Seediscussions,stats,andauthorprofilesforthispublicationat:http://www.researchgate.net/publication/278795446

Efficienteliminationoferroneousnodesincooperativesensingforcognitiveradionetworks

ARTICLEinCOMPUTERS&ELECTRICALENGINEERING·JUNE2015

ImpactFactor:0.82·DOI:10.1016/j.compeleceng.2015.05.004

READS

22

2AUTHORS,INCLUDING:

SeshamSrinu

UniversityofCapeTown

22PUBLICATIONS77CITATIONS

SEEPROFILE

Allin-textreferencesunderlinedinbluearelinkedtopublicationsonResearchGate,

lettingyouaccessandreadthemimmediately.

Availablefrom:SeshamSrinu

Retrievedon:29December2015

Computers and Electrical Engineering xxx (2015) xxx–xxx

Contents lists available at ScienceDirect

Computers and Electrical Engineering

journal homepage: www.elsevier .com/ locate/compeleceng

Efficient elimination of erroneous nodes in cooperative sensingfor cognitive radio networks q

http://dx.doi.org/10.1016/j.compeleceng.2015.05.0040045-7906/� 2015 Elsevier Ltd. All rights reserved.

q Reviews processed and recommended for publication to the Editor-in-Chief by Associate Editor Dr. Sabu Thampi.⇑ Corresponding author.

E-mail addresses: seshamsrinu83@gmail.com (S. Srinu), akmishra@ieee.org (A.K. Mishra).

Please cite this article in press as: Srinu S, Mishra AK. Efficient elimination of erroneous nodes in cooperative sensing for cognitivnetworks. Comput Electr Eng (2015), http://dx.doi.org/10.1016/j.compeleceng.2015.05.004

Sesham Srinu ⇑, Amit Kumar MishraDepartment of Electrical Engineering, University of Cape Town, Cape Town, South Africa

a r t i c l e i n f o a b s t r a c t

Article history:Available online xxxx

Keywords:Cognitive radio networksCooperative spectrum sensingRandom cognitive usersShapiro–Wilk testExtended generalized extreme studentizeddeviateErroneous cognitive user

Cooperative spectrum sensing is a process of achieving spatial diversity gain to make globaldecision for cognitive radio networks. However, accuracy of global decision effects owingto the presence of malicious users/nodes during cooperative sensing. In this work, anextended generalized extreme studentized deviate (EGESD) method is proposed to elimi-nate malicious nodes such as random nodes and selfish nodes in the network. The randomnodes are carried off based on sample covariance of each node decisions on differentframes. Then, the algorithm checks the normality of updated soft data using Shapiro–Wilk test and estimates the expected number of malicious users in cooperative sensing.These are the two essential input parameters required for classical GESD test to eliminatesignificant selfish nodes accurately. Simulation results reveal that the proposed algorithmcan eliminate both random and frequent spectrum sensing data falsification (SSDF) attacksin cooperative sensing and outperforms the existing algorithms.

� 2015 Elsevier Ltd. All rights reserved.

1. Introduction

In cognitive radio networks (CRN), cooperative spectrum sensing (CSS) is an effective technique to combat the multipathfading, shadowing and the receiver uncertainty present in the channel [1,2]. Cooperative sensing is a way of getting spatialdiversity gain by receiving signal from different cognitive users in the vicinity of a fusion center (central node). A wide rangeof fusion techniques have been proposed to achieve spatial diversity gain. All these techniques can be classified into eithersoft decision (EGC and WGC) or hard decision (AND, OR, and MOST) based fusion methods. The classification is based on thetype of data that the central node received from the cooperative cognitive users. Each fusion method has its own pros andcons. In practice, the central node does not have any prior information about signal to noise ratio (SNR) to generate theweights. The soft decision method named as equal gain combining (EGC) assigns equal weight to all nodes to generate aglobal decision. However, it is an unconventional method and degrades the sensing reliability. In weighted gain combining(WGC), different sensing techniques have been reported to estimate the weight for each cognitive radio (CR) user. Most ofthese methods require the signal characteristics a priori. In order to avoid this, differential evolution (DE) optimizationmethod has been considered to estimate the weights for each CR user in [3–5]. Moreover, evolutionary algorithms are betterto consider in the process because of their flexibility to generate proper weights with multiple constraints such as linkbudget, false alarm probability, belief value of each node, and distance of CR from primary user transmitter.

