
Obtain pk = epk , k = 1, . . . ,K
3.3.3 Joint Optimal Power Allocation and Regularization Parameter
Having established an algorithm to determine the optimal power allocation vector p for a fixed ξ,
we now consider the joint optimization of ξ and p. The joint optimization problem can be written
3.3 Physical Layer Security in Isolated Cells: Achievable Rates 35
as
maximizep,ξ
SpBCC(p, ξ)
subject to tr
W†pWp
≤ 1.
(3.38)
Even after using the transformation pk = log pk, k = 1, . . . ,K, the problem (3.38) is non-convex.
To solve this problem, we propose Algorithm 2 in Table 3.2.
At each iteration, Algorithm 2 optimizes the regularization parameter ξ and subsequently the
power allocation vector p. It is straightforward to prove that Algorithm 2 converges monoton-
ically and it thus provides a locally optimal pair (ξ,p) for the RCI precoder. We now show via
simulations that the RCI precoder with jointly optimal regularization parameter and power alloca-
tion vector outperforms RCI precoding with equal power allocation (RCI-EP).
Table 3.2: Algorithm for joint optimal power allocation and regularization parameter.
Algorithm 2
Initialize iteration counter t1 = 0, t2 = 0
Initialize pk = 1/ζ, and set pk = log pk, k = 1, . . . ,K
Initialize ξ0 = 1/ρ
repeat
Increment t1
Obtain the optimal ξ?t1 using steepest descent with ξt1−1 as initial point
Initialize all a(t2)k = 1, b(t2)
k = 0
repeat
Solve (3.37) to obtain p(t2)
Update a(t2)k , b
(t2)k at z0 = γk(ξ
?t1 , p
(t2))
Increment t2
until convergence
Set p = p(t2)
until convergence
Obtain pk = epk , k = 1, . . . ,K
36 Achievable Secrecy Rates in Practical Channels 3.4
In Fig. 3.4 we compare the simulated per-user secrecy rate of the RCI precoder with jointly op-
timized regularization parameter ξ and power allocation vector p to the RCI precoder with sep-
arately optimized ξ and p, and to the RCI precoder with equal power allocation (RCI-EP). We
observe that there is a negligible performance difference between the joint optimization and the
separate optimization. As a result, a low-complexity, near-optimal RCI precoder may be imple-
mented by optimizing the regularization parameter and the power allocation vector separately.
The figure shows that for K = 4, the RCI precoder with optimal power allocation outperforms
the RCI-EP precoder by up to 20%. Figure 3.4 also shows that optimal power allocation reduces
the sum-rate loss due to the secrecy requirements. For an SNR ρ ≥ 15dB, RCI with power al-
location achieves a per-user secrecy rate which is even higher than the per-user rate achieved by
the optimal RCI-EP without secrecy requirements. Furthermore, Fig. 3.4 shows the simulated se-
crecy capacity CMISOME of a MISOME channel with the same per-message transmitted power.
Although CMISOME is obtained in a single-user and interference-free system KHISTI AND WOR-
NELL [2010], at high SNR, RCI precoding with power allocation achieves a per-user secrecy rate
as large as CMISOME. It is not very surprising that the per-user secrecy rate of the RCI precoder
with optimal power allocation is close to the capacity of the MISOME channel, since the RCI
precoder with optimal power allocation should be close to the optimal linear precoder.
3.4 Achievable Secrecy Rates in Practical Channels
3.4.1 Secrecy Rates in the Presence of Imperfect Channel State Information
In the previous sections, we studied the secrecy rates achievable by RCI precoding for the case
when the transmitter has perfect channel state information (CSI). In this section, we consider a
more realistic scenario where only an estimation of the channel is available at the transmitter.
We model the relation between the channel H and the estimated channel H as
H = H + E (3.39)
where the matrix E represents the channel estimation error, and it is independent from H. The
entries of H and E are i.i.d. complex Gaussian random variables with zero mean and variances
1 − τ2 and τ2, respectively. The value of τ ∈ [0, 1] depends on the quality and technique used
3.4 Physical Layer Security in Isolated Cells: Achievable Rates 37
0 5 10 15 20 250
1
2
3
4
5
6
SNR, ρ [dB]
Per-usersecrecyrate
CMISOME
RCI-EP w/o secrecy
RCI w joint optim.
RCI w separate optim.
RCI-EP
Figure 3.4: Per-user secrecy rate vs. SNR ρ for K = 4 users: with separate (dashed) and joint
(circle) optimization of ξ and p, and with equal power allocation (solid). The rate of the RCI-EP
precoder without secrecy requirements (square) and the secrecy capacity of the MISOME channel
(diamond) are also plotted.
38 Achievable Secrecy Rates in Practical Channels 3.4
for channel estimation, and it is the same for all users. When τ = 0 the CSI is perfectly known,
whereas τ = 1 corresponds to the case when no CSI is available at all.
The transmitter uses the knowledge of H to obtain the RCI precoding matrix W, given by
W =1√ζ
H†(HH† +NξIK
)−1
=1√ζ
(H†H +NξIN
)−1H† (3.40)
where ζ = tr(
H†H +NξIN
)−2H†H
is the power normalization constant in the presence of
CSI error.
An achievable secrecy sum-rate in the presence of imperfect channel state information is therefore
given by
SiBCC =K∑k=1
RiBCC,k (3.41)
where
RiBCC,k =[log2
(1 + γik
)− log2
(1 + γiM,k
)]+, (3.42)
with
γik =ρ∣∣∣h†kwk
∣∣∣21 + ρ
∑j 6=k
∣∣∣h†kwj
∣∣∣2 , (3.43)
γiM,k = ρ ‖Hkwk‖2 , (3.44)
and where wk is the kth column of W.
Frequency Division Duplex Systems
In the case of frequency division duplex (FDD) systems, we can assume that users quantize their
perfectly estimated channel vectors and send the quantization index back to the transmitter over a
limited-rate channel. We assume that the channel magnitude is perfectly known to the transmitter,
since it can be efficiently quantized, and that each channel direction is quantized using B bits
and random vector quantization (RVQ) JINDAL [2006]; RYAN et al. [2008]. In RVQ, each user
independently generates a random codebook with 2B vectors, isotropically distributed on the N -
dimensional unit sphere. RVQ generates a CSI that follows the model in (3.39), where the error
3.4 Physical Layer Security in Isolated Cells: Achievable Rates 39
τ2 can be upper bounded as JINDAL [2006]
τ2 < 2−B
N−1 . (3.45)
Time Division Duplex Systems
In the case of time division duplex (TDD) systems, uplink and downlink transmissions alternate on
the same channel. The channel estimation at the transmitter is obtained from known pilot symbols
sent by the users. Let T be the channel coherence interval, i.e., the number of channel uses for
which the channel is constant. The interval T is divided into Tt uses for uplink training and T −Ttuses for the downlink transmission of data. The channel state information at the users is provided
by a training phase in the downlink. However, a minimal amount of training is sufficient for this
phase, and we can therefore neglect the overhead due to the downlink training MARZETTA et al.
[2009].
Each user transmits the same number Tt > K of orthogonal pilot symbols to the base station,
which estimates all the K channels simultaneously. The channel estimation error at the base
station depends on the number Tt as well as on the SNR ρul on the uplink channel, and it is given
by CAIRE et al. [2010]
τ2 =1
1 + Ttρul. (3.46)
3.4.2 Secrecy Rates Under No CSI or Poisoned CSI
This thesis considers the case when channel state information, either perfect or imperfect, is ob-
tained by the BS with the collaboration of users. In some cases, users might be in standby mode
without providing any CSI at all, or they might provide wrong (poisoned) CSI with the purpose
of improving their eavesdropping capability. In both cases, the secrecy performance of the system
would be affected, because without knowledge of the channel matrix, the BS would not be able to
effectively use linear precoding to control interference and information leakage. The case when
a subset of the users are in standby and do not provide CSI to the BS is discussed in GERACI
et al. [2013], whereas analyzing the achievable secrecy rates in the presence of poisoned CSI is
regarded as an interesting future research direction.
40 Achievable Secrecy Rates in Practical Channels 3.4
3.4.3 Secrecy Rates under Transmit Channel Correlation
In the previous sections, we assumed that the channels at the various users are independent and
identically distributed. In practice, channel correlation might be present due to local scatterers
around the BS and insufficient separation between antenna elements. We now consider correlation
amongst transmit antenna elements, and for tractability in the analysis, we assume a separable (or
Kronecker) correlation model MUHARAR AND EVANS [2009]; AL-NAFFOURI et al. [2009]. The
channel matrix can therefore be written as
H = HR12 (3.47)
where H ∈ CK×N contains i.i.d. circularly-symmetric complex Gaussian random variables, with
zero mean and unit variance, and where R ∈ CN×N is the non-singular transmit correlation ma-
trix, given by R = E[H†H]. We assume that each user experiences the same transmit correlation.
The channel vector for the kth user is h†k = h†kR12 ∈ CN×1.
An achievable secrecy sum-rate is therefore given by
ScBCC =K∑k=1
RcBCC,k (3.48)
with
RcBCC,k =[log2 (1 + γck)− log2
(1 + γcM,k
)]+, (3.49)
where
γck =ρA2
k
ρBk + ζ(
1 + Ak
)2 , (3.50)
γiM,k =ρBk
ζ(
1 + Ak
)2 , (3.51)
and where we have introduced the quantities NGUYEN AND EVANS [2008]
Qk =(H†kHk +NξR−1
)−1, (3.52)
Ak = h†kQkhk, (3.53)
Bk = h†kQkH†kHkQkhk, (3.54)
ζ = tr
QkHR−1
H†Qk
. (3.55)
3.5 Physical Layer Security in Isolated Cells: Achievable Rates 41
3.5 Conclusion
In this chapter, we considered the MISO broadcast channel with confidential messages, where a
multi-antenna base station simultaneously transmits independent confidential messages to several
single-antenna users in the presence of Rayleigh fading. We studied the secrecy rates achievable
by a linear precoder based on regularized channel inversion (RCI). We showed that RCI precoding
outperforms plain channel inversion precoding, and that the secrecy requirements result in a loss
in terms of the sum-rate. We then proposed an algorithm to jointly optimize the regularization
parameter of the precoder and the power allocation vector, and showed that optimal power alloca-
tion increases the secrecy sum-rate compared to equal power allocation. Finally, we extended the
secrecy rate analysis to more practical scenarios where only imperfect channel state information is
available at the transmitter, and where channel correlation is present among the transmit antenna
elements.
42
Chapter 4
Physical Layer Security in Isolated
Cells: A Large-System Analysis
Summary In this chapter, we obtain deterministic approximations for the large-system secrecy
rate achievable by RCI precoding in the MISO BCC under Rayleigh fading. We also derive the
optimal regularization parameter and the optimal network load, and simulations confirm that our
analysis is accurate even for finite systems. We find that the RCI precoder performs poorly for
large network loads and large SNR. Therefore, we propose a precoder based on RCI and power
reduction (RCI-PR) that significantly increases the performance in such regime. We compare the
secrecy rate of our proposed RCI-PR precoder to two upper bounds obtained without secrecy re-
quirements and without interference, respectively, and show that it has the same high-SNR scaling
factor as the two bounds. We further study the secrecy rates achievable under imperfect CSI, and
determine the minimum CSI quality in order to maintain a given rate gap to the case with perfect
CSI. We finally extend our large-system analysis to the case of transmit-side channel correlation.
43
44 Large-System Analysis of the Secrecy Rates 4.1
4.1 Large-System Analysis of the Secrecy Rates
I n this chapter, we consider the MISO BCC as in Chapter 3. For this system model, we study the
secrecy rates achievable by the RCI precoder in the large-system regime, where both the number
of receivers K and the number of transmit antennas N approach infinity, with their ratio β =
K/N being held constant. We then derive the optimal regularization parameter ξ that maximizes
the secrecy sum-rate and an approximation for the optimal network load β. As in Chapter 3,
we assumed homogeneous users, i.e., their distances from the serving BS are equal. Our large-
system analysis can be extended to the general case where the mobile users have different distances
from the serving BS. In this case, the large-system deterministic equivalents of the useful signal,
interference, and information leakage, are given in the form of fixed-point equations, and can be
found in YANG et al. [2013].
4.1.1 Large-System Secrecy Rates with RCI Precoding
In the following we provide a deterministic approximation of the secrecy rates, which is almost
surely exact as N → ∞. The proof of this deterministic approximation can be found in Ap-
pendix A.1, and it holds only for values of ξ such that the minimum eigenvalue of(
1NHH† + ξI
)is bounded away from zero for all large N , almost surely. It can be shown that such condition is
verified only if ξ ∈ DN , where the set DN is defined as
DN = R\[−(
1 +√β)2− C
N12−ε,−(
1−√β)2
+C
N12−ε
], (4.1)
with constants C > 0 and ε > 0.
Theorem 4.1. Let ρ > 0 and β > 0. Let RBCC,k be the secrecy rate achievable by user k with
RCI precoding defined in (3.7). Then
supξ∈DN
|RBCC,k (ξ)−RBCC (ξ)| a.s.−→ 0, as N →∞, ∀k. (4.2)
RBCC denotes the secrecy rate in the large-system regime, given by
RBCC =
[log2
1 + γ
1 + γM
]+
, if ξ 6= 0, (4.3)
and
RBCC(0) = limξ→0
RBCC(ξ) =
log2
[1 + (1−β)ρ
β
]if β ≤ 1
log2β3[β+ρ(β−1)]
[β2+ρ(β−1)2]2
+
if β > 1(4.4)
4.1 Physical Layer Security in Isolated Cells: A Large-System Analysis 45
with
γ = g (β, ξ)ρ+ ρξ
β [1 + g (β, ξ)]2
ρ+ [1 + g (β, ξ)]2, (4.5)
γM =ρ
(1 + g (β, ξ))2 , (4.6)
and
g (β, ξ)=1
2
sgn(ξ) ·√
(1−β)2
ξ2+
2 (1+β)
ξ+1+
1−βξ−1
. (4.7)
Proof. See Appendix A.1.
The secrecy sum-rate SBCC can be therefore approximated by the large-system secrecy sum-rate
SBCC, given by
SBCC = KRBCC. (4.8)
4.1.2 Secrecy Sum-Rate Maximizing Regularization Parameter
The value of the regularization parameter ξ has a significant impact on the large-system secrecy
sum-rate SBCC. In the following, we derive the regularization parameter ξBCC that maximizes
SBCC.
Theorem 4.2. Let ρ > 0, β > 0. Let ξ?BCC,N = arg maxξ∈DN
SBCC(ξ) be the optimal regularization
parameter in DN , and denote S?BCC , SBCC(ξ?BCC,N ) the optimal secrecy sum-rate. Then
1
N[S?BCC − SBCC(ξBCC)]
a.s.−→ 0, as N →∞, (4.9)
where ξBCC ∈ DN is the optimal large-system regularization parameter, given, for N large
enough, by
ξBCC =−2ρ2 (1−β)2+6ρβ+2β2−2 [β (ρ+1)−ρ] ·
√β2 [ρ2+ρ+1]−β [2ρ (ρ−1)] + ρ2
6ρ2 (β + 2) + 6ρβ.
(4.10)
46 Large-System Analysis of the Secrecy Rates 4.1
Proof. The value of ξBCC can be found by setting the derivative of SBCC to zero and studying its
maxima in each of the intervals which compose the set DN . Then we have
0(a)
≤ 1
N[S?BCC − SBCC(ξBCC)]
=1
N
[SBCC(ξ?BCC,N )− SBCC(ξ?BCC,N ) + SBCC(ξ?BCC,N )− SBCC(ξBCC)
+SBCC(ξBCC)− SBCC(ξBCC)]
(b)
≤ 1
N
[SBCC(ξ?BCC,N )− SBCC(ξ?BCC,N ) + SBCC(ξBCC)− SBCC(ξBCC)
] (c)−→ 0, (4.11)
where (a), resp. (b), follows from the definition of ξ?BCC,N , resp. ξBCC, and (c) follows from
Theorem 4.1.
When β = 1, the value of ξBCC in (4.10) reduces to
ξBCC =1
3ρ+ 1 +√
3ρ+ 1, for β = 1. (4.12)
We note that the value ξBCC that maximizes the secrecy sum-rate can be negative, and it differs
from the value ξBC = β/ρ that maximizes the sum-rate in the MISO BC without secrecy require-
ments NGUYEN et al. [2009]. Unlike ξBC, which grows unbounded as ρ → 0, ξBCC is upper
bounded by
ξBCC ≥ limρ→0
ξBCC = 1− β
2, ∀β > 0, (4.13)
although when β ≥ 2 it can be shown that SBCC = 0 irrespective of ξ and ρ. Similarly to ξBC, the
value of ξBCC decreases as the SNR increases. In the high-SNR regime, we have
limρ→∞
ξBCC − ξ∞BCC = 0, (4.14)
where ξ∞BCC approximates the high-SNR behavior of ξBCC and is given by
ξ∞BCC =
β2ρ for β < 1
13ρ for β = 1
−2(β−1)2
3(β+2) + β(2−β)2ρ(β+2) for β > 1
. (4.15)
We then have by the continuous mapping theorem
limρ→∞
S?BCC − S?∞BCC
S?BCC
= 0 (4.16)
with
S?BCC , SBCC (ξBCC) and S?∞BCC , limρ→∞
S?BCC. (4.17)
4.1 Physical Layer Security in Isolated Cells: A Large-System Analysis 47
4.1.3 Optimal Secrecy Sum-Rate
By substituting ξBCC from (4.10) in (4.5) and (4.6), and then in (4.3) and (4.8), it is possible
to obtain the optimal secrecy sum-rate S?BCC achievable by RCI precoding in the large-system
regime. The secrecy sum-rate S?BCC is a function of N , β and ρ. For β = 1, we obtain the
following compact result.
Theorem 4.3. For β = 1, the optimal secrecy sum-rate S?BCC achievable by the RCI precoder in
the large-system regime is given by
S?BCC = K log2
9ρ+ 2 + (6ρ+ 2)√
3ρ+ 1
4 (4ρ+ 1). (4.18)
Proof. Equation (4.18) follows by substituting (4.12) in (4.5) and (4.6), and then in (4.3) and
(4.8).
The secrecy sum-rate in (4.18) satisfies
S?BCC > 0 ∀ρ > 0, for β = 1, (4.19)
and S?BCC is monotonically increasing with the SNR ρ. It can be shown that the same is true for
β < 1. However when β > 1, the secrecy sum-rate does not monotonically increase with ρ. It will
be shown in Section 4.2 that there is an optimal value of the SNR beyond which the achievable
secrecy sum-rate S?BCC starts decreasing, until it becomes zero for large SNR. When β ≥ 2 no
positive secrecy sum-rate is achievable at all.
These results can be explained as follows. In the worst-case scenario, the alliance of cooperating
malicious users can cancel the interference, and its received SINR is the ratio between the signal
leakage and the thermal noise. In the limit of large SNR, the thermal noise vanishes, and the
only means for the transmitter to limit the SINR at the malicious users is by reducing the signal
leakage to zero by inverting the channel matrix. This can only be accomplished when the number
of transmit antennas is larger than or equal to the number of users, hence only if β ≤ 1. When
β > 1 this is not possible, and no positive secrecy sum-rate can be achieved. When β ≥ 2,
the eavesdroppers are able to drive the secrecy sum-rate to zero irrespective of ρ. This result is
48 Large-System Analysis of the Secrecy Rates 4.1
expected and consistent with the ones obtained for a single-user system in KHISTI AND WORNELL
[2010].
Since for β > 1 the RCI precoder performs poorly in the high-SNR regime, we have studied the
use of both user scheduling and artificial noise generation for possible secrecy enhancement. In the
presence of secrecy constraints, the strategy of reducing the network load β by simply selecting a
subset of the users to be served, commonly known as user scheduling (US), is not effective. User
scheduling would reduce the interference generated, thus increasing the SINR at the legitimate
receivers. However, US would not be effective to limit the number of malicious receivers in the
network, as the discarded users would still be able to eavesdrop. Moreover, a precoding matrix
designed by taking into account only the channels of the selected users would perform poorly,
generating large information leakage at the nearby discarded users. We have studied the use of
various US algorithms, including the optimal exhaustive search. We have omitted these results
since US does not perform well. We have also investigated the transmission of artificial noise
(AN) to limit the eavesdropping ability of the discarded users, while being harmless to selected
users. However, even combining US with AN seems not to achieve better performance than the
RCI-PR precoder proposed in this thesis, and we have omitted these results as well.
Remark 4.1. In order for Theorem 1 to hold with ξ = ξBCC, it is sufficient that ξBCC ∈ DN . Since
ξBCC in (4.10) depends on β and ρ, so does the accuracy of the deterministic approximation S?BCC
for finite N . We can distinguish the two following cases. (i) When β 6= 1, we have ξBCC ∈ DN∀ρ, and the approximation is accurate uniformly on ρ. (ii) When β = 1, ξBCC ∈ DN for all finite
ρ; if ρ→∞, then it is required that N = O(ρ2+ε), for some ε > 0, otherwise the approximation
gets weaker as ρ→∞ for N fixed. This means for instance that the approximation with N = 10
and ρ = 17dB is as accurate as the approximation with N = 40 and ρ = 20dB.
Figure 4.1 compares the secrecy sum-rate S?BCC of the RCI precoder from the large-system anal-
ysis to the simulated ergodic secrecy sum-rate SBCC with a finite number of users, for different
values of β. The value of S?BCC was obtained by (4.8) with ξBCC as in (4.10). The value of SBCC
was obtained by using the regularization parameter that maximizes the average secrecy sum-rate.
We observe that when β = 0.8 and when β = 1.2 the large-system analysis is accurate for all
values of N and SNR. When β = 1, the analysis is accurate at low SNR for all values of N , and
for high SNR larger values of N are required to increase the accuracy. The previous observations
are consistent with Remark 4.1.
