ieice communications express, vol.4, no.7, 223 227 ... · 4 conclusion a theoretical method for...

Analysis of transmission lineloaded with BCI probe usingcircuit concept approach

Kimitoshi Murano1a), Naoki Takata1, Majid Tayarani2,Fengchao Xiao3, and Yoshio Kami31 School of Engineering, Tokai University,

4–1–1 Kitakaname, Hiratsuka-shi, Kanagawa 259–1292, Japan2 School of Electrical Engineering, Iran University of Science and Technology,

Narmak, Tehran 1684613114, Iran3 Graduate School of Informatics and Engineering, The University of

Electro-Communications, 1–5–1 Chofugaoka, Chofu, Tokyo 182–8585, Japan

a) [email protected]

Abstract: Bulk current injection (BCI) test is adopted as an immunity

testing method for automotive electronic equipment. In this letter, an

analytical method for obtaining the terminal output of the transmission line

excited by using the BCI probe is proposed. The test setup is analytically

solved by using a circuit concept approach because it is considered to be a

transmission line externally excited by an electromagnetic field. To confirm

the validity of the proposed method, a single-ended transmission line loaded

with the BCI probe is considered as an example. The comparison with the

analytical solution and our experimental results shows a good agreement.

Keywords: BCI test, immunity, electromagnetic coupling, transmission

line, circuit concept approach

Classification: Electromagnetic Compatibility (EMC)

References

[1] Road vehicles – Component test methods for electrical disturbances fromnarrowband radiated electromagnetic energy – Part 4: Harness excitationmethods, International Standard, ISO 11452-4, Dec. 2011.

[2] A. Orlandi, G. Antonini, and R. M. Rizzi, “Equivalent circuit model of a bundleof cables for bulk current injection (BCI) test,” IEEE Trans. Electromagn.Compat., vol. 48, no. 4, pp. 701–713, Nov. 2006. DOI:10.1109/TEMC.2006.882850

[3] H. Tanaka, A. Takahashi, Y. Hattori, and M. Izumichi, “A modeling meth-odology for simulation of BCI (bulk current injection) test,” IEICE Trans.Commun., vol. J96-B, no. 4, pp. 458–466, April 2013 (in Japanese).

[4] Y. Kami and R. Sato, “Circuit-concept approach to externally excited transmis-sion lines,” IEEE Trans. Electromagn. Compat., vol. EMC-27, no. 4, pp. 177–183, Nov. 1985. DOI:10.1109/TEMC.1985.304288

[5] N. Takata, Y. Kami, F. Xiao, M. Tayarani, and K. Murano, “Susceptibilitycharacteristics of transmission line in BCI test,” IEICE Tech. Rep., EMCJ2014-84, vol. 114, no. 398, pp. 1–4, Jan. 2015 (in Japanese).© IEICE 2015

DOI: 10.1587/comex.4.223Received May 25, 2015Accepted June 18, 2015Published July 13, 2015

223

IEICE Communications Express, Vol.4, No.7, 223–227

http://dx.doi.org/10.1109/TEMC.2006.882850








1 Introduction

Bulk current injection (BCI) test is usually conducted to evaluate the immunity

characteristics of automotive electronic equipment [1]. In the BCI test, an electric

current is induced in a wire harness connecting to the equipment under test (EUT)

by using a BCI probe in a toroidal coil generating a magnetic field linking the wire

harness. In recent years, some studies on simulation model of the BCI-test system

have been carried out. A circuit model of a bundle of cables in the BCI-test setup

was suggested to predict the induced signals at the terminations in [2], and a

simulation model of BCI-test system was reported to predict the resonant frequen-

cies of the injected current occurring on the wire harness [3].

In contrast, because the BCI-test model can be considered as a kind of

electromagnetic (EM) coupling phenomenon between the external-EM field and

a transmission line, the output of the transmission line can be analytically derived

using a circuit concept approach [4]. If the model can be analytically solved, the

optimum position of the BCI probe on the wire harness can be easily found, and

over or underestimation of the immunity of the EUT can be avoided [5]. This letter

describes a new analytical method of the BCI-test system based on the circuit

concept approach. The method makes it possible to derive the induced current or

voltages at the terminals of the transmission line under all conditions. In this letter,

the proposed method is validated by an experimental result of a single transmission

line which is loaded with a BCI probe.