e radio

2 S. Srinu, A.K. Mishra / Computers and Electrical Engineering xxx (2015) xxx–xxx

In practice, most of the current schemes assume that secondary/cognitive users send the correct measurement/decisionto the fusion center (FC) to make the global decision [1,6]. This opens a window for malicious users to use the vacant spec-trum selfishly. The malicious cognitive user (MCR) can send false information and mislead the spectrum sensing machineleading to collision or inefficient spectrum usage. In particular, the performance/reliability of the CSS degrades as the numberof malicious users increases. There are two types of malicious attacks in CSS [7,8]. The first one is the incumbent emulationattack (IEA), where some malicious users know the characteristics of the primary signal and transmit a signal with similarcharacteristics so that other secondary users would believe that a primary user is present [9–11]. The second one is thespectrum sensing data falsification attack (SSDF) also termed as Byzantine attack, where malicious users send false sensinginformation intentionally to a central node [12,13]. The data falsification attacks associated with the malicious/erroneoususers exist in the network is mainly due to malfunction of sensing hardware or/and presence of selfish nodes that intendto use the radio spectrum selfishly.

To defend the basic SSDF attacks, the generalized extreme studentized deviate (GESD) test is the prominent method fordetection and elimination of multiple selfish users in a cognitive radio network [14,2]. It has been also reported that, if themodel follow a normal distribution, GESD test is the best method for multiple erroneous cognitive user elimination [14,15].To detect multiple MCR user, the GESD method requires two essential input parameters a priori, those are, (1) distribution ofthe soft decision data (since the test is more efficient for normally distributed data), (2) estimation of expected number ofmalicious user in the data. A modified largest gap method to estimate the exact number of malicious users or upper limit ofoutliers in the cooperative sensing under attack is presented in [12]. However, GESD test, modified largest gap method, andTietjen-Moore tests can not eliminate the random cognitive nodes in the cooperation.

In the current work, we propose an algorithm named as extended generalized extreme studentized deviate test (EGESD)which can eliminate probable SSDF attacks that comes from failure of sensing hardware and the presence of selfish nodesin the CRNs. To achieve this, we studied and modeled the three possible cases of soft decision data that a hardware failurenode can send, those are, the random data that follow uniform distribution, random data that follow Gaussian distribution(also occurs due to strong fading environment), and the data that is random between ‘always high or low’. The failure nodesare termed as random cognitive radios (RCR). In the case of selfish nodes, we consider two well-known basic attacks termedas ‘always Yes/No’ in the analysis. The proposed algorithm can estimate two essential input parameters required for efficientelimination of erroneous nodes in the network. Hard and soft decision fusion methods are considered to analyze the perfor-mance of cooperative sensing.

Rest of this paper is organized as follows. Cooperative sensing algorithm using DE and EGESD test is presented inSection 2. Simulation results are given in Section 3. Finally, our conclusions are drawn in Section 4.

2. CSS algorithm with elimination of erroneous nodes

2.1. CSS based on DE

Assuming that there are M nodes in the cooperation that contains both genuine and malicious nodes. In addition, receivedsignal of all nodes are statistically independent. Then, the composite hypothesis test can be written as

Pleasenetwo

H0 : rmðnÞ ¼ wmðnÞ; m ¼ 0;1;2 ‘ . . . ’ ðM � 1ÞH1 : rmðnÞ ¼ hmsmðnÞ þwmðnÞ; n ¼ 0;1 ‘ . . . ’ ðN � 1Þ

where H0 and H1 denote the null and alternative hypothesis. The null hypothesis states that there is only noise present in afrequency band to be scanned. The alternative hypothesis states that there is a primary/incumbent user signal present alongwith the noise in the frequency band to be scanned. The received signal sequence by the mth secondary user is denoted asrmðnÞ, whereas smðnÞ is the primary user’s transmitted signal sequence, wmðnÞ is the additive white Gaussian noise (AWGN)observed by mth node, and hm is the channel gain. It is assumed that the channel is slowly varying such that the channelfrequency response or channel gain remains constant during the sensing duration.