4.1 Physical Layer Security in Isolated Cells: A Large-System Analysis 49
0 5 10 15 200
1
2
3
4
SNR, ρ [dB]
Per-antennasecrecysum-rate
SBCC/N (N = 10)
SBCC/N (N = 20)
SBCC/N (N = 40)
S⋆BCC/N
β = 1.2
β = 0.8
β = 1
Figure 4.1: Comparison between the secrecy sum-rate with RCI precoding in the large-system
regime (4.8) and the simulated ergodic secrecy sum-rate for finite N . Three sets of curves are
shown, each one corresponds to a different value of β.
Figure 4.2 shows that using the regularization parameter ξBCC, obtained from large-system anal-
ysis, does not cause a significant loss in the secrecy sum-rate compared to using ξ?BCC,N , opti-
mized for each channel realization. The figure shows the normalized secrecy sum-rate difference
(S?BCC − SBCC(ξBCC)) /S?BCC, simulated for finite-size systems, β = 0.8 and various values of
the SNR. The value of S?BCC was obtained by using ξ?BCC,N , whereas SBCC(ξBCC) was obtained
by using ξBCC. We observe that the average normalized secrecy sum-rate difference is less than
2% for all values of N and ρ. As a result, one can avoid the calculation of ξ?BCC,N for every
channel realization, and ξBCC can be used with only a small loss of performance.
4.1.4 Optimal Network Load
Figure 4.3 depicts the per-antenna secrecy sum-rate S?BCC/N as a function of the network load
β, for several values of the SNR. We denote by β? the value of β ∈ R+ that maximizes the
per-antenna large-system secrecy sum-rate. It is possible to see from Fig. 4.3 that the value of β?
falls between 0 and 1, and that it is an increasing function of the SNR. An approximation for β?
in the large-SNR regime is given in the following.
50 Large-System Analysis of the Secrecy Rates 4.1
5 10 15 20 25 30 35 4010
−5
10−4
10−3
10−2
10−1
Norm
alizedthroughputdifferen
ce
Number of transmit antennas, N
ρ = 0dBρ = 5dBρ = 10dBρ = 15dBρ = 20dB
Figure 4.2: Mean normalized secrecy sum-rate difference between S?BCC (obtained using the op-
timal ξ?BCC,N ) and SBCC(ξBCC) (obtained with ξBC from large-system analysis), for β = 0.8 and
various values of the SNR.
Proposition 4.1. In the limit of large SNR, the value β? of the optimal network load can be found
by solving the following fixed point equation
β? = ρ(
1− β?)e− 1
1−β? , (4.20)
and the network load β? tends to one for large SNR.
Proof. From (4.16), we have that S?∞BCC approximates S?BCC in the large-SNR regime. We then
obtain (4.20) by noticing that it must be β? ∈ [0, 1], and by setting ∂(S?∞BCC/N)/∂β = 0.
Figure 4.4 shows the optimal number of users K? obtained via simulations, for N = 10, 20, and
40 antennas. This is compared to K?, obtained from an exhaustive search on the large-system
secrecy sum-rate S?BCC, and to the closed-form approximation K? = Nβ?, obtained from (4.20)
in the high-SNR regime. We note that K? is accurate across the whole range of SNR, whereas
K? is accurate for medium-to-large values of the SNR.
4.1 Physical Layer Security in Isolated Cells: A Large-System Analysis 51
0 0.5 1 1.5 20
1
2
3
4
5
Network load, β
S⋆
BCC/N
ρ = 5dBρ = 10dBρ = 15dBρ = 20dB
0.65
0.70
0.75
0.57
Figure 4.3: Large-system secrecy sum-rate per transmit antenna as a function of β for RCI pre-
coding. The value of β? is indicated next to each curve.
0 5 10 15 20 25 300
5
10
15
20
25
30
35
40
45
50
SNR, ρ [dB]
Optimalnumber
ofusers
K⋆ (simulations)
K⋆ (exhaustive search from S⋆BCC)
K⋆ = Nβ⋆ (analysis)
N = 10
N = 20
N = 40
Figure 4.4: Comparison between K? (obtained via simulations), K? (obtained via exhaustive
search and large-system analysis), and the analytical approximation K? (obtained via large-
system and large-SNR analysis).
52 RCI Precoder with Power Reduction 4.2
4.2 RCI Precoder with Power Reduction
We have found that for β > 1 the RCI precoder performs poorly in the high-SNR regime. In this
section, we first derive the optimal value of the SNR ρ? that maximizes the achievable secrecy
sum-rate S?BCC for β > 1. We then propose a linear precoder based on RCI and power reduction
which significantly increases the high-SNR secrecy sum-rate for 1 < β < 2.
4.2.1 Optimal Transmit SNR
When 1 < β < 2, there is an optimal value of the transmit SNR ρ?, provided in the following.
Proposition 4.2. The value of the SNR ρ? that maximizes the secrecy sum-rate S?BCC for 1 <
β < 2, and the corresponding maximum value of S?BCC are respectively given by
ρ? = arg maxρ
S?BCC (ρ) =β (2− β)
(β − 1)2 (4.21)
and
S?BCC (ρ?) = K log2
β2
4 (β − 1). (4.22)
Proof. If 1 < β < 2, then ρ? is the only stationary point of S?BCC, which can be found by setting
its derivative ∂S?BCC/∂ρ to zero. We note that limρ→ρ? ξBCC = 0. Therefore, S?BCC (ρ?) can be
obtained by considering ρ→ ρ? and ξ → 0 in (4.8) and after some algebraic manipulations.
4.2.2 Power Reduction Strategy
We now propose a power reduction strategy to prevent the secrecy sum-rate from decreasing at
high SNR, for 1 < β < 2. This is achieved by reducing the transmit power, and therefore
reducing the SNR to the value ρ? that maximizes the secrecy sum-rate. We denote this scheme as
the RCI precoder with power reduction (RCI-PR), whose precoding matrix WPR is given by
WPR =
1√ζH†(HH† +NξIK
)−1if β ≤ 1
1√κζ
(H†H +NξIN
)−1H† if 1 < β < 2
0 if β ≥ 2
(4.23)
4.2 Physical Layer Security in Isolated Cells: A Large-System Analysis 53
where κ = max(
ρρ? , 1
)is the power reduction constant used for 1 < β < 2, and where ξ is
chosen from (4.10) evaluated with an SNR of min(ρ, ρ?). We note that (4.23) generalizes the
RCI precoder in (3.15) to the case when the power reduction strategy is employed.
Remark 4.2. We note from (4.10) that ξBCC(ρ?) = 0. Therefore when ρ ≥ ρ?, the optimal
value of ξ for the RCI-PR precoder is zero, and it reduces to a CI-PR precoder. Even if β > 1, it
is still possible to calculate WPR by expressing it as in (4.23) for 1 < β < 2.
We denote by RrBCC,k and RrBCC the secrecy rates achievable by the proposed RCI-PR pre-
coder (4.23) in finite systems and in the large-system regime, respectively. Similarly, SrBCC =∑Kk=1R
rBCC,k and SrBCC = KRrBCC denote the respective secrecy sum-rates. The following
theorem provides a high-SNR approximation of SrBCC.
Theorem 4.4. In the high-SNR regime, we have limρ→∞SrBCC−S
r∞BCC
SrBCC= 0, where Sr∞BCC approxi-
mates the large-system secrecy sum-rate SrBCC achieved by the RCI-PR precoder, and it is given
by
Sr∞BCC =
K log21−ββ +K log2 ρ for β < 1
K2 log2
2764 + K
2 log2 ρ for β = 1
K log2β2
4(β−1) for 1 < β < 2
0 for β ≥ 2
(4.24)
Proof. When β ≤ 1, the RCI-PR precoder reduces to the RCI precoder. Therefore, in this case
we have Sr∞BCC = S?∞BCC, with the latter defined in (4.17). The value of (4.24) for 1 < β < 2
is obtained by noting that for large SNR, RCI-PR forces ρ = ρ?, and by using Proposition 4.2.
The value for β ≥ 2 arises from the fact that no positive secrecy sum-rate is achievable in such a
condition, and the RCI-PR precoder (4.23) transmits zero power.
The abrupt change in the large-system secrecy rate as β = 1 is due to the fact that, in this case, the
minimum eigenvalue of the quantity 1NHH†, which affects the secrecy rate, becomes zero for all
large N , almost surely.
From (4.24) we can conclude that the behavior of our proposed RCI-PR precoder can be classified
into four regions. When β < 1, any secrecy sum-rate can be achieved, as long as the transmitter
54 RCI Precoder with Power Reduction 4.2
0 5 10 15 20 250
2
4
6
8
10
12
SNR, ρ [dB]
Secrecy
sum-rate
SBCC
SrBCC
β = 1.6
β = 1.2
β = 1.4
Figure 4.5: Comparison between the ergodic secrecy sum-rates SBCC and SrBCC achieved by the
RCI precoder and by the proposed RCI-PR precoder, respectively, for N = 10 transmit antennas.
Three values of β are considered: 1.2, 1.4, and 1.6, corresponding to K = 12, 14, and 16 users.
has enough power available, and the secrecy sum-rate scales linearly with the factor K. When
β = 1, the linear scaling factor reduces to K/2. When 1 < β < 2, the cooperating eavesdroppers
have more antennas than the transmitter, and thus they can limit the achievable secrecy sum-rate
regardless of how much power is available at the transmitter. When β ≥ 2, the eavesdroppers are
able to prevent secret communications, and the secrecy sum-rate is zero even if unlimited power
is available.
Figure 4.5 shows the simulated ergodic secrecy sum-rates with and without the power reduction
strategy for N = 10 transmit antennas and three values of β > 1. The figure shows that the
proposed RCI-PR precoder in (4.23) increases the secrecy sum-rate compared to the RCI precoder
in (3.15). By using the proposed power reduction strategy, it is possible to prevent the secrecy
sum-rate from decreasing at large values of the SNR ρ. For large ρ, the achieved secrecy sum-rate
equals the maximum across all values of ρ. Moreover, this is achieved by using a lower transmit
power, and the amount of power saved equals 10 log10 κ−1 dB.
4.3 Physical Layer Security in Isolated Cells: A Large-System Analysis 55
4.3 Performance Comparison
In this section, we first compare the secrecy sum-rate SrBCC achieved by the proposed RCI-PR
precoder to the sum-rate S?BC achieved by the optimized RCI precoder in the BC without se-
crecy requirements, in the large-system regime. The gap between SrBCC and S?BC represents the
secrecy loss, i.e., how much the secrecy requirements cost in terms of the achievable sum-rate.
Furthermore, we compare the per-user secrecy rate RrBCC achieved by the proposed precoder to
the secrecy capacity CMISOME of a single-user MISOME wiretap channel KHISTI AND WOR-
NELL [2010]. The gap betweenRrBCC and CMISOME represents a multi-user loss, i.e., the loss due
to the requirement of serving multiple users at the same time.
4.3.1 Secrecy Loss
The secrecy sum-rate SrBCC for the MISO BCC is obtained by using the precoder in (4.23). The
optimal sum-rate S?BC in the MISO BC without secrecy requirements is obtained by using the
precoder in (3.15), and it is given by NGUYEN et al. [2009]
S?BC = K log2 [1 + g (β, ξBC)], (4.25)
with ξBC = β/ρ. Similarly to the secrecy sum-rate, there is an optimal value for the ratio β
that maximizes the per-antenna sum-rate S?BC/N without secrecy requirements NGUYEN et al.
[2009]; HOCHWALD AND VISHWANATH [2002]; WAGNER et al. [2012]. It is easy to show that
S?BC ≥ 0 for all values of β and ρ, with equality only for ρ = 0, and that S?BC tends to zero as
β → ∞. Hence, there is no limit to the number of users per transmit antenna β that the system
can accommodate with a non-zero sum-rate. However if we impose the secrecy requirements,
the secrecy sum-rate SrBCC is zero for β ≥ 2. Therefore, introducing the secrecy requirements
will limit the number of users that can be served with a non-zero rate to two times the number of
transmit antennas.
We now compare the secrecy sum-rate SrBCC to the sum-rate S?BC in the limit of large SNR. Again
by using the regularization parameter ξBC = β/ρ we obtain limρ→∞S?BC−S
?∞BC
S?BC= 0, with
S?∞BC =
K log2
1−ββ +K log2 ρ for β < 1
K2 log2 ρ for β = 1
K log2ββ−1 for β > 1
. (4.26)
56 Performance Comparison 4.3
By comparing (4.26) to (4.24), we can draw the following conclusions regarding the large-SNR
regime. If the number of transmit antennas N is larger than the number of users K, then Sr∞BCC =
S?∞BC and the secrecy requirements do not decrease the sum-rate of the network. Therefore by us-
ing the RCI-PR precoder in (4.23), one can achieve secrecy while maintaining the same sum-rate,
i.e., there is no secrecy loss. If N = K, then the secrecy loss is 12 log2(64
27) ≈ 0.62 bits per user,
but the linear scaling factorK/2 remains unchanged. Alternatively, one can achieve secrecy while
maintaining the same sum-rate, by increasing the transmit power by a factor 64/27 ≈ 3.75dB.
If N < K < 2N , then the secrecy loss is (2 − log2 β) bits per user, but the proposed precoder
transmits a lower power, which is always upper bounded by β(2−β)
(β−1)2 . Finally if K ≥ 2N , then
the secrecy requirements force the sum-rate to zero, whereas the sum-rate S?BC remains positive,
though it also tends to zero for large β.
We note that in the BC, i.e., when there are no secrecy requirements, user scheduling can be used to
achieve a higher multiplexing gain. This is not possible in the BCC, since discarding users does not
prevent them from eavesdropping. In the presence of secrecy constraints, the strategy of reducing
the network load β by simply selecting a subset of the users to be served, commonly known as
user scheduling (US), is not effective. User scheduling would reduce the interference generated,
thus increasing the SINR at the legitimate receivers. However, it would not be effective to limit
the number of malicious receivers in the network, as the discarded users would still be able to
eavesdrop. Moreover, a precoding matrix designed by taking into account only the channels of the
selected users would perform poorly, generating large information leakage at the nearby discarded
users. We have studied the use of various US algorithms, including the optimal exhaustive search.
We have omitted these results since US does not achieve better performance than the RCI-PR
precoder proposed in this thesis.
Since discarding users to reduce the network load β is not an effective strategy in the BCC, the
case β > 1 has practical significance, as it arises whenever the number of users is larger than
the number of transmit antennas. At high SNR, e.g. 20dB and above, when large interference
and information leakage limit the secrecy performance, the proposed RCI-PR precoder achieves
reasonably good performance by controlling the transmit power (see Fig. 4.5 and Fig. 4.6). At low
SNR, e.g. 5dB and below, when noise becomes a limiting factor, the performance is still partially
affected by the value of β, and RCI precoding takes it into account and achieves a nonnegative
secrecy rate (see Fig. 4.3 and Fig. 4.5).
4.3 Physical Layer Security in Isolated Cells: A Large-System Analysis 57
4.3.2 Multi-User Loss
We now consider the multi-user loss, i.e., the loss due to the interference caused by the presence
of multiple users in the system. This is given by the gap between the secrecy rate RrBCC achieved
by the proposed RCI-PR precoder and the secrecy capacity CMISOME of the MISOME wiretap
channel, where one user is served at a time and the remaining users can eavesdrop KHISTI AND
WORNELL [2010]. We compare Rr∞BCC to CMISOME in the large-SNR regime. The former is
obtained by dividing (4.24) byK. The value ofCMISOME was obtained in KHISTI AND WORNELL
[2010], and for large SNR we have limρ→∞CMISOME−C∞MISOME
CMISOME= 0, where
C∞MISOME =
log2 ρ for β < 1
12 log2 ρ for β = 1
log21
(β−1) for 1 < β < 2
0 for β ≥ 2
. (4.27)
We remark that in CMISOME from KHISTI AND WORNELL [2010] a single-user system is con-
sidered. Therefore, only one message is transmitted to one legitimate user, and the user does not
experience any interference. By comparing (4.27) to Rr∞BCC, we can conclude that the multi-user
loss is log21−ββ and 0.62 bits per user for β < 1 and β = 1, respectively. Hence for β ≤ 1, the
proposed RCI-PR precoder achieves a secrecy rate which has the same linear scaling factor as the
secrecy capacity of a single-user system with no interference. When 1 < β < 2, the proposed pre-
coder suffers a multi-user loss of (2− 2 log2 β) bits, but again it has the advantage of transmitting
a limited power.
In Fig. 4.6 we compare the simulated per-user ergodic secrecy rate SrBCC/K of the RCI-PR pre-
coder to the rate SBC/K of the RCI precoder in the MISO BC without secrecy requirements.
These were obtained by using the regularization parameters ξBCC and ξBC, respectively. For
β < 1, the difference between SrBCC/K and SBC/K becomes negligible at large SNR, and se-
crecy can be achieved without additional costs. For β = 1, the two curves tend to have the same
slope at large SNR, but there is a residual gap between them. Therefore, secrecy can be achieved
at a lower rate. We note that in order to achieve secrecy without decreasing the rate, the required
additional power is less than 4dB at all SNRs. For 1 < β < 2, the sum-rate SBC tends to satu-
rate for large SNR, and so does the secrecy sum-rate SrBCC. In the simulations, for β = 1.2 and
ρ = 25dB, the gap is about 1.79 bits, close to 2− log2 β ≈ 1.74 bits. Moreover, we note that the
58 Performance Comparison 4.3
0 5 10 15 20 250
2
4
6
8
SNR, ρ [dB]
Per-userrate
SrBCC/K
SBC/K
CMISOME
β = 1.2
β = 0.8
β = 1
Figure 4.6: Comparison between the simulated ergodic per-user secrecy rate with RCI-PR (solid)
and the two upper bounds: (i) per-user rate without secrecy requirements (dashed) and (ii) MIS-
OME secrecy capacity (dotted), for K = 12 users. Three values of β are considered: 0.8, 1, and
1.2, corresponding to N = 15, 12, and 10 antennas.
proposed precoder saves 92% of the transmit power. The gap is smaller for smaller values of the
SNR, e.g., it reduces to about 0.72 bits when we set the transmit power to 10dB.
Figure 4.6 also shows the simulated secrecy capacity CMISOME of the MISOME wiretap channel.
For β ≤ 1, the RCI-PR precoder achieves a per-user secrecy rate which has the same linear
scaling factor as CMISOME. When 1 < β < 2, also CMISOME saturates at high SNR. In particular,
for β = 1.2 and ρ = 25dB, the gap with the RCI-PR precoder is about 1.47 ≈ 2−2 log2 β bits, but
the RCI-PR precoder saves 92% of the power. The gap is smaller for smaller values of the SNR,
e.g., it reduces to about 0.4 bits when we set the transmit power to 10dB. All these numerical
results confirm the ones obtained from the large-system analysis.
4.3.3 Comparison to Other Linear Schemes
We now compare the secrecy sum-rate achieved by the RCI precoder with ξBCC to the secrecy
sum-rates obtained from (4.8) by using: (i) ξ = 0 (CI precoder), (ii) ξ → ∞ (single user beam-
4.3 Physical Layer Security in Isolated Cells: A Large-System Analysis 59
former) and (iii) ξ = ξBC (optimum RCI precoder in the BC without secrecy requirements). The
analytical results provided in this section are obtained as special cases of the RCI precoder, for
special values of the regularization parameter ξ. Although these results can be obtained for any
value of β, in the following we restrict ourselves to β = 1, which leads to compact expressions
that allow us to easily compare the performance of the different schemes.
The aim of the CI precoder is to cancel all the interference and information leakage, therefore
yielding to a secrecy sum-rate that coincides with the sum-rate when β ≤ 1. We note that for the
CI precoder it is ξ = 0, and the precoding matrix is given by
W =1
ζ0H†(HH†
)−1, (4.28)
with ζ0 = tr(
H†H)−1
.
The secrecy sum-rate achieved by CI precoding in the large-system regime grows at most sublin-
early with N →∞. In fact,
limξ→0
limN→∞
SBCC
N= 0, β = 1. (4.29)
This result is consistent with NGUYEN et al. [2009], where it was shown that the CI precoder
performs poorly in the large-system regime when the number of antennas equals the number of
users.
Similarly, we calculate the secrecy sum-rate achieved when ξ → ∞, i.e., for the single user
beamformer (SUB). Here, the transmitter beamforms in a direction such as to maximize the signal
strength of each user, without taking into account the interference it creates and the amount of
resulting information leakage. The secrecy sum-rate achieved by SUB in the large-system regime
is zero. In fact,
limξ→∞
limN→∞
SBCC
N=
[log2
2ρ+ 1
(ρ+ 1)2
]+
= 0, β = 1. (4.30)
Clearly, SUB performs poorly compared to the optimal RCI precoder. This is due to the intended
user suffering from a large amount of interference, while the malicious users may cancel the
interference by cooperating.
Finally, we consider ξ = ξBC = 1/ρ, which is the value that maximizes the sum-rate in a MISO
BC without secrecy requirements PEEL et al. [2005]. The secrecy sum-rate SBCC achieved by
60 Imperfect CSI 4.4
RCI with ξ = ξBC in the large-system regime is given by
SBCC(ξBC) = K log2
4ρ+ 1 + (2ρ+ 1)√
4ρ+ 1
2 (4ρ+ 1), β = 1. (4.31)
We observe that RCI precoding with ξ = ξBC outperforms CI precoding and SUB in the large-
system regime, but it is suboptimal compared to the use of ξBCC. For high SNR, the per-antenna
secrecy sum-rate gain provided by using ξ = ξBCC in place of ξ = ξBC is given by
limρ→∞
SBCC(ξBCC)− SBCC(ξBC)
N= log2
3√
3
4≈ 0.38 bits. (4.32)
4.4 Imperfect CSI
In this section, we study the large-system secrecy rates achievable by RCI precoding in the pres-
ence of imperfect CSI, and we determine how the CSI error must scale with the SNR in order to
ensure a constant high-SNR rate gap to the case with perfect CSI, i.e., to maintain the same multi-
plexing gain. We derive the minimum number of feedback bits required for FDD systems, and we
study the optimum amount of channel training for TDD systems. For simplicity, in the following
we will consider the RCI precoder without power reduction strategy, although the results obtained
can be easily extended to the case of RCI-PR precoding. We also focus on the case β ≤ 1, which
is the only one that yields a non-zero high-SNR multiplexing gain.