2 Formulation for BCI test setup

Consider a transmission line of length ‘ externally excited by an EM-plane wave in

Fig. 1(a). By modifying the telegrapher’s equation, a relation between two terminal

outputs on both sides of the transmission line can be expressed as follows [4]:

Vð0ÞIð0Þ

" #�Z ‘

0

FðxÞVfðxÞIfðxÞ

" #dx ¼ Fð‘Þ Vð‘Þ

Ið‘Þ

" #ð1Þ

where V and I are the line voltage and current, and Vf and If are the distributed

voltage and current source on the transmission line expressed as

VfðxÞ ¼ �j!Z h

0

Bezdy IfðxÞ ¼ j!C

Z h

0

Eeydy

where h and C are the line height and the line capacitance, respectively. Vf and If

are equivalently generated by the external-EM fields, Ee and He ¼ Be=�0. F is the

chain matrix of the transmission line as follows:

FðxÞ ¼cos �x jZ0 sin �x

j1

Z0sin �x cos �x

24

35

where β and Z0 are the phase constant and the characteristic impedance of the

transmission line, respectively. In Eq. (1), the 2nd term on the left side indicates the

voltage and current induced by Ee and Be, so that those effects are equivalently

expressed at the starting point x ¼ 0 in Eq. (1). Thus an equivalent circuit cor-

responding to Eq. (1) can be expressed as Fig. 1(b) where the effects of Ee and Be© IEICE 2015DOI: 10.1587/comex.4.223Received May 25, 2015Accepted June 18, 2015Published July 13, 2015

224


are expressed by using the equivalent source Eeq0 and Jeq0. Eeq0 and Jeq0 are as

follows.

Eeq0

Jeq0

" #¼ �

Z ‘

0

FðxÞVfðxÞIfðxÞ

" #dx

Here, we assume that x0 is an arbitrary position on the transmission line.

Multiplying on both sides of Eq. (1) by the unit matrix I ¼ Fðx0ÞF�1ðx0Þ, wefinally obtain the following expression:

Vðx0ÞIðx0Þ

" #�Z ‘

0

Fðx � x0ÞVfðxÞIfðxÞ

" #dx ¼ Fð‘ � x0Þ

Vð‘ÞIð‘Þ

" #: ð2Þ

The equivalent circuit corresponding to Eq. (2) is expressed as Fig. 1(c). Namely,

Eq. (2) shows that the effects of Ee and Be are expressed at the arbitrary position x0

on the transmission line.

In the case of BCI-test setup, the BCI probe locates at an arbitrary point on the

transmission line. The probe generates a magnetic field locally linking the line, so

that its effect can be expressed in an equivalent voltage source EBCIeq at the point as

shown in Fig. 2(a). In this case, Eq. (2) can be written as

Vðx0ÞIðx0Þ

" #� EBCI

eq

0

" #¼ Fð‘ � x0Þ

Vð‘ÞIð‘Þ

" #: ð3Þ

Multiplying both sides of above equation by Fðx0Þ yieldVð0ÞIð0Þ

" #�

EBCIeq0

JBCIeq0

" #¼ Fð‘Þ Vð‘Þ

Ið‘Þ

" #ð4Þ

where

Fig. 1. Externally excited transmission line.

© IEICE 2015DOI: 10.1587/comex.4.223Received May 25, 2015Accepted June 18, 2015Published July 13, 2015

225


EBCIeq0

JBCIeq0

" #¼ EBCI

eq

cos �x0

j1

Z0sin �x0

24

35:

Eq. (4) shows that the equivalent voltage and current sources are expressed at the

terminal x ¼ 0 on the transmission line as shown in Fig. 2(b). The terminal output

can be provided theoretically by solving Eq. (4): the induced currents at both

terminal loads, R0 and R‘, are

Ið0ÞIð‘Þ

" #¼

�R0 �ðR‘ cos �‘ þ jZ0 sin �‘Þ

1 � cos �‘ þ jR‘

Z0sin �‘

� �264

375

�1EBCIeq0

JBCIeq0

" #:

3 Experimental validation

Experimental evaluation of a transmission line loaded with a BCI probe have been

conducted to validate the proposed method. Experimental setup is shown in

Fig. 3(a). The length ‘ and height h of the transmission line examined here are

1m and 40mm, respectively. In this case, the characteristic impedance Z0 is about

222.7Ω. One terminal of the transmission line is terminated with a load of 50Ω,

and the other is connected to port#2 of a network analyzer. In addition, the terminal

of the BCI probe is connected to port#1 of the network analyzer so that the

transmission coefficient jS21j is measured as relative values of the terminal output

for a constant current injected to the transmission line.

Fig. 3(b) shows the relative terminal outputs measured for various frequency

and probe position x0. In this figure, jS21j is normalized by the maximum value for

obtaining the relative output variation. The theoretical results under the same

condition mentioned above are shown in Fig. 3(c). From these results, we can

see that the theoretical results accord well with tendency of Fig. 3(b). Figs. 3(d)

and (e) show a comparison between experimental and theoretical values for the

frequency of 250.25MHz as an example. Both results show that the terminal output

may greatly change by the probe position x0. From these result, it is found that the

terminal output can be expected by using the proposed method. Moreover, it is

indicated that the most suitable probe position where over or underestimation of the

conducted immunity of the EUT is avoided can be found.

Fig. 2. Two kinds of equivalent circuits of transmission line with BCIprobe.