In this work, both hard and soft decision logics are considered. In case of WGC fusion, the weight vector is evaluated usingDE algorithm [16]. Mathematically, the problem can be expressed as

maxXM

m¼1

wmHm; s:tXM

m¼1

Hm ¼ 1; 0 < Hm < 1: ð1Þ

In DE algorithm, the sum of the product (soft decision (wm) and its corresponding weight (HmÞ) of all cooperative nodesare considered as the objective function. The notation wm represents the energy measurement of the mth node, given as

wm ¼PN�1

n¼0 jr½n�j2.

Since the DE optimization algorithm is the development of Genetic algorithm, the proposed method generates an optimalweight vector based on the three important steps: mutation, crossover, and selection. Let it be, ðHopt ¼ ½Hoptð1Þ;Hoptð2Þ‘ . . . ’HoptðMÞ�Þ. Then, the cooperative detection probability with the optimal weights can be computed as [16]

Qd�wgcðoptÞ ¼XM

m¼1

wmHoptðmÞ?H1

H0

ke; ð2Þ

where wm is the soft measurement of mth node and ke is the threshold value.

cite this article in press as: Srinu S, Mishra AK. Efficient elimination of erroneous nodes in cooperative sensing for cognitive radiorks. Comput Electr Eng (2015), http://dx.doi.org/10.1016/j.compeleceng.2015.05.004

S. Srinu, A.K. Mishra / Computers and Electrical Engineering xxx (2015) xxx–xxx 3

2.2. Proposed extended GESD test

One of the issues in the cooperative sensing is the presence of few MCR users in the network. Hence, an extended gen-eralized ESD test is proposed to counteract the data falsification attacks generated by malicious CR users. The basic flow chartof introduced algorithm is shown in Fig. 1. Different functionalities are introduced to eliminate all possible basic attacks,which are highlighted and can be seen in Fig. 1. In cooperative sensing model, MCR users are considered along with the gen-uine cognitive radio (GCR) users. Each cognitive user in the network senses the desired frequency band using energy detec-tion algorithm and transmits the soft decision information to the FC. The collected information at FC is considered as theinput to eliminate random cognitive users based on test statistic Sm

cov� �

and threshold value. If the test statistic value is lessthan threshold value, the data is free from random nodes which is considered for normality check using SW test. If the datadoes not follow the normal distribution, the number of iterations is increased until the data follows normal. Once the datadistribution follows normal, the expected number of outliers is estimated using box plot criterion given in Algorithm 1. Then,we apply GESD test to detect and suppress the malicious nodes in the soft decision data to make global decision with genuinenodes.

To achieve this, the soft decision data of all cooperative nodes are accumulated by sensing the frequency band on snumber of frames. The soft decision data set (X) based on s frames can be expressed in matrix form as,

Pleasenetwo

X ¼

wð1;1Þ wð1;2Þ . . . wð1;sÞwð2;1Þ wð2;2Þ . . . wð2;sÞ

..

. ... ..