4.4.1 Minimum Required CSI
We now provide a deterministic approximation for the secrecy rate achievable by RCI precoding
in the presence of imperfect CSI. The following result is generic, and it holds irrespective of the
channel estimation technique, therefore it holds for both FDD and TDD systems.
Theorem 4.5. Let ρ > 0, β > 0, and ξ ∈ DN . Let RiBCC,k be the secrecy rate for user k in the
presence of channel estimation error with variance τ2, defined in (3.42). Define ρ ,ρ(1−τ2)ρτ2+1
and
ξ , ξ1−τ2 . Then ∣∣RiBCC,k −RiBCC
∣∣ a.s.−→ 0, as N →∞, ∀k (4.33)
4.4 Physical Layer Security in Isolated Cells: A Large-System Analysis 61
where RiBCC is the large-system secrecy rate in the presence of imperfect CSI, given by
RiBCC =
log2
1 + g(β, ξ)ρ+ ξρ
β [1+g(β,ξ)]2
ρ+[1+g(β,ξ)]2
1 + ρ
[τ2 + 1−τ2
(1+g(β,ξ))2
]
+
. (4.34)
Proof. See Appendix A.2.
In the case of perfect CSI, i.e., τ = 0, eq. (4.34) reduces to (4.3). The large-system secrecy
sum-rate in the presence of imperfect CSI is given by
SiBCC = KRiBCC. (4.35)
We now determine how the CSI estimation error must scale with the SNR, to maintain a given rate
gap to the case with perfect CSI. In the following, we assume that the regularization parameter
ξBCC from (4.10) is used. We define the per-user gap ∆RBCC as the difference
∆RBCC , RBCC −RiBCC (4.36)
where RBCC and RiBCC are the large-system secrecy rates obtained by the RCI precoder under
perfect CSI and under CSI distortion τ2, respectively. We now derive the scaling of τ2 required to
maintain a constant secrecy rate gap for high SNR, so that the multiplexing gain is not affected.
Proposition 4.3. For β ≤ 1, b > 1, a CSI distortion τ2 = Cρ , with
C =
12
(√4b− 3− 1
)for β < 1
23
(√3b− 2− 1
)for β = 1
(4.37)
produces a high-SNR rate gap of log2 b bits.
Proof. For β ≤ 1, define
κτ , τ2ρ+τ4ρ(1 + g(β, ξBCC))2
1− τ2+τ2(1 + g(β, ξBCC))2
1− τ2(4.38)
with g(β, ξ) defined as in (4.7). We have
limρ→∞
∆RBCC =
limρ→∞
log2
[1 + β2
4ρ(1−β)2κτ
]= log2 b for β < 1
limρ→∞
log2
[1 + 1
4κτ]
= log2 b for β = 1(4.39)
62 Imperfect CSI 4.4
0 5 10 15 200
1
2
3
4
SNR, ρ [dB]
Per-antennasecrecysum-rate
SiBCC/N (N = 10)
SiBCC/N (N = 20)
SiBCC/N (N = 40)
SiBCC/N
β = 1.2
β = 1
τ = 0.1
β = 0.8
Figure 4.7: Comparison between the per-antenna secrecy sum-rate SiBCC/N with RCI precoding
in the large-system regime and the simulated ergodic secrecy sum-rate SiBCC/N , for finite N , and
in the presence of a channel estimation error τ = 0.1. Three sets of curves are shown, each one
corresponds to a different value of β.
where the logarithmic quantities arise since the rate gap ∆RBCC is given by the difference of
RBCC and RiBCC, which are both logarithmic functions.
Figure 4.7 compares the secrecy sum-rate SiBCC of the RCI precoder from the large-system anal-
ysis to the simulated ergodic secrecy sum-rate SiBCC for finite N , in the presence of a CSI error
τ = 0.1 and for different values of β. The purpose of this figure is to confirm the accuracy of our
large-system analysis, and the channel estimation error τ was chosen as a constant independent of
the SNR for the sake of simplicity. The values of SiBCC and SiBCC were obtained by (4.35) and
(3.41), respectively, with ξ = ξBCC. As expected, the accuracy of the large-system approximation
increases as N grows.
We note that it is possible to design the number of feedback bits or the training length to achieve
the desired value of τ , according to (3.45) and (3.46). However, the number of feedback bits will
impact the overhead on the uplink channel, and the training length will impact the fraction of time
used to transmit useful information.
4.4 Physical Layer Security in Isolated Cells: A Large-System Analysis 63
4.4.2 Channel Feedback in FDD Systems
For the case of FDD systems, we obtain the following result on the minimum number of feedback
bits required in order to meet a given rate gap with the case of perfect CSI.
Corollary 4.6. In order to maintain a high-SNR secrecy rate offset of log2 b bits per user in the
large-system regime, it is sufficient to scale the number of feedback bits B per user as
B ≈
N−13 ρdB − (N − 1) [log2
(√4b− 3− 1
)− 1] if β < 1
N−13 ρdB − (N − 1) [log2
√3b−2−1
3 + 1] if β = 1. (4.40)
Proof. The result follows from (3.45) and Proposition 4.3.
Figure 4.8 shows the ergodic per-user secrecy rate SiBCC/K, achieved in the presence of a channel
estimation error that scales as τ2 = Cρ , with C obtained from Proposition 4.3 for log2 b = 1 bit.
This is compared to the ergodic rate SBCC/K, achieved in the presence of perfect CSI (τ = 0),
for N = 10. The simulations show a high-SNR gap of nearly 1 bit, which confirms the claims
made in Proposition 4.3.
4.4.3 Channel Training in TDD Systems
For the case of TDD systems, part of the transmission time is used for channel training. Therefore,
although a longer training interval improves the quality of the CSI, it also reduces the interval for
data transmission, and we cannot directly apply the result in Proposition 4.3. In the following, we
study the optimal value of the training interval Tt for high SNR, in the case when both the uplink
SNR ρul and the downlink SNR ρ grow with a finite ratio rρ , ρ/ρul.
Theorem 4.7. In the large-system regime, let ρ, ρul grow large with rρ = ρ/ρul constant. Then, an
approximation of the secrecy sum-rate maximizing amount of channel training Tt can be obtained
as a solution of the equations
T 3t q + T 2
t (rρq −Krρ) + Tt(r2ρq +KrρT − 2Kr2
ρ) + 2Kr2ρT = 0, (4.41)
T 3t 4q + T 2
t (4rρq − 4Krρ) + Tt(3r2ρq + 4KrρT − 6Kr2
ρ) + 6KTr2ρ = 0, (4.42)
for β < 1 and β = 1, respectively, and with q , −SBCC log 2.
64 Imperfect CSI 4.4
0 5 10 15 200
1
2
3
4
5
6
SNR, ρ [dB]
Per-usersecrecyrate
SiBCC/K (τ2 = C/ρ)
SBCC/K (τ2 = 0)
β = 0.8
β = 1
Figure 4.8: Comparison between the ergodic per-user secrecy rates SBCC/K and SiBCC/K with
RCI-PR precoder, for N = 10, in the presence of perfect CSI and in the presence of a channel
estimation error τ2 = Cρ , with C obtained from (4.37) for log2 b = 1 bit.
Proof. By using (3.46), a large-system approximation for the secrecy sum-rate in a TDD system
is given by
SiBCC =T − TtT
K
log2
1+g(β, ξ)ρ+ ξρ
β [1+g(β,ξ)]2
ρ+[1+g(β,ξ)]2
1+ρ
[τ2 + 1−τ2
(1+g(β,ξ))2
]
+
(4.43)
with
ξ =ξ(1 + Ttρul)
Ttρuland ρ =
ρTtρulρ+ 1 + Ttρul
. (4.44)
Rewrite SiBCC = T−TtT (SBCC −K∆), where
limρ,ρul→∞
∆ =
log2
(1 +
r2ρ
T 2t
+rρTt
)for β < 1
log2
(1 +
3r2ρ
4T 2t
+rρTt
)for β = 1
. (4.45)
Then (4.41) and (4.42) can be obtained by setting ∂SiBCC/∂Tt = 0 and after further high-SNR
approximations.
Although it is difficult to formally prove that equations (4.41) and (4.42) always admit a solution
Tt ≤ T , in the following we show that a solution can be obtained in the high-SNR regime by
4.5 Physical Layer Security in Isolated Cells: A Large-System Analysis 65
0 5 10 15 20 25 300
0.1
0.2
0.3
0.4
0.5
SNR, ρ [dB]
Tt/T
Simulations
Analysis
T = 300
T = 100
T = 1000
Figure 4.9: Optimal relative amount of training Tt/T vs high-SNR approximation, for N = K =
10 and rρ = ρ/ρul = 10.
solving equations (4.41) and (4.42) numerically, and that this solution well approximates the values
obtained via simulation.
Figure 4.9 shows the simulated optimal relative amount of training Tt/T versus the downlink
SNR ρ, for a system with N = K = 10, and rρ = 10. This is compared to the high-SNR
approximations obtained from Theorem 4.7. Three sets of curves are shown, each for a different
coherence time T . The figure shows that for increasing SNR, the channel estimation becomes more
accurate, and less resources should be allocated to channel training. At high SNR, the optimal
amount of training converges to the value predicted by the analysis and provided in Theorem 4.7.
4.5 Channel Correlation
In this section, we analyze the performance of the RCI precoder under transmit-side channel cor-
relation in the large-system regime, and we determine the optimal regularization parameter. Al-
though we do not consider power reduction strategy, the results derived can be easily extended
to the RCI-PR precoder. Furthermore, we obtain the large-system secrecy rates achievable by
channel inversion precoding and single user beamforming under transmit channel correlation.
66 Channel Correlation 4.5
4.5.1 Large-System Secrecy Sum-Rates
The following theorem provides a deterministic approximationRcBCC for the per-user secrecy rates
RcBCC,k in (3.49), which is almost surely exact as N →∞. RcBCC is independent of user k.
Theorem 4.8. Let ρ > 0, β > 0, and ξ > 0. Then
∣∣RcBCC,k −RcBCC
∣∣ a.s.−→ 0, as N →∞, ∀k. (4.46)
RcBCC is the large-system secrecy rate under transmit channel correlation, given by
RcBCC =
log2
1 + ωρE22+ ξρ
β(1+ω)2E12
ρE22+(1+ω)2E12
1 + ρE22
(1+ω)2E12
+
, (4.47)
where
ω = E[
T(1 + ω)
ξ(1+ω)+βT
], Eij =E
[Ti
(ξ(1+ω)+βT)j
]. (4.48)
The expectations in (4.48) are taken over the random variable T, whose distribution function Λ(t)
is the limiting eigenvalue distribution of the correlation matrix R.
Proof. The secrecy rate in the presence of channel correlation is given in (3.49). As N,K →∞,
the quantities in (3.53), (3.54), and (3.55), respectively converge almost surely to MUHARAR AND
EVANS [2009]
A = ω, B =βE22
1− βE22, and ζ = − β
(1 + ω)2
∂ω
∂ξ, (4.49)
where ∂ω∂ξ = −(1+ω)2E12
1−βE22and ω, E12 and E22 are defined in (4.48). Then (4.46) follows from
(3.49), (4.49), by applying the continuous mapping theorem, the Markov inequality, and the Borel-
Cantelli lemma COUILLET AND DEBBAH [2011].
We note that (4.47) is a non-random function of ρ, β, and ξ. The large-system secrecy sum-rate
ScBCC in the presence of channel correlation is given by
ScBCC = KRBCC. (4.50)
Corollary 4.9. In the case of no channel correlation, RcBCC reduces to RBCC in (4.3).
4.5 Physical Layer Security in Isolated Cells: A Large-System Analysis 67
Proof. When no channel correlation is present, we have R = I and Λ(t) = δ(t − 1), which
implies ω = g(β, ξ) and E12 = E22 = [ξ(1 + g(β, ξ)) + β]−2. Therefore A = g(β, ξ),
B = βg(β,ξ)ξ(1+g(β,ξ))2+β
, and ζ = βg(β,ξ)ξ(1+g(β,ξ))2+β
, and Corollary 4.9 follows.
4.5.2 Selection of the Optimal Regularization Parameter
The value of ξ affects the large-system secrecy rate RcBCC under transmit channel correlation in
(4.47). We now study the optimal regularization parameter ξcBCC that maximizes RcBCC under
transmit-side channel correlation.
Theorem 4.10. The regularization parameter ξcBCC that maximizes the secrecy rate RcBCC under
transmit-side channel correlation can be obtained by solving the fixed-point equation
ξ = β(ω2 − 1)E12 − ρE22
2ρω(1 + ω)E12. (4.51)
Proof. We obtain (4.51) by defining MUHARAR AND EVANS [2009]
c1 = (1 + ω)2E12, (4.52)
c2 = (βE22 + ξc1)(ρE22 + c1), (4.53)
c3 =βE22 + ξ(1 + ω)2E12
ρE22 + (1 + ω)2E12, (4.54)
c4 =ρ2c3ωc5
βc2, (4.55)
c5 =2ω′(1 + ω)E12E22 + (1 + ω)2(E′12E22 − E′22E12)
ω′, (4.56)
with
E′12E22 − E′22E12 = −2β(1 + ω + ξω′)(E13E33 − E223), (4.57)
and solving
∂RcBCC
∂ξ=
ω′
2RcBCCρc2
1β(1+ ρE22
c1)2 log 2
[c4(ρξ−β)(c2
1+ρc1E22)β + ρ2c5(β + ρωc3)]
= 0.
(4.58)
68 Channel Correlation 4.5
We note that ω, E12, and E22 in (4.51) depend on ξ. It is easy to verify that the value of the
regularization parameter ξcBCC that maximizes RcBCC differs from the value ξcBC = β/ρ that max-
imizes the rates in the MISO BC without secrecy requirements MUHARAR AND EVANS [2009].
Moreover in the presence of secrecy requirements, the channel correlation affects the optimal reg-
ularization parameter.
4.5.3 Comparison to Other Linear Schemes
We now provide large-system approximations for the secrecy rates achievable by channel inversion
precoding and single user beamforming in the presence of channel correlation. These approxima-
tions were obtained from some of the results given in MUHARAR AND EVANS [2009].
Channel Inversion Precoding
As the regularization parameter ξ → 0, we have the channel inversion (CI) precoder. The aim of
the CI precoder is to cancel all the interference and information leakage. With CI precoding the
transmitted signal is
x =
1√ζ0
H†(HH†
)−1u for β ≤ 1
1√ζ0
(H†H
)−1H†u for β > 1
(4.59)
where ζ0 is the power normalization constant. It can be shown that in the large-system regime, the
secrecy rate achievable by CI precoding under transmit-side channel correlation is given by
RcBCC =
log2
(1 + ρ
κ0
)for β ≤ 1
log2
1+ ρ
ρ(β−1)+β2ζ01+
ρ(β−1)
ζ0β2
+
for β > 1, for ξ → 0 (4.60)
with κ0 = β(∫ tdΛ(t)
1+c1t)−1 and ζ0 = (β − 1)(
∫ dΛ(t)t )−1.
Single User Beamforming
As the regularization parameter ξ → ∞, we have the single user beamformer (SUB). Here, the
transmitter beamforms in a direction such as to maximize the signal strength of each user, without
4.5 Physical Layer Security in Isolated Cells: A Large-System Analysis 69
taking into account the interference it creates and the amount of resulting information leakage.
The SUB is given by x =(ζ∞
)− 12
H†u, where ζ∞ is the power normalization constant. It can be
shown that the secrecy rate achievable by SUB in the large-system regime and under transmit-side
channel correlation is
RcBCC =
log2
1 + ρE2[T]β(ρE[T2]+E[T])
1 + ρE[T2]E[T]
+
, for ξ →∞. (4.61)
We note that when β ≥ 1, SUB achieves RcBCC = 0 ∀ρ. When β < 1, there is always a value
of ρ beyond which RcBCC = 0. Such poor performance is due to the intended user suffering
from a large amount of interference, while the malicious users may cancel the interference by
cooperating.
4.5.4 Numerical Results
We now provide simulation results to confirm the accuracy of the analysis, and to show the per-
formance of RCI precoding in the presence of channel correlation. Calculating RcBCC in (4.47)
requires the limiting eigenvalue distribution of the correlation matrix R. In the following, we
consider the Toeplitz-exponential model GRAY [2005], where R has a Toeplitz structure and its
entries follow the distribution rij = ν|i−j|, governed by the correlation coefficient ν ∈ [0, 1].
It can be shown that under Toeplitz-exponential correlation, the large-system secrecy rate RcBCC
reduces to
RcBCC =
log2
1+ωρcTE+ ξρ
β(1+ω)2
ρcTE+(1+ω)2
1 + ρcTE
(1+ω)2
+
, (4.62)
with cTE = ξ(1+ω)(1+ν2)+β(1−ν2)ξ(1+ω)(1−ν2)+β(1+ν2)
.
Figure 4.10 compares the per-antenna secrecy sum-rate approximation ScBCC/N under transmit-
side correlation to the simulated ergodic per-antenna secrecy sum-rate ScBCC/N with a finite num-
ber of users, for ν = 0.5 and three values of β. The values of ScBCC and ScBCC were obtained from
(4.50) and (3.48), respectively, using ξ = ξcBCC obtained as a positive solution to (4.51). We note
that the accuracy of the analysis decreases with ρ. The loss of accuracy is due to the limitations of
the tools used to derive the deterministic approximations WAGNER et al. [2012]. One can increase
the accuracy at any given SNR ρ <∞ by increasing the dimension N .
70 Channel Correlation 4.5
0 5 10 15 200
1
2
3
4
SNR, ρ [dB]
Per-antennasecrecysum-rate
ScBCC/N (N = 10)
ScBCC/N (N = 20)
ScBCC/N (N = 30)
ScBCC/N
β = 1.2
β = 0.8
β = 1
Figure 4.10: Comparison between the simulated ergodic per-antenna secrecy sum-rate ScBCC/N
from (3.48) and the large-system approximation ScBCC/N from (4.50), for ν = 0.5 and various
values of β.
Figure 4.11 shows the relative secrecy rate loss (RBCC−RcBCC)/RBCC due to channel correlation,
with RBCC and RcBCC from (4.3) and (4.47), respectively. We observe how low-to-moderate
correlation , i.e., ν < 0.4, does not affect the secrecy rate excessively, i.e., the loss is less than
10%. However, higher correlation can significantly degrade the performance of the RCI precoder,
especially at low SNR.
Figure 4.12 shows that using the large-system regularization parameter ξcBCC from (4.51) does not
cause a significant loss in the secrecy sum-rate compared to using a finite-system regularization
parameter ξ = ξc?BCC optimized by bi-sectional search for each channel realization. The comple-
mentary cumulative distribution function (CCDF) of the normalized secrecy sum-rate difference
(RcBCC(ξc?BCC)−RcBCC(ξcBCC))/RcBCC(ξc?BCC) between using ξc?BCC and ξcBCC depends on the val-
ues of SNR and β, where large SNR and β = 1 represent the worst case. The CCDFs in Fig. 4.12
were obtained for β = 1, and ν = 0.5. Since the average normalized secrecy sum-rate difference
is 2% percent or less for all values of N , the large-system regularization parameter ξcBCC from
(4.51) may be used instead of the finite-system regularization parameter with only a small loss of
performance. Such choice avoids the computation of ξc?BCC for each channel realization.
4.6 Physical Layer Security in Isolated Cells: A Large-System Analysis 71
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
Correlation coefficient, ν
Relative
secrecyrate
loss
ρ = 0dB
ρ = 5dB
ρ = 10dB
ρ = 15dB
ρ = 20dB
Figure 4.11: Relative secrecy rate loss (RBCC − RcBCC)/RBCC as a function of the correlation
coefficient ν, for β = 0.8.
4.6 Conclusion
In this chapter, we studied the large-system performance of RCI precoding in the MISO BCC.
We obtained deterministic approximations for the achievable secrecy rates under Rayleigh fading,
and we derived expressions for the optimal regularization parameter ξ and network load β. The
analysis proved to be accurate even for finite-size systems. We found that for β > 1 the RCI
precoder performs poorly in the high-SNR regime. We therefore proposed a linear precoder based
on RCI and power reduction (RCI-PR) to increase the high-SNR performance for network loads
in the range 1 < β < 2. The proposed RCI-PR precoder was showed to achieve a secrecy rate
with the same high-SNR scaling factor as both the following upper bounds: (i) the sum-rate of the
optimal RCI precoder in the MISO BC without secrecy requirements, and (ii) the secrecy capacity
of a single-user system without interference. We then studied the performance of RCI precoding
in the presence of CSI estimation error, and determined: (i) how the CSI error must scale with
the SNR in order to ensure a constant high-SNR rate gap to the case with perfect CSI, so that
the multiplexing gain is not affected, (ii) the minimum amount of channel feedback required to
the users in an FDD system , and (iii) the optimum amount of channel training in a TDD system.
We further extended the large-system analysis to the MISO BCC under transmit-side channel
72 Conclusion 4.6
0 0.01 0.02 0.03 0.04 0.050
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Normalized throughput difference
CCDF
N = 5, ρ = 10dB, mean = 2.1 × 10−2
N = 10, ρ = 10dB, mean = 6.2 × 10−3
N = 20, ρ = 10dB, mean = 1.5 × 10−3
N = 40, ρ = 10dB, mean = 4.2 × 10−4
N = 40, ρ = 30dB, mean = 4.3 × 10−3
Figure 4.12: CCDF of the normalized secrecy rate difference between using: (i) ξcBCC obtained
from (4.51) and (ii) ξc?BCC obtained by bi-sectional search for every channel realization, for β = 1,
and ν = 0.5.
correlation. We found that low-to-moderate correlation only partially affects the secrecy rates.
However, high correlation degrades the performance of the RCI precoder, especially at low SNR.
Chapter 5
Physical Layer Security in a Random
Field of Eavesdroppers
Summary In this chapter, we introduce the MISO broadcast channel with confidential messages
and external eavesdroppers (BCCE), where a multi-antenna base station simultaneously commu-
nicates to multiple malicious users, in the presence of randomly located external eavesdroppers.