.

wðM;1Þ wðM;2Þ . . . wðM;sÞ

0BBBBB@

1CCCCCA ð3Þ

Fig. 1. Flow chart of Basic SSDF attacks elimination method.

cite this article in press as: Srinu S, Mishra AK. Efficient elimination of erroneous nodes in cooperative sensing for cognitive radiorks. Comput Electr Eng (2015), http://dx.doi.org/10.1016/j.compeleceng.2015.05.004

4 S. Srinu, A.K. Mishra / Computers and Electrical Engineering xxx (2015) xxx–xxx

Then, we plot the Quantile–Quantile (QQ) plot to visualize the soft data distribution. If it is not normal, then, RCR nodesare eliminated based on sample covariance and mean absolute deviation with the global mean value, as shown in Fig. 2. Thetest statistic and threshold are formulated as

Pleasenetwo

Smcov ¼

1ðs� 1Þ

Xs

a¼1

wðm;aÞ � lm

� �2; m ¼ 1;2 ‘ . . . ’ M ð4Þ

kr ¼1

2Mðs� 1ÞXM

m¼1

Xs

a¼1

wðm;aÞ � Gm

��� ���� E wðm;aÞ � Gm

��� ���� �� �2þ 1

2ðs� 1ÞXs

a¼1

wðm;aÞ � Y��� ���� �2

ð5Þ

where Y ¼ ½wðm;aÞ;wðm;aþsÞ;wðm;aþ2sÞ ‘ . . . ’ wðm;nÞ�, is a set contains the soft decisions of mth node after each s frame.In the case of classical GESD test, it is normally assumed that the number of malicious users is at most ‘u’, can be taken

randomly such that u < M [15,14]. In practice, the information about the normality of the data and number of malicioususers for the GESD test is unknown. Hence, in this work, we have addressed the normality verification usingSwapiro-Wilk (SW) test and estimated the number of malicious users u in the soft decision data. The process of normalitycheck and estimating u to apply GESD test for cooperative sensing can be seen in Fig. 1. The observed data set X contains theelements in the first frame for all cooperative nodes as, fwð1;1Þ;wð2;1Þ‘. . .’ wððM�1Þ;1Þ;wðM;1Þg. Let the soft decision data set on sframes for all cooperative nodes be denoted as X. Then, normality check is verified using SW test using test statistic givenas [17,18],

W ¼ b2

dð6Þ

where d ¼Ph

i¼1ðXðm;aÞ � XÞ2 and b ¼Pj

i¼1aiðXnþ1�i � XiÞ2. The symbols X; h ¼ Ms denote the sample mean and size of X;j isthe index of the median value. ai and p values are the weight and power of the test (p-value), can be taken from SW test-tablewith the prior information of size of the data (h) and test statistic ‘W’ value [18].

The discrimination of data for normality check can be expressed as,

p P as; support N0

p < as; support N1

Fig. 2. Data collection to evaluate sample covariance and threshold.

cite this article in press as: Srinu S, Mishra AK. Efficient elimination of erroneous nodes in cooperative sensing for cognitive radiorks. Comput Electr Eng (2015), http://dx.doi.org/10.1016/j.compeleceng.2015.05.004

S. Srinu, A.K. Mishra / Computers and Electrical Engineering xxx (2015) xxx–xxx 5

where N0 denotes the null hypothesis that the data are distributed normally and N1 denotes the alternative hypothesis. Thesymbol as is the significance level. If the data is distributed normally, it estimates the number of malicious users with thefollowing steps,

Algorithm 1. Estimating upper limit of outliers

1: procedure PLOT THE PROBABILITY DENSITY FUNCTION (HISTOGRAM) OF (X)2: sort the data of X in ascending order3: Find the first and third quartiles, Q1 and Q3.4: Calculate the lower fence P1 ¼ Q1� ns � ðIQRÞ and upper fence P3 ¼ Q3þ ns � ðIQRÞ, where IQR is the inter

Quartile range5: Estimate ‘u’ based on the decisions below the P1 and above the P3

6: end procedure

The single ESD in the data X, can be calculated as

Pleasenetwo

Rj ¼maxi

Xðm;aÞ � X�� ��

SX�dev

( );

j ¼ 1;2 ‘ . . . ’ ui ¼ 1;2 ‘ . . . ’ ðhÞ

ð7Þ

where SX�dev denotes the sample standard deviation of data set X. Eq. (7) is used repeatedly to compute R1;R2‘. . .’Ru andeach of this is individually compared with a critical value vj as follows

vj ¼ðM � jÞtM�j�1;pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

M � j� 1þ t2M�j�1;p

� �ðM � jþ 1Þ

r ð8Þ

where tM�j�1;p is the 100a percentage point from the t-distribution with ðM � j� 1Þ degrees of freedom, and

p ¼ 1� a2ðM�jþ1Þ

n o, where a ¼ 0:05 is the significance level for the overall test. The number of MCR users are determined

by finding the largest ‘j’ such that Rj P vj

The final MCR users can be determined based on the following equality for a fixed value of m [2],

Xs

itr¼1

IðRÞ½ �itr;m ¼ s ðs� 1Þ2

M þm�

ð9Þ

The effective cooperative detection probability for soft decision fusion can be calculated using Eq. (2) with the setE ¼ fx : x 2 X and x R Rmcrg. Where ‘‘Rmcr ’’ is the set of MCR user.

3. Results and discussions

In the simulations, Digital Video Broadcasting-Terrestrial (DVB-T) and Quadrature Phase Shift Keying (QPSK) signals withvariable SNR are considered as a primary user signal, which includes all possible noise impairments of wireless channels. Theenergy detection algorithm is used for local or initial spectrum sensing in each cognitive radio. The detector estimates thetest statistic of the received signal and compares it with the corresponding threshold. The threshold is computed to achieve atolerable false alarm probability (Pf ¼ 0:1). To eliminate random nodes, the sample covariance of each node decisions ondifferent frames are considered. The number of samples in a frame is taken as 256. The number of nodes considered asvariable from 3 to 10. The number of random nodes considered as 3. The simulation parameters are given in Table 1.Owing to the non-existence of closed form solution for Pd, the performance of the detector is analyzed using Monte Carlomethods for 10,000 iterations.

The process of calculating the probability of detection can be seen in Fig. 3. It explains the evaluation of detection prob-ability that can be applied for all kind of detection methods. In practice, two types of procedures are being used to evaluatethe detection accuracy, (i) sequential detection method, (ii) snap shot detection method. In the former case, the hypothesistest is performed on n contiguous frames in a primary user (PU) signal stream. In the latter case, the hypothesis test is per-formed by selecting a single frame of desired length (or a snap shot) from the PU signal stream. Based on the samples in thesnap shot and white Gaussian noise, different frames (n) of desired length can be generated. This is one of the processes ofboot strapping method in Monte Carlo techniques. Then, the detection test has been done for each frame and accumulatedthe decisions. Finally, the detection probability can be computed using the basic probability formula which is ratio betweennumber of times the detector supports alternate hypothesis (H1) over the number of times (n) the test is performed. In thesimulation, cooperative cognitive users are assumed to have configurations as shown in Fig. 2. The performance of the CSS isevaluated with DE and proposed EGESD algorithms.

cite this article in press as: Srinu S, Mishra AK. Efficient elimination of erroneous nodes in cooperative sensing for cognitive radiorks. Comput Electr Eng (2015), http://dx.doi.org/10.1016/j.compeleceng.2015.05.004

Table 1Simulation parameters.

PU signal DVB-T signal, QPSKCarrier frequency, Observed time duration 4.8 MHz, 10–120 lsSample size or frame size (N) 256Number of frames (n) 10,000Number of nodes (M) 3–10Significance level for GESD test (a) 0.05Channel Band width (W) 6 MHzPf ; Pd values 0.1, 0.9RCR nodes 3Always ‘Yes=No’ nodes 3SNR Range, noise uncertainty �20 dB to 0 dB, 0 dB

Fig. 3. Simulation set up for evaluating detection probability.

−2.5 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2 2.50

100

200

300

400

500

600

700

800

900

1000

Standard Normal Quantiles

Qua

ntile

s of

Inpu

t Sam

ple

QQ Plot of Sample Data versus Standard Normal

normal data

random data

Fig. 4. Quantile–Quantile plot of collected soft decision data.