We study the performance of RCI precoding in the BCCE and provide explicit expressions for the
large-system probability of secrecy outage and mean secrecy rate with respect to the spatial distri-
bution of the nodes and to the fluctuations of their channels. Our analysis is shown to be tight via
simulations. We find that, irrespective of the collusion strategy at the external eavesdroppers, a
large number of transmit antennas drives both the probability of secrecy outage and the rate loss
due to the presence of external eavesdroppers to zero. Increasing the density of eavesdroppers by
a factor n, requires n2 as many antennas to meet a given probability of secrecy outage and a given
mean secrecy rate.
73
74 Introduction to the BCCE 5.1
5.1 Introduction to the BCCE
I n this section we introduce the MISO broadcast channel with confidential messages and ex-
ternal eavesdroppers (BCCE), where not only malicious users but also nodes external to the
network can act as eavesdroppers. This can be the case in a practical system, where external
nodes are randomly scattered in space. These nodes must be regarded as potential eavesdroppers,
otherwise the system would be vulnerable to secrecy outage. The BCCE therefore represents a
practical scenario that needs to be addressed. Similarly to the BCC, we assume homogeneous
users, i.e., each user experiences the same average received signal power. This models a scenario
where users are located at equal distances from the BS, or where the BS employs power control
to guarantee the same average received signal power. Without loss of generality, such power can
be assumed unitary. Although the results presented in this chapter can be easily extended to the
non-homogeneous case, the assumption of homogeneous users allows analytical tractability, and
it yields to closed form expressions for the SINRs at the legitimate user and at the malicious users
in the large-system regime.
The BCCE can be obtained from the BCC by including external single-antenna eavesdroppers in
the system. For the sake of tractability, in this thesis the external eavesdroppers are assumed to be
distributed on the two-dimensional plane according to a Poisson point process (PPP) Φe of density
λe STOYAN et al. [1996]. Figure 5.1 shows an example of BCCE, where the BS is at the origin,
and the users lie on a disc of radius 1. As a worst-case scenario, we assume that each eavesdropper
can cancel the interference caused by the remaining K − 1 messages. Assuming that the BS lies
at the origin, the SINR γe,k for the kth message at a generic eavesdropper located in e is then given
by
γe,k =
∣∣∣h†ewk
∣∣∣2‖e‖ησ2
(5.1)
where wk is the precoding vector for user k, h†e is the channel vector between the base station
and the eavesdropper in e, and it takes into account the Rayleigh fading, and η is the path loss
exponent. Some of the results provided in this chapter assume a path loss exponent η = 4. In this
special case, which is a reasonable value for η in a shadowed urban area RAPPAPORT [1996], it is
possible to obtain more compact expressions for quantities of interest, such as the probability of
secrecy outage and the mean secrecy rate.
5.1 Physical Layer Security in a Random Field of Eavesdroppers 75
−5 0 5−5
−4
−3
−2
−1
0
1
2
3
4
5
y
x
Multi-antenna base station
Malicious users
External eavesdroppers
Figure 5.1: Example of a BCCE with K = 5 malicious users and a density of external eavesdrop-
pers λe = 0.2.
The precoding vector wk is calculated independently of h†e, therefore they are independent isotropic
random vectors. The channel h†e has unit norm, whereas the precoding vector wk has norm 1√K
be-
cause it is obtained after the normalization ‖W‖2 =∑K
k=1 ‖wk‖2 = 1. The inner product h†ewk
is a linear combination of N complex normal random variables, therefore∣∣∣h†ewk
∣∣∣2 ∼ exp( 1K ).
In the following, we consider two types of external eavesdroppers, namely non-colluding eaves-
droppers and colluding eavesdroppers. In the non-colluding case, the eavesdroppers individually
overhear the communication without centralized processing. In the colluding eavesdroppers case,
all eavesdroppers are able to jointly process their received messages at a central data processing
unit. The secrecy rate RBCCE,k achievable by the kth user in the BCCE is given by
RBCCE,k =[log2
(1 + γk
)− log2
(1 + max (γM,k, γE,k)
)]+, (5.2)
where γE,k is the resulting SINR of the PPP of external eavesdroppers for the kth message. The
secrecy rate RBCCE,k is therefore affected by the maximum of the SINR γM,k at the alliance of
malicious users and the SINR γE,k at the external eavesdroppers. In the case of non-colluding
eavesdroppers, γE,k is the SINR at the strongest eavesdropper. In the case of colluding eavesdrop-
pers, all eavesdroppers can perform joint processing, and they can, therefore, be seen as a single
multi-antenna eavesdropper. After interference cancellation, each eavesdropper receives the use-
76 Probability of Secrecy Outage 5.2
ful signal embedded in noise, and the optimal receive strategy at the colluding eavesdroppers is
maximal ratio combining (MRC) which yields to an SINR γE,k =∑
e∈Φeγe,k given by the sum
of the SINRs γe,k at all eavesdroppers generated by the Poisson point process (PPP) Φe.
The achievable secrecy sum-rate is denoted by SBCCE and defined as
SBCCE =K∑k=1
RBCCE,k. (5.3)
5.2 Probability of Secrecy Outage
In this section, we derive the secrecy outage probability, i.e., the probability that the secrecy rate
RBCCE,k achievable by user k with RCI precoding in the BCCE is zero, for both cases of non-
colluding and colluding eavesdroppers. Then we study the secrecy outage probability in the large-
system regime, and determine how the number of antennas N must scale in order to guarantee a
given secrecy outage probability. The secrecy outage probability for user k is defined as
OBCCE,k , P(RBCCE,k = 0) =
1 if γk ≤ γM,k
P(γE,k ≥ γk | γk) otherwise(5.4)
As discussed in Chapter 4, RCI precoding ensures γk > γM,k in most cases. Therefore, the secrecy
outage probability is often given by the probability that RBCCE,k is driven to zero by the presence
of external eavesdroppers.
The results provided in this chapter can be extended to obtain the probability that the achievable
secrecy rate RBCCE,k is smaller than a target rate RT . In this case, the probability of outage
would be a function of RT . We note that for Gaussian fading channels, there is always a non-
zero probability that the legitimate user’s channel is better than that of the eavesdroppers, but
this probability is smaller than one, hence there is always a non-zero secrecy outage probability.
Moreover, we note that the analysis provided in this section leads to a secrecy outage probability
which is a function of the SINR at the legitimate user and at the malicious users. These SINRs can
be calculated at the BS since the channels at the users are known and deterministic.
5.2 Physical Layer Security in a Random Field of Eavesdroppers 77
5.2.1 Non-colluding Eavesdroppers
In the case of non-colluding eavesdroppers, γE,k is the SINR at the strongest eavesdropper E,
given by
γE,k = maxe∈Φe
γe,k = maxe∈Φe
∣∣∣h†ewk
∣∣∣2‖e‖ησ2
. (5.5)
In the case of non-colluding eavesdroppers, OBCCE,k is the probability that any eavesdropper has
an SINR greater than or equal to the SINR of the legitimate user k. We obtain the following result.
Lemma 5.1. The secrecy outage probability for user k in the presence of non-colluding eaves-
droppers is given by
OBCCE,k =
1 if γk ≤ γM,k
1− exp
[−
2πλeΓ(
2η
)η(Nβσ2γk)
2η
]otherwise
(5.6)
where Γ(·) is the gamma function defined as
Γ(z) ,∫ ∞
0tz−1e−tdt. (5.7)
Proof. See Appendix B.1.
By applying results from random matrix theory (RMT) COUILLET AND DEBBAH [2011], we
now obtain the large-system secrecy outage probability OBCCE in the presence of non-colluding
eavesdroppers. Like in Chapter 4, the large-system results are obtained for the case when both
the number of transmit antennas N and the number of users K grow to infinity in a fixed ratio
β = K/N . In order to simplify the notation in the rest of this chapter, we find it useful to define
the quantity µ , π32
2√βσ2
.
Theorem 5.2. The secrecy outage probability in the presence of non-colluding eavesdroppers
satisfies
|OBCCE,k −OBCCE|a.s.−→ 0, as N →∞, ∀k (5.8)
where
OBCCE =
1 if γ ≤ γM
1− exp
[−
2πλeΓ(
2η
)η(Nβσ2γ)
2η
]otherwise
(5.9)
and with γ and γM given by (4.5) and (4.6), respectively.
78 Probability of Secrecy Outage 5.2
Proof. Theorem 5.2 follows from Lemma 5.1, by noting that |γk−γ| a.s.−→ 0 and |γM,k−γM |a.s.−→
0 as N →∞, and by the continuous mapping theorem BILLINGSLEY [1995].
Corollary 5.3. If γ > γM and η = 4, then (i) the number of transmit antennas required in
order to guarantee a large-system secrecy outage probability OBCCE < ε in the presence of non-
colluding eavesdroppers is N >(µλeε√γ
)2, and (ii) the large-system secrecy outage probability
OBCCE decays as 1√N
.
Proof. The proof follows from Theorem 5.2, by noting that Γ(
12
)=√π, and that 1 − e−x > x
for 0 < x < 1.
A special case of the previous scenario is the one where only the eavesdropper which is nearest to
the base station attempts to eavesdrop. In this case we have
γE,k =
∣∣∣h†Ewk
∣∣∣2‖E‖ησ2
(5.10)
where
E = argmine∈Φe
‖e‖. (5.11)
Lemma 5.4. The secrecy outage probability for user k, caused by the external eavesdropper
nearest to the base station, under a path loss exponent η = 4, is given by
OBCCE,k =
1 if γk ≤ γM,k
2µλe√Nγk
exp(µ2λ2
eπNγk
)Q(µλe
√2
πNγk
)otherwise
(5.12)
where Q(·) is the Q-function defined as
Q(x) ,1√2π
∫ ∞x
exp
(−u
2
2
)du. (5.13)
Proof. See Appendix B.2.
By applying results from RMT, we now obtain the large-system secrecy outage probabilityOBCCE
caused by the eavesdropper which is nearest to the base station.
Theorem 5.5. The secrecy outage probability for user k, caused by the external eavesdropper
nearest to the base station, under a path loss exponent η = 4, satisfies
|OBCCE,k −OBCCE|a.s.−→ 0, as N →∞, ∀k (5.14)
5.2 Physical Layer Security in a Random Field of Eavesdroppers 79
where
OBCCE =
1 if γ ≤ γMµλe√N
(1 + µ2λ2
eπN
)(1− 2µλe
π√N
)otherwise
(5.15)
and with γ and γM given by (4.5) and (4.6), respectively.
Proof. Theorem 5.5 follows from Lemma 5.4, by first-order Taylor approximation of (5.12), by
noting that |γk − γ| a.s.−→ 0 and |γM,k − γM |a.s.−→ 0 as N → ∞, and by the continuous mapping
theorem BILLINGSLEY [1995].
5.2.2 Colluding Eavesdroppers
The colluding eavesdroppers case represents a worst-case scenario. In this case, all eavesdroppers
can perform joint processing, and they can therefore be seen as a single multi-antenna eavesdrop-
per. After interference cancellation, each eavesdropper receives the useful signal embedded in
noise, and the optimal receive strategy at the colluding eavesdroppers is maximal ratio combining
(MRC). This yields to an SINR γE,k at the colluding eavesdroppers given by
γE,k =1
σ2
∑e∈Φe
‖e‖−η∣∣∣h†ewk
∣∣∣2 . (5.16)
Lemma 5.6. The secrecy outage probability for user k in the presence of colluding eavesdroppers,
under a path loss exponent η = 4, is given by
OBCCE,k =
1 if γk ≤ γM,k
1− 2Q(µλe
√π
2Nγk
)otherwise
(5.17)
Proof. See Appendix B.3.
By applying results from RMT, we now obtain the large-system secrecy outage probabilityOBCCE
in the presence of colluding eavesdroppers.
Theorem 5.7. The secrecy outage probability in the presence of colluding eavesdroppers, under
a path loss exponent η = 4, satisfies
|OBCCE,k −OBCCE|a.s.−→ 0, as N →∞, ∀k (5.18)
80 Mean Secrecy Rates 5.3
where
OBCCE =
1 if γ ≤ γM1− 2Q
(µλe
√π
2Nγ
)otherwise
(5.19)
and with γ and γM given by (4.5) and (4.6), respectively.
Proof. Theorem 5.7 follows from Lemma 5.6, by noting that Γ(
12
)=√π, that |γk − γ| a.s.−→ 0
and |γM,k − γM |a.s.−→ 0 as N → ∞, and by the continuous mapping theorem BILLINGSLEY
[1995].
Corollary 5.8. Let γ > γM and η = 4, then (i) the number of transmit antennas required in order
to guarantee a large-system secrecy outage probability OBCCE < ε in the presence of colluding
eavesdroppers is N >(µλeε√γ
)2, and (ii) the large-system outage probability OBCCE decays as
1√N
.
Proof. The proof follows from Theorem 5.7 and by using 1− 2Q(x) <√
2πx for 0 < x < 1.
Remark 5.1. By comparing the results in Corollary 5.3 and Corollary 5.8, we can conclude
that (i) the collusion among eavesdroppers does not significantly affect the number of transmit
antennas N required to meet a given probability of secrecy outage in the large-system regime, and
(ii) increasing the density of eavesdroppers λe by a factor n requires increasing N by a factor n2
in order to meet a given probability of secrecy outage.
5.3 Mean Secrecy Rates
In this section, we derive the mean secrecy rates, averaged over the location of the external eaves-
droppers, achievable by RCI precoding in the BCCE, for both cases of non-colluding and collud-
ing eavesdroppers. We then study the mean secrecy rates in the large-system regime, and derive a
bound on the secrecy rate loss due to the presence of external eavesdroppers. Finally, we propose
a rule for the choice of the regularization parameter of the precoder that maximizes the mean of
the large-system secrecy rate.
5.3 Physical Layer Security in a Random Field of Eavesdroppers 81
5.3.1 Mean Secrecy Rate
We now obtain the following result for the mean secrecy rate at user k.
Lemma 5.9. The mean secrecy rate achievable at user k by RCI precoding in the BCCE is given
by
EΦe [RBCCE,k] =
0 if γk ≤ γM,k
log2(1+γk)
1−OBCCE,k
(1+γM,k)1−PBCCE,k
−∫ γkγM,k
log2(1 + y)fγE,k(y) dy otherwise
(5.20)
In (5.20), PBCCE,k is the probability that the SINR γE,k at the external eavesdroppers is greater
than or equal to the SINR γM,k at the malicious users, and for a path loss exponent η = 4 it is
given by
PBCCE,k , P(γE,k ≥ γM,k) =
1− exp
(− µλe√
NγM,k
)for non-colluding eavesdroppers
1− 2Q
(µλe
√π
2NγM,k
)for colluding eavesdroppers
(5.21)
and fγE,k(y) is the pdf of the SINR at the external eavesdroppers, given by
fγE,k(y) =
µλey
− 32
2√N
exp(− µλe√
Ny
)for non-colluding eavesdroppers
µλey− 3
2
2√N
exp(−πµ2λ2
e4Ny
)for colluding eavesdroppers
(5.22)
Proof. See Appendix B.4.
By applying results from RMT, we now obtain the large-system mean secrecy rate RBCCE achiev-
able by RCI precoding in the BCCE.
Theorem 5.10. The mean secrecy rate achievable for user k by RCI precoding in the BCCE
satisfies
|EΦe [RBCCE,k]−RBCCE|a.s.−→ 0, as N →∞, ∀k. (5.23)
RBCCE denotes the mean secrecy rate in the large-system regime, given by
RBCCE =
0 if γ ≤ γM
log2(1+γ)1−OBCCE
(1+γM)1−P
BCCE−∫ γγM
log2(1 + y)fγE,k(y) dy otherwise(5.24)
82 Mean Secrecy Rates 5.3
In (5.24), PBCCE is the probability that the SINR γE,k at the external eavesdroppers is greater
than or equal to the large-system SINR γM at the malicious users, and for η = 4 it is given by
PBCCE , P(γE,k ≥ γM ) =
1− exp
(− µλe√
NγM
)for non-colluding eavesdroppers
1− 2Q
(µλe
√π
2NγM
)for colluding eavesdroppers
(5.25)
Proof. Theorem 5.10 follows from Lemma 5.9, by replacing γk and γM,k with their respective de-
terministic approximations γ and γM , by applying the continuous mapping theorem, the Markov
inequality, and the Borel-Cantelli lemma BILLINGSLEY [1995].
5.3.2 Secrecy Rate Loss due to the External Eavesdroppers
By comparing the large-system mean secrecy rate of the MISO BCCE in (5.24) to the large-system
secrecy rate of the MISO BCC without external eavesdroppers in (4.3), for a given regularization
parameter ξ, we can evaluate the secrecy rate loss ∆e due to the presence of external eavesdrop-
pers, defined as
∆e , RBCC −RBCCE. (5.26)
We now obtain an upper bound on the secrecy rate loss ∆e.
Corollary 5.11. The secrecy rate loss ∆e due to the presence of external eavesdroppers satisfies
∆e ≤ ∆UBe ,
Cµλe√N, (5.27)
whereCµ is a constant independent ofN , λe, and of the cooperation strategy at the eavesdroppers,
given by
Cµ = µ
[RBCC√γ
+(√
γ −√γM
)+]. (5.28)
Proof. See Appendix B.5
Remark 5.2. It follows from Corollary 5.11 that, irrespective of the collusion strategy at the
external eavesdroppers, (i) as the number N of transmit antennas grows, the secrecy rate loss ∆e
5.3 Physical Layer Security in a Random Field of Eavesdroppers 83
tends to zero as 1√N
, and (ii) increasing the density of eavesdroppers λe by a factor n requires
increasing N by a factor n2 in order to meet a given value of ∆UBe .
5.3.3 Optimal Regularization Parameter
The value of the regularization parameter ξ has a significant impact on the secrecy rates. The
optimal large-system regularization parameters of the RCI precoder ξBC and ξBCC, for the MISO
BC and for the MISO BCC, respectively, were given in Chapter 4. In the MISO broadcast channel
with confidential messages and external eavesdroppers (BCCE), we denote by ξBCCE the regular-
ization parameter that maximizes the large-system mean secrecy rate. The value of ξBCCE can be
obtained by numerically solving the following equation
ξBCCE , arg maxξ
RBCCE (5.29)
with RBCCE given in (5.24). Since the secrecy rate of the MISO BCCE is affected by the SINR at
the external eavesdroppers, the optimal large-system regularization parameter ξBCCE is not just a
function of β and ρ, like ξBC and ξBCC, but it also depends on the number of transmit antennas N ,
the density of the eavesdroppers λe, and their collusion strategy. The value of ξBCCE should be
found as a compromise between: (i) maximizing the SINR γ at the legitimate user, and (ii) trading
off the SINR γM at the malicious users and the probability PBCCE that the external eavesdroppers
are more harmful than the malicious users. We have the following two extreme cases.
Lemma 5.12. The optimal large-system regularization parameter ξBCCE follows the trend:
ξBCCE → ξBCC as λe → 0
ξBCCE → ξBC as λe →∞(5.30)
Proof. For low densities λe, we have by Corollary 5.11 that RBCCE approaches RBCC, therefore
ξBCCE approaches ξBCC. For high densities λe, we have PBCCE,k = P(γE,k ≥ γM,k) → 1, and
the secrecy rateRBCCE,k in (5.2) is determined solely by γk and γE,k. Since γE,k does not depend
on ξ, maximizing the mean rate coincides with the rate maximization problem for the BC, and its
solution in the large-system regime is given by ξBC.
84 Numerical Results 5.4
5.4 Numerical Results
In this section, we provide numerical results to show the performance of RCI precoding in the
BCCE, under a path loss exponent η = 4. We consider finite-size systems, and simulate the
probability of secrecy outage, the secrecy rate, and the optimal regularization parameter of the
precoder, in different scenarios and under different system dimensions, network loads, SNRs, and
densities of eavesdroppers. The simulations show that many results obtained in Section 5.2 and
Section 5.3 by using random matrix theory and stochastic geometry tools hold even for networks
with a small number of users and transmit antennas and randomly located eavesdroppers.
In Fig. 5.2 we compare the simulated probability of outageOBCCE,k under non-colluding and col-
luding eavesdroppers, respectively, to the large-system results OBCCE provided in Theorem 5.2
and Theorem 5.7, respectively. In the simulations, the regularization parameter ξBCC in (4.10)
was used. We observe that for λe = 0.1 and small probabilities of secrecy outage, (i) N >(µλe
0.1√γ
)2= 34 yields to a secrecy outage probability smaller than 0.1, (ii) the secrecy outage
probability decays as 1√N
, and (iii) the collusion of eavesdroppers does not significantly affect
the probability of secrecy outage. All these observations are consistent with Corollary 5.3, Corol-
lary 5.8, and Remark 5.1.
In Fig. 5.3 we compare the simulated ergodic per-antenna secrecy sum-rate under non-colluding
and colluding eavesdroppers, to the large-system results from Theorem 5.10, for λe = 0.1, N =
10, ξ = ξBCC, and various values of β. We note that the accuracy of the large-system analysis
decreases with the SNR. The loss of accuracy is due to the limitations of the tools used from
RMT WAGNER et al. [2012]. Moreover, we note that the per-antenna secrecy sum-rate does not
monotonically increase with the SNR. This is due to the fact that in the worst-case scenario the
malicious users and the external eavesdroppers can cancel the interference, whereas the legitimate
user is interference-limited in the high-SNR regime. This is consistent with the case of β > 1 in
the BCC, studied in Chapter 4.
In Fig. 5.4 we compare the simulated ergodic per-user secrecy rate under non-colluding and col-
luding eavesdroppers, to the large-system results from Theorem 5.10, for β = 1, ρ = 10dB,
ξ = ξBCC, and various values of λe. We note that the accuracy of the large-system analysis in-
creases with N . Moreover, we observe that the expectation of the per-user secrecy rate increases
with N , and this benefit is more for larger values of λe. This happens because the mean received
5.4 Physical Layer Security in a Random Field of Eavesdroppers 85
5 10 15 20 25 30 35 400
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Number of transmit antennas, N
Probabilityofsecrecyoutage
Ok non-coll. (Simulations)O non-coll. (Analysis)Ok colluding (Simulations)O colluding (Analysis)
λe = 0.1, 0.2, 0.4
Figure 5.2: Comparison between the simulated probability of outage OBCCE,k and the large-
system results OBCCE provided in Theorem 5.2 and Theorem 5.7, for a network load β = 1, an
SNR ρ = 10dB, and various values of λe.
0 5 10 15 200
0.5
1
1.5
2
2.5
SNR, ρ [dB]
Per-antennasecrecysum-rate
Simulations (non-coll.)Analysis (non-coll.)Simulations (colluding)Analysis (colluding)
β = 1.2
β = 0.8
β = 1
Figure 5.3: Comparison between the simulated ergodic per-antenna secrecy sum-rate under non-
colluding and colluding eavesdroppers, and the large-system results from Theorem 5.10, for λe =
0.1, N = 10 transmit antennas, and various values of the network load β.
86 Numerical Results 5.4
5 10 15 20 25 30 35 400
0.5
1
1.5
2
Number of transmit antennas, N
Per-usersecrecyrate
Simulations (non-coll.)Analysis (non-coll.)Simulations (colluding)Analysis (colluding)
λe = 0.1, 0.2, 0.4
Figure 5.4: Comparison between the simulated ergodic per-user secrecy rate E[Rk] under non-
colluding and colluding eavesdroppers, and the large-system results R from Theorem 5.10, for a
network load β = 1, an SNR ρ = 10dB, and various values of λe.
power at each external eavesdropper scales as 1βN , hence having more transmit antennas makes
the system more robust against external eavesdroppers.
In Fig. 5.5 we compare the simulated per-user secrecy rate of (i) the BCCE with non-colluding
eavesdroppers, (ii) the BCCE with colluding eavesdroppers, and (iii) the BCC without external
eavesdroppers, for β = 1, ρ = 10dB, ξ = ξBCC, and various values of λe. We note that in the
BCC, the per-user secrecy rate is almost constant with N , for a fixed network load β. On the
other hand, the per-user secrecy rate of the BCCE increases with N . Again, this happens because
the mean received power at each external eavesdropper scales as 1βN , hence having more transmit
antennas makes the system more robust against external eavesdroppers. We also note that for
higher densities of eavesdroppers λe, larger values of N are required to achieve a given per-user
secrecy rate of the BCCE. More precisely, increasing λe by a factor 2, requires increasing N by a
factor 4. Moreover, the collusion of external eavesdroppers does not affect the scaling law of the
mean rate. These observations are consistent with Remark 5.2.
Figure 5.6 compares the large-system regularization parameter ξBCCE given by (5.29) to the value
ξBCCE that maximizes the average simulated secrecy sum-rate SBCCE, for a finite system with
5.4 Physical Layer Security in a Random Field of Eavesdroppers 87
5 10 15 20 25 30 35 400
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
Number of transmit antennas, N
Per-usersecrecyrate
RBCC,k (λe = 0)
RBCCE,k (non-coll.)
RBCCE,k (colluding)
λe = 0.1, 0.2, 0.4
Figure 5.5: Comparison between the simulated ergodic per-user secrecy rates of: (i) the BCCE
with non-colluding eavesdroppers, (ii) the BCCE with colluding eavesdroppers, and (iii) the BCC
without external eavesdroppers, for a network load β = 1, an SNR ρ = 10dB, and various values
of λe.
88 Numerical Results 5.4
10−2
10−1
100
101
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1
0.11
Density of external eavesdroppers, λe
Optimalregularizationparameter
ξBCCE (non-coll.)ξBCCE (non-coll.)
ξBCCE (colluding)ξBCCE (colluding)
ξBCC
ξBC
Figure 5.6: Comparison between the large-system regularization parameter ξBCCE in (5.29) and
the value ξBCCE that maximizes the average simulated secrecy sum-rate SBCCE for a finite system
with N = 10 transmit antennas, a network load β = 1, and an SNR ρ = 10dB.
N = 10, β = 1, and ρ = 10dB. The figure shows that for low densities of eavesdroppers λe,
ξBCCE tends to ξBCC = 0.0273, whereas for high densities λe, it tends to ξBC = 0.1. These ob-
servations are consistent with Lemma 5.12. The finite-system parameter ξBCCE follows a similar
trend. We note that both ξBCCE and ξBCCE are smaller in the case of non-colluding eavesdrop-
pers, and this can be explained as follows. A smaller value of ξ generates a smaller information
leakage to the malicious users. Therefore, it is especially desirable to have a smaller ξ when the
malicious users are the main concern, i.e., when their SINR is larger than the SINR at the ex-
ternal eavesdroppers, and this is more likely to happen when the external eavesdroppers are not
colluding.
Figure 5.7 shows that using the regularization parameter ξBCCE, obtained from large-system anal-
ysis, does not cause a significant loss compared to using the optimal parameter ξ?BCCE, optimized
for each realization of the channels and of the locations of the external eavesdroppers. The fig-
ure shows the mean secrecy sum-rate difference SBCCE(ξ?BCCE) − SBCCE(ξBCCE) normalized
by the mean optimal SBCCE(ξ?BCCE), simulated for finite-size systems, β = 1, various values of
the density of eavesdroppers λe, and various values of the SNR ρ. Figure 5.7 was obtained for
colluding eavesdroppers, but a similar trend was observed for non-colluding eavesdroppers. We
5.5 Physical Layer Security in a Random Field of Eavesdroppers 89
5 10 15 20 25 30 35 400
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
Number of transmit antennas, N
Norm
alizedthroughputdifferen
ce
λe = 0.1, ρ = 0dBλe = 0.4, ρ = 0dBλe = 0.1, ρ = 20dBλe = 0.4, ρ = 20dB
Figure 5.7: Normalized mean secrecy sum-rate difference between using ξ?BCCE, that maximizes
each realization of the secrecy sum-rate SBCCE, and ξBCCE, obtained from large-system analysis
in (5.29), under colluding eavesdroppers, for a network load β = 1, various values of the density
of eavesdroppers λe, and various values of the SNR ρ.
note that calculating the optimal value ξ?BCCE requires the base station to know (i) the channels H
of all users, (ii) the realization of the PPP Φe, i.e., the locations of all external eavesdroppers, and
(iii) the channels h†e of all external eavesdroppers. On the other hand, calculating ξBCCE does not
require the knowledge of any of these quantities. We observe that the normalized mean secrecy
sum-rate difference is less than 7% for all values of N , λe, and ρ, and it decreases when N grows,
e.g., falling under 3% for N = 20. As a result, one can avoid the calculation of ξ?BCCE for every
realization of H, Φe, and h†e, and ξBCCE can be used with only a small loss of performance.
5.5 Conclusion
In this chapter, we considered the broadcast channel with confidential messages and external
eavesdroppers (BCCE), where a multi-antenna base station simultaneously communicates to mul-
tiple malicious users, in the presence of randomly located external eavesdroppers. We showed that,
irrespective of the collusion strategy at the external eavesdroppers, a large number of transmit an-
tennas N drives both the probability of secrecy outage and the rate loss due to the presence of
90 Conclusion 5.5
external eavesdroppers to zero. Increasing the density of eavesdroppers λe by a factor n, requires
n2 as many antennas to meet a given probability of secrecy outage and a given mean secrecy rate.
Using the developed analysis, we clearly established the importance of the number of transmit
antennas at the BS to make communications robust against malicious users and external eaves-
dropping nodes.
Chapter 6
Physical Layer Security in Cellular
Networks
Summary In this chapter, we study physical layer security for the downlink of cellular net-
works, where each BS simultaneously transmits confidential messages to several users, and where
the confidential messages transmitted to each user can be eavesdropped by both (i) other users in
the same cell and (ii) users in other cells. The locations of BSs and mobile users are modeled as
two independent two-dimensional Poisson point processes. Using the proposed model, we analyze
the secrecy rates achievable by RCI precoding by performing a large-system analysis that com-
bines results from stochastic geometry and random matrix theory. We obtain approximations for
the probability of secrecy outage and the mean secrecy rate, and characterize regimes where RCI
precoding achieves a non-zero secrecy rate. We find that unlike isolated cells, the secrecy rate in
a cellular network does not grow monotonically with the transmit power, and the network tends to
be in secrecy outage if the transmit power grows unbounded. Furthermore, we show that there is
an optimal value for the BS deployment density that maximizes the secrecy rate, and this value is
a decreasing function of the SNR.
91
92 System Model 6.1
Figure 6.1: Illustration of a cellular network. The star denotes a typical user. The circles, squares,
and triangles denote BSs, out-of-cell users, and in-cell users, respectively, as discussed in Subsec-
tion 6.1.3.
6.1 System Model
6.1.1 Network Topology
W e consider the downlink of a cellular network, as depicted in Fig. 6.1. Each BS transmits at
power P and is equipped with N antennas. For tractability, we assume that the locations
of the BSs are drawn from a homogeneous PPP Φb of density λb. We consider single-antenna
users, and assume that each user is connected to the closest BS. The locations of the users are
drawn from an independent PPP Φu of density λu. We denote by Kb and by Kb = |Kb| the set of
users and the number of users connected to the BS b, respectively. We denote by
Hb =[‖b− b1‖−ηhb,1, . . . , ‖b− bKb‖−ηhb,Kb
]† (6.1)
the Kb×N channel matrix for the BS b, where hb,j ∼ CN (0, I) is the normalized channel vector
that accounts only for the fading between the BS b and the user j ∈ Kb.
6.1 Physical Layer Security in Cellular Networks 93
6.1.2 RCI Precoding
Transmission takes place over a block fading channel. The signal transmitted by the generic BS
b is xb = [xb,1, . . . , xb,N ]T ∈ CN×1. The vector xb is obtained via RCI precoding from the vec-
tor of confidential messages mb = [mb,1, . . . ,mb,Kb ]T , whose entries are chosen independently,
satisfying E[|mb,j |2] = 1, ∀j. The transmitted signal xb after RCI precoding can be written as
xb =√PWbmb, where Wb = [wb,1, . . . ,wb,Kb ] is the N ×Kb RCI precoding matrix, given by
PEEL et al. [2005]
Wb =1√ζb
H†b
(HbH
†b +NξIKb
)−1, (6.2)
and where ζb = tr
H†bHb
(H†bHb +NξIN
)−2
is a long-term power normalization constant.
The function of the real regularization parameter ξ is to achieve a tradeoff between the signal power
at the legitimate user and the crosstalk at the other users served by the same BS. The optimal value
for the parameter ξ in cellular networks is unknown. Since the results obtained in this chapter
hold for any value of ξ, we will now assume that each BS sets ξ to the value that maximizes the
large-system secrecy rate in an isolated cell, given by ξBCC in (4.10).
6.1.3 Malicious Users
In general, the BSs cannot determine the behavior of the users, i.e., whether they act maliciously
as eavesdroppers or not. As a worst-case scenario, we assume that for each legitimate user, all
the remaining users in the network can act as eavesdroppers. For a user o connected to the BS
b, the set of Kb − 1 malicious users within the same cell is denoted by MIo = Kb\o, and the
set formed by the rest of the malicious users in the network is denoted by MEo = Φu\Kb. In
Fig. 6.1, the legitimate user o, the set of (intra-cell) malicious usersMIo, and the set of (external)
malicious users MEo are represented by star, triangles, and squares, respectively. The total set
of malicious users for the legitimate receiver o is denoted byMo = MIo ∪ME
o = Φu\o. It is
important to make such a distinction between the intra-cell malicious users inMIo and the external
malicious users inMEo . In fact, the BS b knows the channels of the intra-cell malicious users in
MIo ⊂ Kb, and exploits this information by choosing an RCI precoding matrix Wb which is a
function of these channels. The RCI precoding thus controls the amount of information leakage
at the malicious users inMIo. On the other hand, the BS b does not know the channels of all the
94 Achievable Secrecy Rates 6.2
other external malicious users inMEo , and Wb does not depend upon these channels. Therefore,
the signal received by the malicious users inMEo is not directly affected by RCI precoding.
Table 6.1: Notation Summary
Notation Description
Φb; λb A PPP modeling the locations of BSs; deployment density of BSs
Φu; λu An independent PPP modeling the locations of users; density of users
P ; ρ Downlink transmit power for each BS; transmit SNR ρ , Pσ2
N Number of transmit antennas for each BS
c BS which is closest to the origin o
Kb; Kb = |Kb| Set of users associated with BS b; number of users associated with BS b
MIo = Kc\o Set of the Kc − 1 malicious users in the same cell as the typical user
MEo = Φu\Kc Set formed by the malicious users in all the other cells
mb Confidential messages sent by BS b to its users
xb =√PWbmb Signal transmitted by BS b
hb,j ∼ CN (0, I) Channel vector between BS b and user j
η Path loss exponent
gb,o ∼ Γ(Kb, 1) Inter-cell interference power gain from BS b to the typical user in o
gc,e ∼ exp(1) Leakage power gain from BS c to the malicious user e ∈MEo
6.2 Achievable Secrecy Rates
In this section, we derive a secrecy rate achievable by RCI precoding for the typical user in the
downlink of a cellular network.
6.2 Physical Layer Security in Cellular Networks 95
6.2.1 SINR at a Typical User
We consider a typical user o located at the origin, and connected to the closest BS, located in
c ∈ Φb. The distance between the typical user and the closest BS is given by ‖c‖. The typical user
receives self-interference caused by the other messages mc,u, u 6= o transmitted by the BS c, and
inter-cell interference caused by the signal transmitted by all the other BSs b ∈ Φb\c. The signal
received by the typical user is given by
yo =√P ‖c‖−η h†c,owc,omc,o +
√P ‖c‖−η
∑u∈Kc\o
h†c,owc,umc,u
+∑
b∈Φb\c
√P ‖b‖−η
Kb∑j=1
h†b,owb,jmb,j + no (6.3)
where ‖b‖ is the distance between the typical user and the generic BS b, and η is the path loss
exponent. The four terms in (6.3) represent the useful signal, the crosstalk (or self-interference),
the inter-cell interference, and the thermal noise seen at the typical user, respectively. The latter is
given by no ∼ CN (0, σ2), and we define the transmit SNR as ρ , P/σ2.
We assume that the legitimate receiver at o treats the interference power as noise. The SINR γo at
the legitimate receiver o is given by
γo =ρ‖c‖−η
∣∣∣h†c,owc,o
∣∣∣2ρ‖c‖−η∑u∈Kc\o
∣∣∣h†c,owc,u
∣∣∣2 + ρ∑
b∈Φb\cgb,oKb‖b‖−η + 1
, (6.4)
where we define wb,j ,√Kbwb,j and
gb,o ,Kb∑j=1
∣∣∣h†b,owb,j
∣∣∣2 . (6.5)
6.2.2 SINR at the Malicious Users
The cell where the typical user o is located is referred to as the tagged cell. For the typical user
o, the set of malicious users is denoted byMo = MIo ∪ME
o , whereMIo = Kc\o is the set of
remaining users in the tagged cell, andMEo = Φu\Kc is the set of all users in other cells.
We assume that each malicious user can communicate directly to any other malicious user within
a cooperation radius rc around it, i.e., cooperation is possible for distances smaller than rc. This
96 Achievable Secrecy Rates 6.2
Figure 6.2: Example of percolation in a random plane network. Dots represent malicious users,
and discs represents the cooperation range of malicious users. Two malicious users can cooperate
when their respective discs overlap.
assumption comes from the following model: assume that malicious users can transmit at a certain
power PM and that their signal is attenuated over distance according to a deterministic decreasing
function l(d). Assume also that malicious users can successfully receive data if the signal is at
least t times stronger than the ambient noise, which has power Pn. Even under the condition that
the interference is perfectly canceled, there is a maximum distance beyond which the two users
cannot cooperate. This distance, referred to as the cooperation radius, is given by
rc , max
d :
PM l(d)
Pn≥ t. (6.6)
By connecting each pair of cooperating malicious users, it is possible to generate a random plane
network GILBERT [1961], which represents an infinite cooperation network of malicious users
with range rc. It is known that if the density of users satisfies λu > 8 log 2r2c
, the random plane
network contains an infinite cluster almost surely. An illustration of this phenomenon, known as
percolation, is provided in Fig. 6.2. As a result, if the density of users λu is large enough, then
percolation will occur, and there will be an infinite set of malicious users cooperating to eavesdrop
the message intended for the typical user.
Motivated by these observations, in the following we will consider the worst-case scenario where
all the malicious users inMo can cooperate to eavesdrop on the message intended for the typical
user in o. Since each malicious user is likely to decode its own message, it can indirectly pass this
6.2 Physical Layer Security in Cellular Networks 97
information to all the other malicious users. In the worst-case scenario, all the malicious users in
Mo can therefore subtract the interference generated by all the messages mj , j 6= o.
After interference cancellation, the signal received at a malicious user i ∈ MIo in the tagged cell
is given by
yi =√P ‖i− c‖−η h†c,iwc,omc,o + ni (6.7)
where ‖i−c‖ is the distance between the BS c and the malicious user i ∈MIo. The signal received
at a malicious user e ∈MEo outside the tagged cell is given by
ye =√P ‖e− c‖−η h†c,ewc,omc,o + ne. (6.8)
We denote by γi and γe the SINRs at the malicious users i ∈MIo and e ∈ME
o , respectively.
Due to the cooperation among all malicious users inMo =MIo∪ME
o , the setMo can be seen as
a single equivalent multi-antenna malicious user, denoted by Mo. After interference cancellation,
Mo sees the useful signal embedded in noise, therefore applying maximal ratio combining is
optimal, and yields to an SINR given by
γM,o =∑i∈MI
o
γi +∑e∈ME
o
γe = ρ∑i∈MI
o
‖i− c‖−η∣∣∣h†c,iwc,o
∣∣∣2 +ρ
Kc
∑e∈ME
o
gc,e‖e− c‖−η, (6.9)
where
gc,e ,∣∣∣h†c,ewc,o
∣∣∣2 , (6.10)
with wc,o ,√Kcwc,o and ni, ne ∼ CN (0, σ2).
6.2.3 Achievable Secrecy Rates
We are now able to obtain an expression for the secrecy rate achievable by RCI precoding for the
typical user of a downlink cellular network.
Proposition 6.1. A secrecy rate achievable by RCI precoding for the typical user o is given by
RCELL ,
log2
(1 +
ρ‖c‖−η∣∣∣h†c,owc,o
∣∣∣2ρ‖c‖−η∑i∈MI
o
∣∣∣h†c,owc,i
∣∣∣2 + ρI + 1
)
− log2
(1 + ρ
∑i∈MI
o
‖i− c‖−η∣∣∣h†c,iwc,o
∣∣∣2 + ρL
)+
, (6.11)
98 Achievable Secrecy Rates 6.2
where I and L represent the interference and leakage term, respectively, given by
I =∑
b∈Φb\c
gb,oKb‖b‖−η (6.12)
L =1
Kc
∑e∈ME
o
gc,e‖e− c‖−η. (6.13)
Proof. The BS c, the user o, and the equivalent malicious user Mo form an equivalent multi-input,
single-output, multi-eavesdropper (MISOME) wiretap channel KHISTI AND WORNELL [2010].
As a result, an achievable secrecy rate is given by
RCELL = log2 (1 + γo)− log2 (1 + γM,o)+ . (6.14)
Substituting (6.4) and (6.9) in (6.14) yields (6.11).
The statistics of the terms gb,o and gc,e in (6.5) and (6.10), respectively, can be characterized as
follows DHILLON et al. [2013].
Proposition 6.2. For regularized channel inversion precoding we have that (i) the inter-cell inter-
ference power gain at the typical legitimate user o is distributed as gb,o ∼ Γ(Kb, 1), and (ii) the
leakage power gain at the malicious user e ∈MEo is distributed as gc,e ∼ exp(1).
Proof. See Appendix C.1.
We now define the probability of secrecy outage and the mean secrecy rate for the typical user.
Definition 6.1. The probability of secrecy outage for the typical user o is defined as
OCELL , P(RCELL ≤ 0). (6.15)
The probability of secrecy outage also denotes the fraction of time for which a BS cannot transmit
to a typical user at a non-zero secrecy rate.
Definition 6.2. The mean secrecy rate for the typical user o is defined as
RCELL , E [RCELL] . (6.16)
6.3 Physical Layer Security in Cellular Networks 99
6.3 Large-system Analysis
In this section, we derive approximations for (i) the secrecy outage probability, i.e., the probability
that the secrecy rate RCELL achievable by RCI precoding for the typical user o is zero, and (ii) the
mean secrecy rate achievable by RCI precoding in the downlink of a cellular network.
6.3.1 Preliminaries
Throughout the analysis, we make the following assumptions.
Assumption 6.1. For uniformity of notation, we assume Kc = Kb = K , λuλb
, ∀b, i.e., we
approximate the number of users served by each BS by its average value, given by the ratio between
the density of users and the density of BSs. In order for this equivalence to hold, we ignore a small
bias that makes the tagged cell bigger than a typical cell. This bias is a result of Feller’s paradox,
also known as waiting bus paradox in one dimension BACCELLI AND BREMAUD [2003].
Assumption 6.2. We assume ‖i− c‖ ≈ ‖c‖, ∀i ∈MIo, i.e., we approximate the distance between
the tagged BS c and each user connected to c by the distance between the BS c and the typical
user o. We then approximate the Voronoi region of the tagged BS c by a ball centered at c and with
radius r = 1√πλb
, i.e., B(c, r) ,m ∈ R2, ‖m− c‖ ≤ r
. For the sake of consistency, the value
of r is chosen to ensure that B(c, r) has the same area as the average cell.
Note that despite these assumptions, which are necessary to maintain tractability, our analysis
captures all the key characteristics of the cellular networks that affect physical layer security, as
discussed in the sequel. The simplified model also provides some fundamental insights into the
dependence of key performance metrics, such as secrecy outage and mean secrecy rate, on the
transmit power and BS deployment density.