6 S. Srinu, A.K. Mishra / Computers and Electrical Engineering xxx (2015) xxx–xxx

Please cite this article in press as: Srinu S, Mishra AK. Efficient elimination of erroneous nodes in cooperative sensing for cognitive radionetworks. Comput Electr Eng (2015), http://dx.doi.org/10.1016/j.compeleceng.2015.05.004

S. Srinu, A.K. Mishra / Computers and Electrical Engineering xxx (2015) xxx–xxx 7

The QQ plot of collected soft decision data with s ¼ 20 and number of nodes M = 5 can be seen in Fig. 4. It is a graphicalrepresentation of the collected soft decision sample quantiles of (X) vs. theoretical quantiles from a normal distribution. Ifthe distribution of X is normal, the plot will be close to linear. From the figure, it can be observed that the distribution ofsoft decision data (X) approaches to standard normal except few data points on either side due to random data. But, the nor-mality test using QQ plot is a graphical method and is alone not appropriate for the considered problem. Hence, the numer-ical method using SW test is considered for normality verification. From the graph, it is clear that the data is not normal andcontains the data of malicious/random nodes. Hence, there is a need to remove the random data to update the X for efficientcooperative sensing.

Fig. 5 shows the sample variance against number of frames. The variance is large for random cognitive radios (RCRs)compared to genuine cognitive nodes (GCRs). In the simulations, three different random nodes such as nodes that generatedecisions either always Yes or always NO, the nodes decision which follows uniform and Gaussian distributions areconsidered. The SNR at each malicious node is in between �20 dB to 0 dB for random nodes. In the case of ‘always NO’and ‘always YES’ nodes the SNR is varied between �20 dB to 0 dB. The a, significance level is varied between 0.05 (95%)to 0.25 (75%) and the ns value is taken as 2. From the figure it can be seen that the random nodes has higher sample variancecompared to the genuine nodes. Moreover, the random data which follows Gaussian distribution have the higher value dueto more randomness as compared to other distributions.

Fig. 6 illustrates the complementary receiver operating characteristics (CROC) of different nodes considered in the sim-ulation. The genuine CR node has less false alarm probability compared to other random nodes. For instance, at the desired

101 102 103102

103

104

105

106

107

108

No.of. frames

Sam

ple

varia

nce

CR1RCRUniform

RCRGaussian

RCRYes or No

CR2CR3

Fig. 5. Sample covariance of cooperative cognitive radios.

10−4 10−3 10−2 10−1 100

10−4

10−3

10−2

10−1

100

Probability of false alarm

Prob

abili

ty o

f mis

dete

ctio

n

CR Genuine

RCRUniform

RCRGaussian

RCRYes or No

Fig. 6. CROC curves of different RCRs.

Please cite this article in press as: Srinu S, Mishra AK. Efficient elimination of erroneous nodes in cooperative sensing for cognitive radionetworks. Comput Electr Eng (2015), http://dx.doi.org/10.1016/j.compeleceng.2015.05.004

10−4 10−3 10−2 10−1 10010−20

10−15

10−10

10−5

100

Probability of false alarm (Pf)

Prob

abili

ty o

f mis

dete

ctio

n (P

d)M=5,WMCRM=3,WMCRM=5,WOMCRM=3,WOMCR

Fig. 7. Cooperative sensing performance with and without random nodes using hard decision logic.

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Probability of false alarm

Prob

abili

ty o

f mis

dete

ctio

n

EGC,M=3,WMCREGC,M=5,WMCREGC,M=3,WOMCREGC,M=5,WOMCRDEWGC,M=3,WOMCRDEWGC,M=5,WOMCR

Fig. 8. Cooperative sensing performance with and without random nodes.