Under Assumptions 6.1 and 6.2, we obtain the approximations MIo ≈ MI
o and MEo ≈ ME
o ,
where MIo is a set of K − 1 malicious users located at distance ‖c‖ from the BS c, and ME
o is a
set given by
MEo = e ∈ Φu ∩ B(c, r), (6.17)
100 Large-system Analysis 6.3
with B denoting the complement of the set B. We can then approximate the interference and
leakage terms in (6.12) and (6.13) as follows
I ≈ I =1
K
∑b∈Φb\c
gb,o‖b‖−η (6.18)
L ≈ L =1
K
∑e∈ME
o
gc,e‖e− c‖−η. (6.19)
We now carry out a large-system analysis by assuming that both (i) the average number of users
K in each cell, and (ii) the number of transmit antennas N at the each BS grow to infinity in a
fixed ratio β , KN . We can thus approximate the remaining random variables in (6.11) with the
following deterministic quantities NGUYEN AND EVANS [2008]; WAGNER et al. [2012]∣∣∣h†c,owc,o
∣∣∣2 ≈ α, ∑i∈MI
o
∣∣∣h†c,owc,i
∣∣∣2 ≈ χ, and∑i∈MI
o
∣∣∣h†c,iwc,o
∣∣∣2 ≈ χ, (6.20)
where
α =g (β, ξ)
1 + ξ
β [1 + g (β, ξ)]2
[1 + g (β, ξ)]2, χ =
1
[1 + g (β, ξ)]2, (6.21)
g (β, ξ) is given in (4.7), and where it follows from (4.10) that
limρ→∞
χ = 0, for β ≤ 1. (6.22)
An approximated secrecy rate can be therefore obtained as follows.
Definition 6.3. An approximation for the achievable secrecy rate RCELL is given by
RCELL ≈ RCELL ,
log2
(1+
ρα‖c‖−ηρχ‖c‖−η + ρI + 1
)−log2
(1+ρχ‖c‖−η+ρL
)+
. (6.23)
In Fig. 6.3 we compare the simulated ergodic secrecy rate RCELL in (6.11) to the approximation
RCELL in (6.23), obtained in the large-system regime under Assumptions 6.1 and 6.2. The secrecy
ratesRCELL and RCELL are plotted versus the transmit SNR ρ, for a system withN = 20 transmit
antennas, an average numberK = 20 of users per cell, a path loss exponent η = 4, and two values
of the density of BS λb. Figure 6.3 shows that RCELL and RCELL follow the same trend, and
that the approximation RCELL ≈ RCELL is reasonable. The figure also shows that in a cellular
network the secrecy rate does not monotonically increase with the transmit SNR. A more detailed
discussion on this phenomenon will be provided in Subsection 6.3.4.
6.3 Physical Layer Security in Cellular Networks 101
0 5 10 15 20 25 300
0.05
0.1
0.15
0.2
SNR, ρ [dB]
Per-usersecrecyrate
RCELL
RCELL
λb = 0.01
λb = 0.1
Figure 6.3: Comparison between the simulated ergodic secrecy rate RCELL in (6.11) and the
approximation RCELL in (6.23) versus the SNR, for N = 20 transmit antennas, an average of
K = 20 users per BS, and η = 4.
6.3.2 Characterization of Interference and Leakage
We now provide some results on the Laplace transforms of the terms I and L which will be useful
in the remainder of the chapter.
Lemma 6.1. The Laplace transform of the interference term LI(s, ‖c‖) = E[e−sI ] is
LI(s, ‖c‖) = exp
(−( sK
) 2ηλbCη,K (s, ‖c‖)
)(6.24)
where
Cη,K (s, ‖c‖) =2π
η
K∑n=1
(K
n
)[B
(1;K − n+
2
η, n− 2
η
)−B
((1 +
sP
K‖c‖−η)−1;K − n+
2
η, n− 2
η
)], (6.25)
and B(x; y, z) =∫ x
0 ty−1(1− t)z−1dt is the incomplete Beta function.
Proof. See Appendix C.2.
102 Large-system Analysis 6.3
Lemma 6.2. The Laplace transform of the leakage term LL(s) = E[e−sL] is
LL(s) = exp
(−λu
( sK
) 2ηDη(s)
)(6.26)
where
Dη(s) =2π
η
[B
(1;
2
η, 1− 2
η
)−B
(1
1 + sK r−η ;
2
η, 1− 2
η
)]. (6.27)
Proof. The proof is omitted since it is similar to the proof of Lemma 6.1.
The probability density functions (pdfs) fI and fL of I and L, respectively, can be obtained by
inverting the respective Laplace transforms LI and LL. We now propose simple approximations
for fI and fL, using the following well-known results BACCELLI AND BŁASZCZYSZYN [2009].
Proposition 6.3. The mean and the variance of the interference term I are respectively given by
µI =2πλb‖c‖−(η−2)
η − 2, (6.28)
σ2I
=πλb
(K +K2
)‖c‖−2(η−1)
K2 (η − 1), (6.29)
whereas the mean and the variance of the leakage term L are respectively given by
µL =2πλur
−(η−2)
K(η − 2), (6.30)
σ2L
=2πλur
−2(η−1)
K2 (η − 1). (6.31)
Proof. See Appendix C.3.
We then approximate the pdfs of I and L by lognormal distributions with the same respective
mean and variance, as follows.
Definition 6.4. The probability density functions of I and L can be approximated as follows
fI(x) ≈ 1
xσI,N√
2πexp
−(
log x− µI,N)2
2σ2I,N
, x > 0 (6.32)
fL(z) ≈ 1
zσL,N√
2πexp
−(
log z − µL,N)2
2σ2L,N
, z > 0 (6.33)
6.3 Physical Layer Security in Cellular Networks 103
10−4
10−2
100
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
CDF
FI (sim.)
FI (approx.)
FL (sim.)
FI (approx.)
λb = 0.1
λb = 0.03
λb = 0.01
Figure 6.4: Comparison between the simulated cumulative distribution functions (CDFs) of I and
L and the lognormal approximations in (6.32) and (6.33), for an SNR ρ = 10dB,N = 20 transmit
antennas, K = 20 users per BS, ‖c‖ = r, and η = 4.
where
µI,N = logµI −1
2log
(1 +
σ2I
µ2I
), σ2
I,N= log
(1 +
σ2I
µ2I
)(6.34)
µL,N = logµL −1
2log
(1 +
σ2L
µ2L
), σ2
L,N= log
(1 +
σ2L
µ2L
). (6.35)
In Fig. 6.4 we compare the simulated cumulative distribution functions (CDFs) of I and L to the
lognormal approximations provided in (6.32) and (6.33). The CDFs are plotted for a transmit SNR
ρ = 10dB, N = 20 transmit antennas, an average of K = 20 users per BS, ‖c‖ = r, η = 4, and
three values of the density of BS λb. Figure 6.4 shows that the lognormal approximations provided
in Definition 6.4 are accurate for all values of λb.
6.3.3 Probability of Secrecy Outage
We now obtain an approximation for the probability of secrecy outage with RCI precoding.
104 Large-system Analysis 6.3
Theorem 6.3. The probability of secrecy outage with RCI precoding can be approximated as
OCELL ≈ OCELL =
∫ ∞0
∫ ∞−∞
∫ ∞−∞
1(z≥τ(x,y)) fL(z) dz fI(x, y) dx 2λbπye−λbπy2
dy, (6.36)
where fI(x, y) is the probability density function of the interference I for ‖c‖ = y, fL(z) is the
probability density function of the leakage L, and where we have defined
τ(x, y) ,αy−η
ρχy−η + ρx+ 1− χy−η. (6.37)
Proof. By using approximation (6.23) in (6.15), we obtain
OCELL ≈ P(RCELL ≤ 0) = P
(ρχ‖c‖−η + ρL ≥ α‖c‖−η
χ‖c‖−η + I + 1ρ
)
= P
(L ≥ α‖c‖−η
ρχ‖c‖−η + ρI + 1− χ‖c‖−η
)(a)=
∫ ∞−∞
∫ ∞−∞
P(L ≥ τ(x, y)
)fI(x, y | ‖c‖ = y) f‖c‖(y) dx dy
=
∫ ∞0
∫ ∞−∞
E[1(L≥τ(x,y))
]fI(x, y) dx 2λbπye
−λbπy2dy
=
∫ ∞0
∫ ∞−∞
∫ ∞−∞
1(z≥τ(x,y)) fL(z) dz fI(x, y) dx 2λbπye−λbπy2
dy, (6.38)
where (a) holds by defining τ(x, y) as in (6.37), and by noting that the distance ‖c‖ between the
typical user and the nearest BS c has distribution HAENGGI [2005]
f‖c‖(y) = 2λbπy exp(−λbπy2), y > 0. (6.39)
The result provided in Theorem 6.3 allows to evaluate the probability of secrecy outage without
the need for Monte-Carlo simulations, which can be computationally expensive to account for all
users and all exact Voronoi cells. Moreover, Theorem 6.3 yields to the following asymptotic result
without the need to solve the integral.
In Chapter 5, we showed that for an isolated cell in a random field of eavesdroppers, a sufficient
number of transmit antennas allows the BS to cancel the intra-cell interference and leakage, and
to drive the probability of secrecy outage to zero. In a cellular network, the secrecy outage is also
6.3 Physical Layer Security in Cellular Networks 105
caused by the inter-cell interference and leakage, which cannot be controlled by the BS. It is easy
to show that limρ→∞ τ(x, y) ≤ 0, which from Theorem 6.3 implies limρ→∞ OCELL = 1. We
therefore have the following result.
Remark 6.1. In cellular networks, RCI precoding can achieve confidential communication with
probability of secrecy outage OCELL < 1. However unlike an isolated cell, cellular networks tend
to be in secrecy outage w.p. 1 if the transmit power grows unbounded, irrespective of the number
of transmit antennas.
6.3.4 Mean Secrecy Rate
In the following, we derive the mean secrecy rate achievable by RCI precoding.
Theorem 6.4. The mean secrecy rate achievable by RCI precoding can be approximated as
RCELL ≈ RCELL, with
RCELL =
∫ ∞0
∫ αρχ− 1ρ−χy−η
−∞
log2
(1 +
ραy−η
ρχy−η + ρx+ 1
)∫ τ(x,y)
−∞fL(z)
−∫ τ(x,y)
−∞log2
(1 + ρχy−η + ρz
)fL(z) dz
fI(x, y) dx 2λbπye
−λbπy2dy. (6.40)
Proof. See Appendix C.4.
The result provided in Theorem 6.4 allows to evaluate the mean secrecy rate without the need
for computationally expensive Monte-Carlo simulations. Moreover, Theorem 6.4 yields to the
following asymptotic result without the need to solve the integral.
In Chapters 4 and 5 we showed that in an isolated cell, even in the presence of a random field
of eavesdroppers, a sufficient number of transmit antennas allows the BS to cancel the intra-cell
interference and leakage, and the secrecy rate increases monotonically with the SNR. In a cellular
network, the secrecy rate is also affected by the inter-cell interference and leakage, which cannot be
controlled by the BS. It is easy to show that limρ→∞αρχ− 1
ρ−χy−η ≤ 0, which from Theorem 6.4
implies limρ→∞ RCELL = 0. We therefore have the following result.
106 Large-system Analysis 6.3
Remark 6.2. In cellular networks, RCI precoding can achieve a non-zero secrecy rate RCELL.
However unlike an isolated cell, the secrecy rate in a cellular network is interference-and-leakage-
limited, and it cannot grow unbounded with the transmit SNR, irrespective of the number of trans-
mit antennas.
Theorem 6.4 shows that an optimal value for the BS deployment density λb should be found as a
tradeoff between (i) increasing the useful power αy−η, and (ii) reducing the intra-cell interference
χy−η and leakage χy−η, and the inter-cell interference x and leakage z. We know from (6.22) that
χ vanishes at high SNR, thus the terms x and z become dominant in (6.40). For a given cell load
K = λuλb
, the terms x and z are minimized by small densities λb. We therefore have the following
result which we will validate by simulations in Section 6.4.
Remark 6.3. In a cellular network with a fixed load, i.e., average number of users per BS, there
is an optimal value for the deployment density of BSs that maximizes the mean secrecy rate, and
this value is a decreasing function of the SNR. The optimal value of λb can be found from (6.40)
by performing a linear search.
In order to calculate the mean secrecy rate in (6.40), one must obtain expressions for fI and fL
via Laplace anti-transform or via approximations, as discussed in Section 6.3.2. We now derive a
lower bound on RCELL which can be calculated without knowledge of fI and fL.
Corollary 6.5. The approximated mean secrecy rate RCELL can be lower bounded as RCELL ≥RLBCELL, with
RLBCELL =
∫ ∞0
∫ ∞−∞
[F∗1 (φ, y)LI(−i2πφ, y)−F∗2 (φ, y)LL(−i2πφ)
]dφ2λbπye
−λbπy2dy
+
,
(6.41)
and
F1(φ, y) =sgn(φ) e
2πi(χy−η+ 1ρ
)φ
2φ log 2
(1− e2πiαφy−η
), (6.42)
F2(φ, y) =−e2πi(χy−η+ 1
ρ)φ
log 2
[1
2 |φ| +γ
ρδ(φ)
]. (6.43)
Proof. See Appendix C.5.
6.4 Physical Layer Security in Cellular Networks 107
0 5 10 15 20 25 300.5
0.6
0.7
0.8
0.9
1
SNR, ρ [dB]
Probabilityofsecrecyoutage
OCELL (analysis)
OCELL (simulations)
λb = 0.01
λb = 0.03
λb = 0.1
Figure 6.5: Comparison between the simulated probability of secrecy outage OCELL and the ana-
lytical result from Theorem 6.3, for N = 20 transmit antennas, K = 20 users per BS, and three
values of the density of BSs λb.
6.4 Numerical Results
In Fig. 6.5 we compare the simulated probability of secrecy outage OCELL to the analytical result
given in Theorem 6.3, forN = 20 transmit antennas,K = 20 users per BS, and three values of the
density of BSs λb. The analytical curves were obtained by using lognormal approximations for the
pdfs fI(x, y) and fL(z). The figure shows that the result provided in Theorem 6.3 is accurate for
all values of λb at relatively low values of SNR. Due to the lognormal approximations, the result
is slightly less accurate at relatively high values of SNR, when the BS can cancel the intra-cell
interference and leakage, and the secrecy outage is mostly determined by I and L.
In Fig. 6.6 we compare the simulated mean secrecy rate RCELL to the analytical result given in
Theorem 6.4, for N = 20 transmit antennas, K = 20 users per BS, and two values of the den-
sity of BSs λb. The analytical curves were again obtained by using lognormal approximations for
the pdfs fI(x, y) and fL(z). The figure shows that the simulations and the analytical result from
Theorem 6.4 follow the same trend. Therefore, the analysis provides insights into the behavior
of the secrecy rate as a function of λb and the SNR. The result from Theorem 6.4 is accurate for
all values of λb at relatively low values of SNR. Again due to the lognormal approximations, the
108 Numerical Results 6.4
0 5 10 15 20 25 300
0.05
0.1
0.15
0.2
SNR, ρ [dB]
Per-usersecrecyrate
RCELL (analysis)
RCELL (simulations)
λb = 0.01
λb = 0.1
Figure 6.6: Comparison between the simulated mean secrecy rate RCELL and the analytical result
from Theorem 6.4, for N = 20 transmit antennas, K = 20 users per BS, and two values of the
density of BSs λb.
analytical curve is less accurate at relatively high values of SNR, since χ vanishes as reported in
(6.22), and the secrecy rate in (6.40) is dominated by I and L. This inaccuracy could be avoided
by employing the exact pdfs of fI(x, y) and fL(z), obtained from their Laplace transforms in
Lemma 6.1 and Lemma 6.2, but the anti-transform operation is computationally expensive. Find-
ing better approximations for fI(x, y) and fL(z) is therefore identified as a promising research
problem.
In Fig. 6.7 we plot the simulated probability of secrecy outage versus the transmit SNR, for K =
10 users per BS and three values of the number of transmit antennasN . In this figure, two cases are
considered for the density of BSs λb, namely 0.01 and 0.1, while the density of users is given by
λu = Kλb. Figure 6.7 shows that RCI precoding achieves confidential communications in cellular
networks with probability of secrecy outage OCELL < 1, and that having more transmit antennas is
beneficial as it reduces the probability of secrecy outage. However unlike an isolated cell, cellular
networks tend to be in secrecy outage w.p. 1 if the transmit power grows unbounded, irrespective
of the number of transmit antennas. These observations are consistent with Remark 6.1.
In Fig. 6.8 we plot the simulated per-user ergodic secrecy rate versus the transmit SNR, forK = 10
6.4 Physical Layer Security in Cellular Networks 109
−10 0 10 20 30 400.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
SNR, ρ [dB]
Probabilityofsecrecyoutage
N = 10
N = 12
N = 14
λb = 0.1
λb = 0.01
Figure 6.7: Simulated probability of secrecy outage versus transmit SNR, for K = 10 users per
BS and various values of the number of transmit antennas N and density of BSs λb.
users per BS and three values of the number of transmit antennasN . In this figure, again, two cases
are considered for the density of BSs λb, namely 0.01 and 0.1, while the density of users is given
by λu = Kλb. Figure 6.8 shows that in cellular networks RCI precoding can achieve a non-zero
secrecy rate, and that having more transmit antennas is beneficial as it increases the secrecy rate.
However unlike the secrecy rate in an isolated cell, the secrecy rate in a cellular scenario does not
grow unbounded with the SNR, even with a large number of transmit antennas. These observations
are consistent with Remark 6.2.
In Fig. 6.9 we plot the simulated per-user ergodic secrecy rate as a function of the density of BSs
λb, for N = 20 transmit antennas, K = 20 users per BS, and various values of the transmit
SNR. Figure 6.9 shows that there is an optimal value for the density of BSs λb that maximizes
the secrecy rate, and that such value is smaller for higher values of the SNR. This observation is
consistent with Remark 6.3.
110 Numerical Results 6.4
−10 0 10 20 30 400
0.2
0.4
0.6
0.8
1
SNR, ρ [dB]
Per-usersecrecyrate
N = 10
N = 12
N = 14λb = 0.01
λb = 0.1
Figure 6.8: Simulated ergodic secrecy rate versus transmit SNR, for K = 10 users per BS and
various values of the number of transmit antennas N and density of BSs λb.
10−3
10−2
10−1
100
0
0.05
0.1
0.15
0.2
0.25
Per-usersecrecyrate
Density of BSs, λb
ρ = 0 dB
ρ = 10 dB
ρ = 20 dB
ρ = 30 dB
Figure 6.9: Simulated ergodic secrecy rate versus density of BSs, for N = 20 transmit antennas,
K = 20 users per BS, and various values of the transmit SNR ρ.
6.5 Physical Layer Security in Cellular Networks 111
6.5 Conclusion
In this chapter, we considered physical layer security for the downlink of cellular networks, where
multiple base stations (BSs) generate inter-cell interference, and malicious users of neighboring
cells can cooperate to eavesdrop. We showed that RCI precoding can achieve a non-zero secrecy
rate with probability of outage smaller than one. However we also found that unlike isolated cells,
the secrecy rate in a cellular network does not grow monotonically with the transmit signal-to-
noise ratio (SNR), and the network tends to be in secrecy outage if the transmit power grows
unbounded. We further showed that there is an optimal value for the density of BSs that max-
imizes the secrecy rate, and this value is a decreasing function of the transmit SNR. Using the
developed analysis, we clearly established the importance of designing the transmit power and the
BS deployment density to make communications robust against malicious users in other cells.
112
Chapter 7
Conclusions and Future Work
In this thesis, we studied physical layer security for multi-user multi-antenna wireless systems.
Physical layer security was recently proposed as a method to protect confidential information
without requiring encryption keys.
We started by considering the MISO broadcast channel with confidential messages (BCC) under
Rayleigh fading, where a multi-antenna base station (BS) simultaneously transmits independent
confidential messages to several spatially dispersed malicious users that can eavesdrop on each
other. For this system set-up, we proposed a low-complexity linear precoder based on regularized
channel inversion (RCI), and we derived the achievable secrecy rates, i.e., the rates at which each
message can be reliably transmitted to the intended user while no information is leaked at the
other users. We showed that RCI precoding outperforms plain channel inversion precoding, and
that the secrecy requirements result in a loss in terms of the sum-rate. We then proposed an
algorithm to jointly optimize the regularization parameter of the precoder and the power allocation
vector, and showed that optimal power allocation increases the secrecy sum-rate compared to equal
power allocation by up to 20% at practical values of the signal-to-noise ratio (SNR). Finally, we
extended the secrecy rate analysis to more practical scenarios where only imperfect channel state
information is available at the transmitter, and where channel correlation is present among the
transmit antenna elements.
Fundamental tools from random matrix theory allowed us to carry out a large-system analysis and
obtain closed form approximations for the achievable secrecy rates in the MISO BCC. Simulations
113
114 Conclusions and Future Work 7.0
confirmed that the approximations obtained via our analysis are accurate even for finite systems
of practical size. We derived expressions for the optimal regularization parameter of the precoder
ξ and for the optimal network load β, which maximize the secrecy sum-rate in the MISO BCC.
These values are different to the respective ones that maximize the sum-rate in the absence of
secrecy requirements. We found that for β > 1, i.e., when the number of users K is larger than
the number of transmit antennas N , the RCI precoder performs poorly in the high-SNR regime.
This happens because an alliance of cooperating malicious users can cancel the interference, and
therefore has an unbounded signal-to-interference-plus-noise ratio (SINR), while the legitimate
user’s SINR is interference limited. The only means for the transmitter to limit the eavesdropper’s
SINR is by inverting the channel matrix, but this can only be accomplished when β ≤ 1. In order to
increase the high-SNR performance for network loads β > 1, we proposed a linear precoder based
on RCI and power reduction (RCI-PR). The proposed RCI-PR precoder was showed to achieve a
per-user secrecy rate with the same high-SNR scaling factor as both the following upper bounds:
(i) the per-user rate of the optimal RCI precoder in the MISO broadcast channel (BC) without
secrecy requirements, and (ii) the secrecy capacity of a single-user system without interference.
We then studied the large-system performance of RCI precoding in practical channels. We first
considered the presence of channel state information (CSI) error, and determined how the CSI error
must scale with the SNR in order to ensure a constant high-SNR rate gap to the case with perfect
CSI. Such constant rate gap ensures that the multiplexing gain is not affected. For frequency
division duplex (FDD) systems, where users quantize their estimated channel vectors and send
the quantization index back to the BS over a limited-rate channel, we determined the minimum
amount of feedback bits required in order to meet a given secrecy rate gap. For time division
duplex (TDD) systems, where the channel estimation at the transmitter is obtained from known
pilot symbols sent by the users, we determined the optimum amount of channel training that
maximizes the secrecy sum-rate. We then applied our analysis to the MISO BCC under transmit-
side channel correlation. We obtained large-system approximations for the secrecy rates and for
the optimal regularization parameter of the precoder. We found that low-to-moderate correlation
only partially affects the secrecy rates. However, high correlation degrades the performance of the
RCI precoder, especially at low SNR.
We then introduced the broadcast channel with confidential messages and external eavesdroppers
(BCCE), where a multi-antenna BS simultaneously communicates to multiple malicious users, in
7.0 Conclusions and Future Work 115
the presence of a PPP of external eavesdroppers. Unlike the BCC, in the BCCE not just malicious
users, but also randomly located external nodes can act as eavesdroppers. By using results from
stochastic geometry, we obtained the probability of secrecy outage and the mean secrecy rate
for the RCI precoder in the BCCE under Rayleigh fading, for the two cases of non-colluding
and colluding eavesdroppers. We showed that, irrespective of the collusion strategy at the external
eavesdroppers, a large number of transmit antennasN drives both the probability of secrecy outage
and the rate loss due to the presence of external eavesdroppers to zero. Increasing the density of
eavesdroppers λe by a factor n, requires n2 as many antennas to meet a given probability of
secrecy outage and a given mean secrecy rate. Our analysis demonstrated that the number of
transmit antennas at the BS is a key resource to secure communications against malicious users
and external eavesdropping nodes.
We finally turned our attention to cellular networks where, unlike the case of isolated cells, multi-
ple BSs generate inter-cell interference, and malicious users of neighboring cells can cooperate to
eavesdrop. We characterized the probability of secrecy outage and the mean secrecy rate with RCI
precoding, accounting for the spatial distribution of BSs and users (modeled as independent PPPs)
and the fluctuations of their channels. We found that RCI precoding can achieve a non-zero se-
crecy rate with probability of outage smaller than one. However we also found that unlike isolated
cells, the secrecy rate in a cellular network does not grow monotonically with the SNR, and the
network tends to be in secrecy outage if the transmit power grows unbounded. We further showed
that there is an optimal value for the density of BSs that maximizes the secrecy rate, and this value
is a decreasing function of the SNR. Using the developed analysis, we clearly established the im-
portance of designing the transmit power and the BS deployment density to make communications
robust against malicious users in other cells.
This thesis considered scenarios where mobile users are equipped with a single receive antenna. In
addition, the analysis focused on a model where the noise power at each legitimate user is the same
for all users, and where the legitimate users of each cell have similar distances from the serving
BS, i.e., similar path-loss. A direct extension of this work considers the broadcast of confidential
messages in multiuser networks with multi-antenna receivers, under unequal path loss YANG et al.
[2013]. Furthermore, this thesis studied the performance of RCI precoding in cellular networks
by using approximations for the probability density functions (pdfs) of the inter-cell interference
power I and leakage power L. Obtaining an exact characterization of the pdfs of I and L could
116 Conclusions and Future Work 7.0
be an interesting future research direction. Finally, investigating the secrecy rates achievable in
heterogeneous networks, where small BSs are overlaid within the macro cellular network based
on traffic and coverage demand, is identified as another promising research problem.
Appendix A
Appendix for Chapter 4
A.1 Proof of Theorem 4.1
Proof. By defining Ak as in (3.22) and Bk as in (3.23), the SINRs at the intended user and the
eavesdropper can be expressed as in (3.24) and (3.25). We rewrite Ak = NK−1v†NQNvN , where
Qk,N =
(1
K − 1H†kHk − zI
)−1
∈ CN×N , vk,N =1√N
hk ∈ CN , z = − Nξ
K − 1. (A.1)
By Bai and Silverstein’s Lemma PEACOCK et al. [2008]; COUILLET AND DEBBAH [2011];
COUILLET et al. [2012], we have
E[φk
∣∣∣∣v†k,NQk,Nvk,N −1
NtrQk,N
∣∣∣∣p] ≤ CpNp
(N
λN
) p2
=Cp
Np2 λ
p2N
= fN ∀p ≥ 1 (A.2)
where Cp is a constant depending only on p, and φk = 1|λ1(H†kHk)−Nξ|,...,|λN (H†kHk)−Nξ|>λN,
with λN → 0. Assume ξ ∈ D′N , withD′N = DN for β ≤ 1 andD′N = DN\[− C
N12−ε
,+ C
N12−ε
]for β > 1. Then we have N = O(λ−2−ε
N ), for some ε > 0, and mink≤Kφk a.s.−→ 1 COUILLET
et al. [2012]. It follows from the Markov inequality and the Borel-Cantelli lemma BILLINGSLEY
[1995] that maxk|v†k,NQk,Nvk,N − 1
N trQk,N| a.s.−→ 0, as N → ∞. The term 1N trQk,N is
by definition the Stieltjes transform mH†kHk,N
(z) of H†kHk,N . Similarly, it can be shown that
maxk|m
H†kHk,N(z) − m(z)| a.s.−→ 0, where m(z) can be obtained as the solution of m(z) =[
1− NK−1 − z − z N
K−1m(z)]−1
. This yields
Ak − g (β, ξ)a.s.−→ 0 (A.3)
117
118 Proof of Theorem 4.1 A.1
with
g (β, ξ) = β−1m(ξ) =1
2
±√
(1− β)2
ξ2+
2 (1 + β)
ξ+ 1 +
1− βξ− 1
, (A.4)
and where in order for m to be a Stieltjes transform, the sign of the square root must be chosen the
same as the sign of ξ SILVERSTEIN AND CHOI [1995].
We now rewrite Bk = Ak − N2ξ
(K−1)2 v†k,NQ2k,Nvk,N , and similarly we have
E[φk
∣∣∣∣v†k,NQ2k,Nvk,N −
1
NtrQ2
k,N
∣∣∣∣p] ≤ CpNp
(N
λ2N
) p2
=Cp
Np2 λpN
= gN ∀p ≥ 1. (A.5)
Again, if N = O(λ−2−εN ), for some ε > 0, we have max
k|v†k,NQ2
k,Nvk,N − 1N trQ2
k,N|a.s.−→ 0,
as N →∞. We note that 1N trQ2
k,N is the Stieltjes transform of Q2k,N , given by
1
NtrQ2
k,N =
∫dFN (λ)
(λRN− z)2 =
∂
∂z
∫dFN (λ)
λRN− z = m′RN
(z) (A.6)
where FN (λ) is the distribution of the eigenvalues of RN . Since both mRk,N(z) and m(z) are
analytic functions, we have maxk|m′Rk,N
(z)−m′(z)| a.s.−→ 0, as N →∞, and it follows that
Bk −[
1
βm(z)− N2ξ
(K − 1)2
∂
∂zm(z)
]= Bk −
[g (β, ξ) + ξ
∂
∂ξg (β, ξ)
]a.s.−→ 0. (A.7)
For the power normalization constant ζ we have
ζ = tr
(H†H +NξI
)−1−Nξtr
(H†H +MξI
)−2
=1
βm
HH†N(z′)− ξ
β2m′
HH†N(z′)
(A.8)
where z′ = −NξK . If ξ ∈ D′N , if follows that
ζ −[g (β, ξ) + ξ
∂
∂ξg (β, ξ)
]a.s.−→ 0. (A.9)
By the continuity of RBCC,k and RBCC, it follows that the previous convergence results also hold
for ξ ∈[− C
N12−ε
,+ C
N12−ε
]and β > 1. Equation (4.2) then follows from (3.27), (A.3), (A.7),
(A.9), and by applying the continuous mapping theorem, the Markov inequality, and the Borel-
Cantelli lemma.
A.2 Appendix for Chapter 4 119
A.2 Proof of Theorem 4.5
Proof. From NGUYEN et al. [2009], by defining ρ ,ρ(1−τ2)ρτ2+1
and ξ , ξ1−τ2 , a deterministic
approximation for γik is given by
γi = g(β, ξ)ρ+ ξρ
β
[1 + g(β, ξ)
]2
ρ+[1 + g(β, ξ)
]2 . (A.10)
By defining Ωk =(H†kHk +NξI
)−1and using (3.40), we can rewrite γiM,k in (3.44) as
γiM,k = ρBk + 2
(1 + Ak
)Qk +
(1 + Ak
)2Rk
ζ(
1 + Ak
)2 , (A.11)
with
Ak = h†kΩkhk, (A.12)
Bk = h†kΩkH†kHkΩkhk, (A.13)
Qk = h†kΩkH†kHk
(Ωk −
Ωkhkh†kΩk
1 + Ak
)ek, (A.14)
Rk = e†k
(Ωk −
Ωkhkh†kΩk
1 + Ak
)H†kHk
(Ωk −
Ωkhkh†kΩk
1 + Ak
)ek. (A.15)
If ξ ∈ DN , we have
Ak − g(β, ξ)
a.s.−→ 0, (A.16)
Bk −[g(β, ξ)
+ ξ∂
∂ξg(β, ξ)]
a.s.−→ 0, (A.17)
Qka.s.−→ 0, (A.18)
Rk −τ2
1− τ2
[g(β, ξ)
+ ξ∂
∂ξg(β, ξ)]
a.s.−→ 0, (A.19)
and
ζ − 1
1− τ2
[g(β, ξ)
+ ξ∂
∂ξg(β, ξ)]
a.s.−→ 0, (A.20)
hence a deterministic approximation for γiM,k is given by
γiM = ρ
τ2 +1− τ2(
1 + g(β, ξ))2
. (A.21)
Theorem 4.5 then follows from (3.42), (A.10), (A.21), and from the continuous mapping theorem
BILLINGSLEY [1995].
120
Appendix B
Appendix for Chapter 5
B.1 Proof of Lemma 5.1
Proof. If γk ≤ γM,k, then RBCCE,k in (5.2) is zero w.p. 1. If γk > γM,k, we have for non-
colluding eavesdroppers γE,k = maxe
γe,k, therefore
OBCCE,k = P(γE,k ≥ γk
∣∣∣ γk) = 1− EΦe
[ ∏x∈Φe
P(γx,k < γk
∣∣∣ γk)]
= 1− EΦe
[ ∏x∈Φe
[1− P
(γx,k ≥ γk
∣∣∣ γk)]]
(a)= 1− EΦe
[ ∏x∈Φe
[1− exp
(−Nβσ2γk‖x‖η
)]](b)= 1− exp
[−2πλe
∫ ∞0
y exp(−Nβσ2γk y
η)dy
](c)= 1− exp
[−πλe
∫ ∞0
exp(−Nβσ2γk u
η2
)du
](d)= 1− exp
[− 2πλe
η(Nβσ2γk)2η
∫ ∞0
e−tt2η−1dt
]
(e)= 1− exp
− 2πλeΓ(
2η
)η(Nβσ2γk)
2η
(B.1)
where (a) follows from the distribution of γe,k, and (b) follows by using ‖x‖ = y, by applying the
probability generating functional (PGFL) for the PPP Φe, given by STOYAN et al. [1996]
EΦe
[ ∏x∈Φe
f(x)
]= exp
−∫R2
[1− f(x)]λedx
(B.2)
121
122 Proof of Lemma 5.4 B.2
and by changing to polar coordinates. Moreover, in (c) we have used u = y2, in (d) we have used
t = Nβσ2γkuη2 , and (e) follows from the definition of the gamma function
Γ(z) ,∫ ∞
0tz−1e−tdt. (B.3)
B.2 Proof of Lemma 5.4
Proof. If γk ≤ γM,k, then RBCCE,k in (5.2) is zero with probability one. If γk > γM,k, we have
for the eavesdropper nearest to the BS
OBCCE,k = P(γE,k ≥ γk
∣∣∣ γk)=
∫ ∞0
P(γE,k ≥ γk
∣∣∣ γk, ‖E‖ = x)f‖E‖(x)dx
=
∫ ∞0
P(x−ησ2
∣∣∣h†Ewk
∣∣∣2 ≥ γk ∣∣∣ γk, ‖E‖ = x)f‖E‖(x)dx
=
∫ ∞0
P( ∣∣∣h†Ewk
∣∣∣2 ≥ σ2γkxη∣∣∣ γk, ‖E‖ = x
)f‖E‖(x)dx
(a)=
∫ ∞0
exp(−Nβσ2γkx
η)f‖E‖(x)dx
(b)= 2πλe
∫ ∞0
x exp(−Nβσ2γk x
η − λeπx2)dx, (B.4)
where (a) holds because∣∣∣h†Ewk
∣∣∣2 ∼ exp( 1Nβ ), and (b) holds because the distance ‖E‖ between
the base station and the nearest eavesdropper E has distribution HAENGGI [2005]
f‖E‖(x) = 2λeπx exp(−λeπx2). (B.5)
For a path loss exponent η = 4, (B.4) reduces to
OBCCE,k = 2πλe
∫ ∞0
x exp(−Nβσ2γk x
4 − λeπx2)dx
(c)= πλe
∫ ∞0
exp(−Nβσ2γk u
2 − λeπu)du
(d)=
π32λe
2√Nβσ2γk
exp
[(πλe)
2
4Nβσ2γk
]erfc
(πλe
2√Nβσ2γk
)(B.6)
where in (c) we have used u = x2, and (d) follows from∫ ∞0
exp(−ax2 − bx)dx =1
2
√π
aexp
(b2
4a
)erfc
(b
2√a
). (B.7)
B.3 Appendix for Chapter 5 123
B.3 Proof of Lemma 5.6
Proof. For the case of colluding eavesdroppers, the Laplace transform of the SINR is HAENGGI
et al. [2009]
LγE,k(s) = E
[exp
(− s
σ2
∑x∈Φe
‖x‖−η∣∣∣h†xwk
∣∣∣2)](a)= exp
−2πλe
∫R2
Eh
[1− exp
(− s
σ2
∣∣∣h†xwk
∣∣∣2 ‖x‖−η)]x dx(b)= exp
−πλe Eh
[∣∣∣∣ 1σh†xwk
∣∣∣∣ 4η
]Γ
(1− 2
η
)s
2η
(c)= exp
−πλe
(Nβσ2
)− 2η Γ
(1 +
2
η
)Γ
(1− 2
η
)s
2η
(B.8)
where (a) holds since Φe is a PPP HAENGGI et al. [2009], (b) follows since the fading is indepen-
dent of the point process, and (c) follows since∣∣∣h†xwk
∣∣∣2 ∼ exp( 1Nβ ). Under a path loss exponent
η = 4, (B.8) reduces to
LγE,k(s) = exp(−π
2λe2
√s
Nβσ2
). (B.9)
By inverse transform one can obtain the probability distribution function SOUSA AND SILVESTER
[1990]
fγE,k(y) =π
32λey
− 32
4√Nβσ2
exp(− π4λ2
e
16Nβσ2y
), (B.10)
which integrated yields the cumulative distribution function
FγE,k(y) = erfc
[π2λe
4√Nβσ2y
], (B.11)
from which the secrecy outage probability in (5.17) can be calculated as OBCCE,k = FγE,k(γk).
124 Proof of Lemma 5.9 B.4
B.4 Proof of Lemma 5.9
Proof. We note from (5.2) that when γk ≤ γM,k, the secrecy rate RBCCE,k is zero ∀ γE,k. When
γk > γM,k, the mean secrecy rate is given by
EΦe [RBCCE,k|γk>γM,k]
= EΦe
[max
[log2
(1 + γk
)− log2
(1 + max (γE,k, γM,k)
), 0]]
= EΦe
[[log2
(1 + γk
)− log2
(1 + max (γE,k, γM,k)
)]1(γE,k<γk)
]= EΦe
[log2
(1 + γk
)1(γE,k<γk) − log2
(1 + max (γE,k, γM,k)
)1(γE,k<γk)
]= P (γE,k < γk) log2
(1 + γk
)− EΦe
[log2
(1+max (γE,k, γM,k)
)1(γE,k<γk)
]= P (γE,k < γk) log2
(1 + γk
)− EΦe
[log2
(1 + γM,k
)1(γE,k<γM,k) + log2
(1 + γE,k
)1(γM,k<γE,k<γk)
]= P (γE,k < γk) log2
(1 + γk
)− P (γE,k < γM,k) log2
(1 + γM,k
)−∫ γk
γM,k
log2(1 + y)fγE,k(y) dy
= log2
(1 + γk
)1−OBCCE,k−log2
(1 + γM,k
)1−PBCCE,k−∫ γk
γM,k
log2(1 + y)fγE,k(y) dy
= log2
(1 + γk
)1−OBCCE,k
(1 + γM,k
)1−PBCCE,k−∫ γk
γM,k
log2(1 + y)fγE,k(y) dy (B.12)
where (i) 1(·) is the indicator function, (ii) OBCCE,k , P (γE,k ≥ γk) is given by the secrecy
outage probability; (iii) PBCCE,k , P (γE,k ≥ γM,k) is the probability that the SINR at the exter-
nal eavesdroppers is greater than or equal to the SINR at the malicious users, given in (5.21) and
obtained by calculations similar to the ones in Lemma 5.1 and Lemma 5.6; and (iv) fγE,k(y) is the
pdf of the SINR at the external eavesdroppers, given by (B.10) for colluding eavesdroppers, and
by
fγE,k(y) =∂P (γE,k < y)
∂y=π
32λey
− 32
4√Nβσ2
exp
(− π
32λe
2√Nβσ2y
)(B.13)
for non-colluding eavesdroppers.
B.5 Appendix for Chapter 5 125
B.5 Proof of Corollary 5.11
Proof. For γ ≤ γM , we have RBCC = 0 and RBCCE = 0, therefore ∆e = 0. For γ > γM and
fixed ξ, irrespective of the cooperation strategy at the eavesdroppers, we have
∆e = OBCCE log(1 + γ)− PBCCE log(1 + γM ) +
∫ γ
γM
log2(1 + y)fγE,k(y) dy
(a)
≤ OBCCE ·RBCC +µλe
2√N
∫ γ
γM
y−12 dy
=
[1− exp
(− µλe√
Nγ
)]RBCC +
µλe√N
(√γ −
√γM
)≤ µλe√
NγRBCC +
µλe√N
(√γ −
√γM
)= µ
[RBCC√γ
+(√
γ −√γM
)] λe√N
(B.14)
where (a) holds because PBCCE > OBCCE, log2(1 + y) ≤ y, and fγE,k(y) ≤ µλey− 3
2
2√N
.
126
Appendix C
Appendix for Chapter 6
C.1 Proof of Proposition 6.2
Proof. Under RCI precoding, the BS b multiplies the confidential message mb,j destined for user
j, for 1 ≤ j ≤ Kb, by wb,j , so that the transmitted signal is a linear function of the message
mb,j , i.e., xb =√P∑Kb
j=1 wb,jmb,j . The inter-cell interference power gain at the typical user o
is given by gb,o =∑Kb
j=1
∣∣∣h†b,owb,j
∣∣∣2, with wb,j =√Kbwb,j . The normalized precoding vectors
wb,j have unit-norm on average, and they are calculated independently of h†b,o. Therefore, h†b,o and
wb,j are independent isotropic unit-norm random vectors, and∣∣∣h†b,owb,j
∣∣∣2 is a linear combination
of N complex normal random variables, i.e., exponentially distributed. As a result, we have that
gb,o ∼ Γ(Kb, 1), since it is the sum of Kb i.i.d. exponential r.v.
The leakage power gain at the malicious user e ∈ MEo is given by gc,e =
∣∣∣h†c,ewc,o
∣∣∣2, with
wc,o =√Kcwc,o. Similarly, we have that h†c,e and wc,o are independent isotropic unit-norm
random vectors. As a result, we have that gc,e ∼ exp(1) since it is a linear combination of N
complex normal r.v.
127
128 Proof of Lemma 6.1 C.2
C.2 Proof of Lemma 6.1
Proof. The Laplace transform of the interference term I can be derived as follows
E[e−sI
]= E
[e− sK
∑b∈Φb\c
gb,o‖b‖−η]
= E
∏b∈Φb\c
e−sKgb,o‖b‖−η
(a)= EΦb
∏b∈Φb\c
Lgb,o( sK‖b‖−η
)(b)= exp
−λb
∫R2∩B(o,‖c‖)
[1− Lgb,o
( sK‖b‖−η
)]db
(c)= exp
−λb
∫R2∩B(o,‖c‖)
[1− 1(
1 + sK ‖b‖−η
)K]
db
= exp
−λb
∫R2∩B(o,‖c‖)
(1 + s
K ‖b‖−η)K − 1(
1 + sK ‖b‖−η
)K db
(d)= exp
−λb∫R2∩B(o,‖c‖)
∑Kn=1
(Kn
) (sK ‖b‖−η
)n(1 + s
Kb‖b‖−η
)K db
= exp
−λb
K∑n=1
(K
n
)∫R2∩B(o,‖c‖)
(sK ‖b‖−η
)n(1 + s
K ‖b‖−η)K db
(e)= exp
−2πλb
( sK
) 2η
K∑n=1
(K
n
)∫ ∞‖c‖( s
K)− 1η
v−nη
(1 + v−η)Kvdv
(f)= exp
−λb
( sK
) 2ηCη,K (s, ‖c‖)
, (C.1)
where (a) follows since the channel powers gb,o are independent of the locations of the BSs, (b)
follows from the PGFL of a PPP STOYAN et al. [1996], (c) follows from the Laplace transform
of gb,o ∼ Γ(K, 1), (d) follows from the Binomial theorem, (e) follows by converting to polar
coordinates, and (f) follows by substituting (1 + ν−η)−1 = t and noting that the integral is the
difference of two incomplete Beta functions B(x; y, z) =∫ x
0 ty−1(1− t)z−1dt.