8 S. Srinu, A.K. Mishra / Computers and Electrical Engineering xxx (2015) xxx–xxx

false alarm rate (0.1), the misdetection probability of GCR is 0.0031 whereas the erroneous nodes have high false alarm prob-ability which is 0.09223, 0.2966, and 0.5258 for nodes that generate decisions either always Yes or always NO, the nodes datawhich follows uniform and Gaussian distributions, respectively. At higher desired performance the sensing performancegoing to be degraded sharply due to random nodes.

Figs. 7 and 8 show the CROC curves for cooperative sensing using hard decision (OR) and soft decision (EGC and WGC)fusion logics. In the simulation, variable number of nodes considered as 3 and 5. Here, cooperation has been done in twoways, without elimination of erroneous nodes (WOMCR) and with elimination of erroneous nodes (WMCR). From the figure,it is clear that the misdetection probability is high when we ignored to eliminate erroneous nodes during CSS. Hence, theelimination must be done before cooperation. From Fig. 8, it can be noted that the proposed method (DEWGC withEGESD) has less misdetection probability compared to EGC fusion logic. For instance, in the case of EGC logic, at false alarmprobability Pf ¼ 0:4, the misdetection probability is 0.94 and 0.81 with M = 3 and 5 with possible attacks of MCR users. Onthe other hand, in same simulation environment, the misdetection probability reduced to 0.85 and 0.41 when we eliminatethe erroneous nodes using proposed method. In the case of weighted gain combining using differential evolution (DEWGC),in same simulation environment, the misdetection probability is further reduced to 0.48 and 0.25 when we eliminate theerroneous nodes using proposed EGESD method. In conclusion, performance of the system is significant even at higher falsealarm rate and the global decision after elimination of malicious user is reliable which is essential for real time application.

Please cite this article in press as: Srinu S, Mishra AK. Efficient elimination of erroneous nodes in cooperative sensing for cognitive radionetworks. Comput Electr Eng (2015), http://dx.doi.org/10.1016/j.compeleceng.2015.05.004

S. Srinu, A.K. Mishra / Computers and Electrical Engineering xxx (2015) xxx–xxx 9

4. Conclusions

In this work, the performance of cooperative spectrum sensing is analyzed by considering the most frequent spectrumsensing data falsification (SSDF) attacks such as random attacks (arise due to failure of sensing hardware) and selfish attacks(due to presence of selfish nodes) in a cognitive radio network. The proposed algorithm based on extended generalizedextreme studentized deviate (EGESD) method can eliminate effects of both random and selfish attacks presents during coop-erative sensing. Moreover, it overcomes the drawbacks of the classical GESD test, which are, estimation of number ofexpected malicious users in the data and prior information of data distribution. From the simulation results, the misdetec-tion probability reduces as we increase the number of nodes in the cooperative sensing. However, owing to the few RCR andselfish nodes misdetection probability increases further in cooperative performance. Hence, random attacks of malicioususer are effectively eliminated using sample covariance and mean absolute deviation methods. The selfish nodes are carriedoff based on Shapiro–Wilk and GESD tests. In conclusion, proposed system can efficiently eliminate all basic attacks com-pared to the classical GESD test because of prior knowledge on distribution and expected number of malicious user in thesoft decision data set. For instance, the misdetection probability of cooperative sensing is reduced to 0.48 to 0.25 by applyingDEWGC and EGESD. This is a significant performance improvement in cooperative sensing over impaired cognitive radio net-works. The applications of proposed method with different detection measurements and for different statistical attacks areunder development.

Acknowledgment

The authors are thankful to the URC/UCT Post-Doctoral Research funding, University of Cape Town, SA, for providing nec-essary support to carry out this research work.