C.3 Appendix for Chapter 6 129
C.3 Proof of Proposition 6.3
Proof. The mean and variance of the interference can be obtained by applying Campbell’s theorem
and are given by BACCELLI AND BŁASZCZYSZYN [2009]
µI = E[I]
(a)= EΦb
∑b∈Φb\c
‖b‖−η = 2πλb
∫ ∞‖c‖
v−ηv dv =2πλb‖c‖−(η−2)
η − 2(C.2)
σ2I
= E[I2]− µ2
I
(b)=K +K2
K2E
∑b∈Φb\c
‖b‖−2η
=
2πλb(K +K2
)K2
∫ ∞‖c‖
v−2ηv dv =πλb
(K +K2
)‖c‖−2(η−1)
K2 (η − 1), (C.3)
where (a) follows from E[gb,o] = K, and (b) follows from E[g2b,o
]= K + K2. Similarly, the
mean and variance of the leakage are given by BACCELLI AND BŁASZCZYSZYN [2009]
µL = E[L]
(c)=
1
KEΦu
∑e∈ME
o
E [gc,e] ‖e− c‖−η =
2πλuK
∫ ∞r
v−ηv dv =2πλur
−(η−2)
K (η − 2)
(C.4)
σ2L
= E[L2]− µ2
L
(a)=
2
K2E
∑e∈ME
o
‖e− c‖−2η
=4πλuK2
∫ ∞r
v−2ηv dv =2πλur
−2(η−1)
K2 (η − 1),
(C.5)
where (c) follows from E[gc,e] = 1, and (d) follows from E[g2c,e
]= 2.
130 Proof of Corollary 6.5 C.5
C.4 Proof of Theorem 6.4
Proof. By using approximation (6.23) in (6.16), we obtain
RCELL ≈ E[RCELL
]= E
[log2
(1+
ρα‖c‖−ηρχ‖c‖−η + ρI + 1
)−log2
(1+ρχ‖c‖−η+ρL
)+]
= E
[[log2
(1 +
ρα‖c‖−ηρχ‖c‖−η + ρI + 1
)− log2
(1 + ρχ‖c‖−η + ρL
)]1(L<τ(I,‖c‖))
]
= EI,‖c‖
[log2
(1 +
ρα‖c‖−ηρχ‖c‖−η + ρI + 1
)P(L < τ
(I , ‖c‖
))]
− EI,L,‖c‖
[log2
(1 + ρχ‖c‖−η + ρL
)1(L<τ(I,‖c‖))
](a)=
∫ ∞0
∫ αρχ− 1ρ−χy−η
−∞
log2
(1 +
ραy−η
ρχy−η + ρx+ 1
)∫ τ(x,y)
−∞fL(z)
−∫ τ(x,y)
−∞log2
(1 + ρχy−η + ρz
)fL(z) dz
fI(x, y) dx 2λbπye
−λbπy2dy, (C.6)
where 1(·) is the indicator function, and where the upper limit in the inner integration in (a) follows
from 0 ≤ L < τ(I , ‖c‖).
C.5 Proof of Corollary 6.5
Proof. The lower bound in (6.41) can be obtained as follows
RCELL
= E[RCELL
] (a)
≥E
[log2
(1 +
ρα‖c‖−ηρχ‖c‖−η + ρI + 1
)− log2
(1 + ρχ‖c‖−η + ρL
)]+
=
∫ ∞0
∫ ∞−∞
[log2
(1 +
ραy−η
ρχy−η + ρx+ 1
)fI(x, y)
− log2
(1 + ρχy−η + ρx
)fL(x)
]dx 2λbπye
−λbπy2dy
+
(b)=
∫ ∞0
∫ ∞−∞
[F∗1 (φ, y)LI(−i2πφ, y)−F∗2 (φ, y)LL(−i2πφ)
]dφ 2λbπye
−λbπy2dy
+
.
(C.7)
C.5 Appendix for Chapter 6 131
Equation (a) follows from Jensen’s inequality E[x+] ≥ E[x]+. Equation (b) follows by Parse-
val’s theorem RUDIN [1987], and since
F1(φ, y) =sgn(φ) e
2πi(χy−η+ 1ρ
)φ
2φ log 2
(1− e2πiαφy−η
)(C.8)
F2(φ, y) =−e2πi(χy−η+ 1
ρ)φ
log 2
[1
2 |φ| +γ
ρδ(φ)
](C.9)
are the respective Fourier transforms of
f1(x, y) = log2
(1 +
ραy−η
ρχy−η + ρx+ 1
)(C.10)
f2(x, y) = log2
(1 + ρχy−η + ρx
), (C.11)
where γ = limn→∞(∑n
k=11k − log n
)is the Euler-Mascheroni constant. The functions F1(φ, y)
and F2(φ, y) can be obtained from the Fourier transforms of 1x and log |x|, and by applying the
differentiation and shift theorems.
132
Bibliography
AL-NAFFOURI, T., M. SHARIF, AND B. HASSIBI [2009]. “How much does transmit correlation
affect the sum-rate scaling of MIMO Gaussian broadcast channels?” IEEE Trans. Commun.,
vol. 57, no. 2, pp. 562–572.
BACCELLI, F. AND B. BŁASZCZYSZYN [2009]. Stochastic Geometry and Wireless Networks,
Volume I: Theory. Hanover, MA: Now Publishers Inc., 1st ed.
BACCELLI, F. AND P. BREMAUD [2003]. Elements of queueing theory: Palm martingale calculus
and stochastic recurrences. New York, NY: Springer-Verlag.
BAGHERIKARAM, G., A. S. MOTAHARI, AND A. K. KHANDANI [2013]. “The secrecy capacity
region of the Gaussian MIMO broadcast channel.” IEEE Trans. Inf. Theory, vol. 59, no. 5, pp.
2673–2682.
BARROS, J. AND M. RODRIGUES [2006]. “Secrecy capacity of wireless channels.” In Proc. IEEE
Int. Symp. on Inform. Theory (ISIT). pp. 356–360.
BILLINGSLEY, P. [1995]. Probability and measure, 3rd ed. Hoboken, NJ: John Wiley & Sons,
Inc.
BLOCH, M., J. BARROS, M. R. D. RODRIGUES, AND S. W. MCLAUGHLIN [2006]. “An oppor-
tunistic physical-layer approach to secure wireless communications.” In Proc. Allerton Conf.
on Commun., Control, and Computing. Monticello, IL.
BUSTIN, R., R. LIU, H. V. POOR, AND S. SHAMAI [2009]. “An MMSE approach to the se-
crecy capacity of the MIMO Gaussian wiretap channel.” EURASIP Journ. Wireless Commun.
Network.
133
134 Bibliography
CAIRE, G., N. JINDAL, M. KOBAYASHI, AND N. RAVINDRAN [2010]. “Multiuser MIMO
achievable rates with downlink training and channel state feedback.” IEEE Trans. Inf. Theory,
vol. 56, no. 6, pp. 2845–2866.
CAIRE, G. AND S. SHAMAI [2003]. “On the achievable throughput of a multiantenna Gaussian
broadcast channel.” IEEE Trans. Inf. Theory, vol. 49, no. 7, pp. 1691–1706.
COUILLET, R. AND M. DEBBAH [2011]. Random Matrix Theory Methods for Wireless Commu-
nications. Cambridge University Press.
COUILLET, R., F. PASCAL, AND J. W. SILVERSTEIN [2012]. “Robust M-estimation for array
processing: A random matrix approach.” IEEE Trans. Signal Process. Submitted. Available:
http://couillet.romain.perso.sfr.fr/docs/articles/robust est submitted.pdf.
CSISZAR, I. AND J. KORNER [1978]. “Broadcast channels with confidential messages.” IEEE
Trans. Inf. Theory, vol. 24, no. 3, pp. 339–348.
DHILLON, H. S., M. KOUNTOURIS, AND J. G. ANDREWS [2013]. “Downlink MIMO Het-
Nets: Modeling, ordering results and performance analysis.” IEEE Trans. Wireless Commun.
Accepted for publication, available arXiv:1301.5034.
EKREM, E. AND S. ULUKUS [2011]. “The secrecy capacity region of the Gaussian MIMO multi-
receiver wiretap channel.” IEEE Trans. Inf. Theory, vol. 57, no. 4, pp. 2083–2114.
FAKOORIAN, S. A. A. AND A. L. SWINDLEHURST [2013]. “Full rank solutions for the MIMO
Gaussian wiretap channel with an average power constraint.” IEEE Trans. Signal Process.,
vol. 61, no. 10, pp. 2620–2631.
FOSCHINI, G. J. AND M. J. GANS [1998]. “On limits of wireless communications in a fading
environment when using multiple antennas.” Wirel. Pers. Commun., vol. 6, no. 3, pp. 311–335.
GERACI, G., H. S. DHILLON, J. G. ANDREWS, J. YUAN, AND I. B. COLLINGS [2013]. “Phys-
ical layer security in downlink multi-antenna cellular networks.” submitted to IEEE Trans.
Commun. Available arXiv:1307.7211.
GILBERT, E. N. [1961]. “Random plane networks.” J. Soc. Indust. Appl. Math., vol. 9, pp.
533–543.
GOEL, S. AND R. NEGI [2008]. “Guaranteeing secrecy using artificial noise.” IEEE Trans.
Wireless Commun., vol. 7, no. 6, pp. 2180–2189.
Bibliography 135
GOLDSMITH, A. [2005]. Wireless Communications. Cambridge University Press.
GOPALA, P. K., L. LAI, AND H. EL-GAMAL [2008]. “On the secrecy capacity of fading chan-
nels.” IEEE Trans. Inf. Theory, vol. 54, no. 10, pp. 4687–4698.
GRAY, R. M. [2005]. “Toeplitz and circulant matrices: A review.” Foundations and Trends in
Commun. and Inf. Theory, vol. 2, no. 3.
HAENGGI, M. [2005]. “On distances in uniformly random networks.” IEEE Trans. Inf. Theory,
vol. 51, no. 10, pp. 3584–3586.
HAENGGI, M., J. ANDREWS, F. BACCELLI, O. DOUSSE, AND M. FRANCESCHETTI [2009].
“Stochastic geometry and random graphs for the analysis and design of wireless networks.”
IEEE J. Sel. Areas Commun., vol. 27, no. 7, pp. 1029–1046.
HE, X. AND A. YENER [2010]. “Providing secrecy irrespective of eavesdropper’s channel state.”
In Proc. IEEE Global Commun. Conf. (Globecom). pp. 1–5.
HERO, A. O. [2003]. “Secure space-time communication.” IEEE Trans. Inf. Theory, vol. 49,
no. 12, pp. 3235–3249.
HOCHWALD, B. AND S. VISHWANATH [2002]. “Space-time multiple access: Linear growth in
the sum rate.” In Proc. Allerton Conf. on Commun., Control, and Computing. Monticello, IL.
HOCHWALD, B. M., C. B. PEEL, AND A. L. SWINDLEHURST [2005]. “A vector-perturbation
technique for near-capacity multiantenna multiuser communication - Part II: Perturbation.”
IEEE Trans. Commun., vol. 55, no. 5, pp. 537–544.
JINDAL, N. [2006]. “MIMO broadcast channels with finite-rate feedback.” IEEE Trans. Inf.
Theory, vol. 52, no. 11, pp. 5045–5060.
JOHAM, M., W. UTSCHICK, AND J. NOSSEK [2005]. “Linear transmit processing in MIMO
communications systems.” IEEE Trans. Signal Process., vol. 53, no. 8, pp. 2700–2712.
KHISTI, A., A. TCHAMKERTEN, AND G. WORNELL [2008]. “Secure broadcasting over fading
channels.” IEEE Trans. Inf. Theory, vol. 54, no. 6, pp. 2453–2469.
KHISTI, A. AND G. WORNELL [2010]. “Secure transmission with multiple antennas I: The
MISOME wiretap channel.” IEEE Trans. Inf. Theory, vol. 56, no. 7, pp. 3088–3104.
136 Bibliography
KHISTI, A., G. WORNELL, A. WIESEL, AND Y. ELDAR [2007]. “On the Gaussian MIMO
wiretap channel.” In Proc. IEEE Int. Symp. on Inform. Theory (ISIT). pp. 2471–2475.
LEUNG-YAN-CHEONG, S. AND M. HELLMAN [1978]. “The Gaussian wire-tap channel.” IEEE
Trans. Inf. Theory, vol. 24, no. 4, pp. 451–456.
LI, Q., G. LI, W. LEE, M. LEE, D. MAZZARESE, B. CLERCKX, AND Z. LI [2010]. “MIMO
techniques in WiMAX and LTE: a feature overview.” IEEE Comms. Mag., vol. 48, no. 5, pp.
86–92.
LI, Z., W. TRAPPE, AND R. YATES [2007a]. “Secret communication via multi-antenna transmis-
sion.” In Proc. IEEE Conf. on Inf. Sciences and Systems (CISS). pp. 905–910.
LI, Z., R. YATES, AND W. TRAPPE [2007b]. “Secret communication with a fading eavesdropper
channel.” In Proc. IEEE Int. Symp. on Inform. Theory (ISIT). pp. 1296–1300.
LIANG, Y., G. KRAMER, H. V. POOR, AND S. SHAMAI [2009a]. “Compound wiretap channels.”
EURASIP Journ. Wireless Commun. Network.
LIANG, Y., H. V. POOR, AND S. SHAMAI [2009b]. Information Theoretic Security. Dordrecht,
The Netherlands: Now Publisher.
LIM, C., T. YOO, B. CLERCKX, B. LEE, AND B. SHIM [2013]. “Recent trend of multiuser
MIMO in LTE-advanced.” IEEE Comms. Mag., vol. 51, no. 3, pp. 127–135.
LIU, R., T. LIU, H. V. POOR, AND S. SHAMAI [2010a]. “Multiple-input multiple-output Gauss-
ian broadcast channels with confidential messages.” IEEE Trans. Inf. Theory, vol. 56, no. 9, pp.
4215–4227.
—— [2010b]. “A vector generalization of Costa’s entropy-power inequality with applications.”
IEEE Trans. Inf. Theory, vol. 56, no. 4, pp. 1865–1879.
—— [2013]. “New results on multiple-input multiple-output broadcast channels with confidential
messages.” IEEE Trans. Inf. Theory, pp. 1346–1359.
LIU, R., I. MARIC, P. SPASOJEVIC, AND R. D. YATES [2008]. “Discrete memoryless interfer-
ence and broadcast channels with confidential messages: Secrecy rate regions.” IEEE Trans.
Inf. Theory, vol. 54, no. 6, pp. 2493–2507.
Bibliography 137
LIU, R. AND H. POOR [2009]. “Secrecy capacity region of a multiple-antenna Gaussian broadcast
channel with confidential messages.” IEEE Trans. Inf. Theory, vol. 55, no. 3, pp. 1235–1249.
LIU, R. AND W. TRAPPE [2010]. Securing Wireless Communications at the Physical Layer. New
York, NY: Springer Verlag.
LIU, T. AND S. SHAMAI [2009]. “A note on the secrecy capacity of the multiple-antenna wiretap
channel.” IEEE Trans. Inf. Theory, vol. 55, no. 6, pp. 2547–2553.
MARZETTA, T. L., N. JINDAL, AND A. LOZANO [2009]. “What is the value of joint processing
of pilots and data in block-fading channels?” In Proc. IEEE Int. Symp. on Inform. Theory
(ISIT). pp. 2189–2193.
MASSEY, J. [1988]. “An introduction to contemporary cryptology.” Proc. IEEE, vol. 76, no. 5,
pp. 533–549.
MUHARAR, R. AND J. EVANS [2009]. “Downlink beamforming with transmit-side channel cor-
relation: A large system analysis.” Technical Report. Available: http://www.cubinlab.
ee.unimelb.edu.au/˜rmuharar/doc/techRepCorr.pdf.
MUKHERJEE, A., S. A. A. FAKOORIAN, J. HUANG, AND A. L. SWINDLEHURST [2013]. “Prin-
ciples of physical layer security in multiuser wireless networks: A survey.” submitted to IEEE
Commun. Surveys and Tutorials. Available arXiv:1011.3754v2.
NEGI, R. AND S. GOEL [2005]. “Secret communication using artificial noise.” In Proc. of the
IEEE Veh. Tech. Conference (VTC), vol. 3. pp. 1906–1910.
NGUYEN, V. AND J. EVANS [2008]. “Multiuser transmit beamforming via regularized channel
inversion: A large system analysis.” In Proc. IEEE Global Commun. Conf. (Globecom). pp.
1–4.
NGUYEN, V. K., R. MUHARAR, AND J. S. EVANS [2009]. “Multiuser transmit
beamforming via regularized channel inversion: A large system analysis.” Technical
Report. Available: http://cubinlab.ee.unimelb.edu.au/˜rmuharar/doc/
manuscriptrevRusdha22-11.pdf.
OGGIER, F. AND B. HASSIBI [2008]. “The secrecy capacity of the MIMO wiretap channel.” In
Proc. IEEE Int. Symp. on Inform. Theory (ISIT). pp. 524–528.
138 Bibliography
PAPANDRIOPOULOS, J., S. DEY, AND J. EVANS [2008]. “Optimal and distributed protocols for
cross-layer design of physical and transport layers in MANETs.” IEEE/ACM Transactions on
Networking, vol. 16, no. 6, pp. 1392–1405.
PARADA, P. AND R. BLAHUT [2005]. “Secrecy capacity of SIMO and slow fading channels.” In
Proc. IEEE Int. Symp. on Inform. Theory (ISIT). pp. 2152–2155.
PEACOCK, M. J. M., I. B. COLLINGS, AND M. L. HONIG [2008]. “Eigenvalues distributions of
sums and products of large random matrices via incremental matrix expansions.” IEEE Trans.
Inf. Theory, vol. 54, no. 5, pp. 2123–2138.
PEEL, C. B., B. M. HOCHWALD, AND A. L. SWINDLEHURST [2005]. “A vector-perturbation
technique for near-capacity multiantenna multiuser communication - Part I: Channel inversion
and regularization.” IEEE Trans. Commun., vol. 53, no. 1, pp. 195–202.
RAPPAPORT, T. S. [1996]. Wireless Communications: Principles and Practice. IEEE Press, 1st
ed.
RUDIN, W. [1987]. Real and Complex Analysis. New York, NY: McGraw-Hill, 3rd ed.
RYAN, D., I. B. COLLINGS, I. V. L. CLARKSON, AND R. W. HEATH JR. [2008]. “Performance
of vector perturbation multiuser MIMO systems with limited feedback.” IEEE Trans. Commun.,
vol. 57, no. 9, pp. 2633–2644.
SCHNEIER, B. [1998]. “Cryptographic design vulnerabilities.” IEEE Computer, vol. 31, no. 9,
pp. 29–33.
SHAFIEE, S. AND S. ULUKUS [2007]. “Achievable rates in Gaussian MISO channels with secrecy
constraints.” In Proc. IEEE Int. Symp. on Inform. Theory (ISIT). pp. 2466–2470.
SHANNON, C. E. [1949]. “Communication theory of secrecy systems.” Bell System Tech. J.,
vol. 28, pp. 656–715.
SILVERSTEIN, J. W. AND S.-I. CHOI [1995]. “Analysis of the limiting spectral distribution of
large dimensional random matrices.” Journal of Multivariate Analysis, vol. 54, no. 2, pp. 295–
309.
SOUSA, E. AND J. SILVESTER [1990]. “Optimum transmission ranges in a direct-sequence
spread-spectrum multihop packet radio network.” IEEE J. Sel. Areas Commun., vol. 8, no. 5,
pp. 762–771.
Bibliography 139
SPENCER, Q. H., C. B. PEEL, A. L. SWINDLEHURST, AND M. HAARDT [2004a]. “An intro-
duction to the multi-user MIMO downlink.” IEEE Comms. Mag., vol. 42, no. 10, pp. 60–67.
SPENCER, Q. H., A. L. SWINDLEHURST, AND M. HAARDT [2004b]. “Zero-forcing methods
for downlink spatial multiplexing in multiuser MIMO channels.” IEEE Trans. Signal Process.,
vol. 52, no. 2, pp. 461–471.
STOYAN, D., W. KENDALL, AND J. MECKE [1996]. Stochastic geometry and its applications.
New York, NY: John Wiley & Sons Ltd., 2nd ed.
SUNG, C. K. AND I. COLLINGS [2010]. “Cooperative transmission with decode-and-forward
MIMO relaying in multiuser relay networks.” In Proc. IEEE Int. Conf. on Comm. (ICC). pp.
1–5.
SWINDLEHURST, A. L. [2009]. “Fixed SINR solutions for the MIMO wiretap channel.” In Proc.
Internat. Conf. Acoust. Speech Signal Process. pp. 2437–2440.
TELATAR, E. [1999]. “Capacity of multi-antenna Gaussian channels.” European Transactions on
Telecommunications, vol. 10, no. 6, pp. 585–595.
TSE, D. N. C. AND P. VISWANATH [2005]. Fundamentals of Wireless Communication. Cam-
bridge University Press.
VUCETIC, B. AND J. YUAN [2003]. Space-time Coding. John Wiley & Sons.
WAGNER, S., R. COUILLET, M. DEBBAH, AND D. T. M. SLOCK [2012]. “Large system analysis
of linear precoding in correlated MISO broadcast channels under limited feedback.” IEEE
Trans. Inf. Theory, vol. 58, no. 7, pp. 4509–4537.
WYNER, A. D. [1975]. “The wire-tap channel.” Bell System Tech. J., vol. 54, pp. 1355–1387.
XIE, J. AND S. ULUKUS [2013]. “Secure degrees of freedom of the gaussian wiretap channel
with helpers and no eavesdropper csi: Blind cooperative jamming.” In Proc. IEEE Conf. on Inf.
Sciences and Systems (CISS). pp. 1–5.
YANG, N., G. GERACI, J. YUAN, AND R. MALANEY [2013]. “Confidential broadcasting via
linear precoding in non-homogeneous MIMO multiuser networks.” submitted to IEEE Trans.
Commun.
140 Bibliography
YOO, T. AND A. GOLDSMITH [2006]. “On the optimality of multiantenna broadcast scheduling
using zero-forcing beamforming.” IEEE J. Sel. Areas Commun., vol. 24, no. 3, pp. 528–541.
ZHOU, X. AND M. MCKAY [2009]. “Physical layer security with artificial noise: Secrecy capac-
ity and optimal power allocation.” In Proc. IEEE Int. Conf. on Sig. Proc. and Commun. Syst.
(ICSPCS). pp. 1–5.