References

[1] Akyildiz IF, Brandon FL, Ravikumar B. Cooperative spectrum sensing in cognitive radio networks: a survey. Phys Commun 2011;4(1):40–62.[2] Srinu S, Sabat SL. Cooperative wideband spectrum sensing in suspicious cognitive radio network. IET Wirel Sensor Syst 2013;3(2):153–61.[3] Axell E, Leus G, Larsson E, Poor H. Spectrum sensing for cognitive radio: state-of-the-art and recent advances. IEEE Signal Process Mag

2012;29(3):101–16.[4] Mishra AK, Johnson DL. In: White space communication. Springer; 2015.[5] Kim J, Andrews J. Sensitive white space detection with spectral covariance sensing. IEEE Trans Wirel Commun 2012;9(9):2945–55.[6] Duan D, Yang L, Jose C. Cooperative diversity of spectrum sensing for cognitive radio systems. IEEE Trans Signal Process 2010;58(6):3218–27.[7] Foukalas FT, Karetsos GT, Merakos L. Spectral efficiency of cognitive radio networks under interference constraint and QoS guarantees. Comput Electr

Eng 2012;38(3):591–602.[8] He X, Dai H, Ning P. HMM-based malicious user detection for robust collaborative spectrum sensing. IEEE J Sel Areas Commun 2013;31(11):2196–208.[9] Haghighat M, Sadough SMS. Cooperative spectrum sensing for cognitive radio networks in the presence of smart malicious users. Int J Electr Commun

2014;68(6):520–7.[10] Min A, Kim KH, Shin K. Robust cooperative sensing via state estimation in cognitive radio networks. In: IEEE symposium on new frontiers in dynamic

spectrum access networks (DySPAN); 2011. p. 185–96.[11] Xie X, Wang W. Detecting primary user emulation attacks in cognitive radio networks via physical layer network coding. Proc Comput Sci

2013;21(0):430–5.[12] Sanket PKS, Kalamkar S, Banerjee A. Block outlier methods for malicious user detection in cooperative spectrum sensing. In: 79th IEEE vehicular

technology conference-spring (VTC-Spring); 2014. p. 1–5.[13] Anand P, Rawat A, Chen H, Varshney P. Collaborative spectrum sensing in the presence of byzantine attacks in cognitive radio networks. In: Second

international conference on communication systems and networks (COMSNETS); 2010. p. 1–9.[14] Yu RC, Teh HW, Jaques PA. Quality control of semi-continuous mobility size-fractionated particle number concentration data. Atmos Environ

2004;38(20):3341–8.[15] Srinu S, Sabat SL. Effective cooperative wideband sensing using energy detection under suspicious cognitive radio network. Comput Electr Eng

2013;39(4):1153–63.[16] Srinu S, Sabat SL. Optimal multinode sensing in a malicious cognitive radio network. IEEE Syst J 2013;PP(99):1–10.[17] Srinu S, Mishra A, Farooq S. Improved gesd test for cooperative sensing over impaired cognitive radio networks. In: 2014 annual IEEE India conference

(INDICON); 2014. p. 1–5.[18] Shapiro SS, Wilk MB. An analysis of variance test for normality (complete samples). Biometrika 1965;52(3/4):591–611.

Sesham Srinu received M.Sc. in Electronics with distinction from Andhra University in 2008. He obtained his Ph.D. degree (Electronics) in 2014 fromHyderabad Central University, India. He is pursuing Post-doctoral research at Department of Electrical Engineering in University of Cape Town, South Africa.His research interests are developing spectrum sensing and radar signal processing algorithms, uncertainty quantification, and hardware implementation.

Amit Kumar MISHRA graduated from REC Rourkela, India in 2001. After working for two years in industry he started his graduation studies in theUniversity of Edinburgh in 2003. He joined Indian Institute of Technology Guwahati, as an Assistant Professor in 2006, where he was an academic till 2011,after which he joined the University of Cape Town. His main areas of interest areRadar information processing and pattern recognition. He is a SeniorMember of IEEE since 2013 and a Y rated researcher in South Africa.

Please cite this article in press as: Srinu S, Mishra AK. Efficient elimination of erroneous nodes in cooperative sensing for cognitive radionetworks. Comput Electr Eng (2015), http://dx.doi.org/10.1016/j.compeleceng.2015.05.004