226


4 Conclusion

A theoretical method for analyzing the transmission line loaded with the BCI probe

was proposed. The validity of this method was demonstrated through the experi-

ments using the single-ended transmission line. Since this method is based on the

circuit concept approach, it is applicable to various kinds of transmission lines

under any terminal conditions such as a multi-conductor transmission line with any

load impedance. And it needs to estimate quantitatively the equivalent voltage

source due to the BCI probe because the theoretical results shown here are in

relative value.

Ω

Fig. 3. Measurement of relative terminal output of transmission lineloaded with BCI probe.


227


OpenStack hypervisor,container and Baremetalservers performancecomparison

Yoji Yamatoa)

Software Innovation Center, NTT Corporation,

3–9–11 Midori-cho, Musashino-shi, Tokyo 180–8585, Japan


Abstract: Recently, IaaS services provide not only virtual machines on

hypervisors but also Baremetal servers or container based virtual servers. In

this paper, we measure performances and start up time of Baremetal server,

container servers, virtual machines on OpenStack with virtual server number

changing and evaluate quantitative performances.

Keywords: performance, cloud computing, IaaS, Baremetal, container,

hypervisor, OpenStack

Classification: Network

References

[1] OpenStack web site, http://www.openstack.org/.[2] W. Fester, A. Ferreria, R. Rajamony and J. Rubio, “An updated performance

comparison of virtual machines and Linux containers,” IBM Research Report,July 2014.

[3] B. Russell, “Passive benchmarking with docker LXC, KVM &OpenStack,” slides in http://www.slideshare.net/BodenRussell/kvm-and-docker-lxc-benchmarking-with-openstack, Apr. 2014.

[4] UnixBench web site, https://github.com/kdlucas/byte-unixbench.

1 Introduction

Recently, cloud technology has been progressed and many providers have started

cloud services. To build IaaS systems, many providers adopt open source software

such as OpenStack [1] and CloudStack. NTT group also has started IaaS services

based on OpenStack since 2013.

Currently, many cloud services provide virtual servers to users using virtual

machines deployed on hypervisors such as Xen or KVM. However, hypervisors

have a demerit of much virtualization overhead. Therefore, some providers have

started to provide non-virtualized Baremetal servers (hereafter, Baremetal) or

container based virtual servers which overheads are small (hereafter, Container).© IEICE 2015DOI: 10.1587/comex.4.228Received June 2, 2015Accepted June 30, 2015Published July 28, 2015

228


http://www.openstack.org/



http://www.slideshare.net/BodenRussell/kvm-and-docker-lxc-benchmarking-with-openstack




https://github.com/kdlucas/byte-unixbench

https://github.com/kdlucas/byte-unixbench

It is generally said that Baremetals and Containers show better performances

than virtual machines on hypervisors. However, there are few works to compare

performances and start up time of those three in same conditions and appropriate

usage discussions based on quantitative data are not enough. For example, [2]

compared performances of Baremetal, Docker and KVM but there is no data of

start up time. The work of [3] includes Baremetal, Docker and KVM results of boot

up time, reboot time and other performance data other than Unixbench data.

However, it did not compare 3 types with virtual server number changing.

Therefore, this paper measures performances and start-up time of Baremetal

using Ironic, Containers by Docker and virtual machines on KVM with virtual

server number changing on OpenStack and shows quantitative data. The previous

work of [2] and [3] do not have enough data with virtual server number changing.

2 Outline of Baremetal, Container and Hypervisor

In this section, we compare Baremetal, Container and Hypervisor qualitatively.

Baremetal is a non-virtualized physical server and same as an existing dedi-

cated hosting server. IBM SoftLayer provides Baremetal cloud services adding

characteristics of prompt provisioning and pay-per-use billing to dedicated servers.

In OpenStack, Ironic component provides baremetal provisioning. Because Bare-

metal is a dedicated server, flexibility and performance are high but provisioning

and start-up time are long and it also cannot conduct live migrations.

Containers’ technology is OS virtualization. OpenVZ or FreeBSD jail were

used for VPS (Virtual Private Server) for many years. Computer resources are

isolated with each unit called container but OS kernel is shared among all contain-

ers. Docker which uses LXC (Linux Container) appeared in 2013 and attracted

many users because of its usability. Containers do not have kernel flexibility but a

container creation only needs a process invocation and it takes a short time for start

up. Virtualization overhead is also small. OpenVZ can conduct live migrations but

Docker or LXC cannot conduct live migrations now.

Hypervisors’ technology is hardware virtualization and virtual machines are

behaved on emulated hardware, thus users can customize virtual machine OS

flexibly. Major hypervisors are Xen, KVM and VMware ESX. Virtual machines

have merits of flexible OS and live migrations but those have demerits of perform-

ances and start up time.

Next, we compare performance and start-up time quantitatively.

3 Performance measurement conditions

This paper measures performances and start up time of 3 types servers with same

conditions. We use OpenStack version Juno as a cloud controller, a physical server

provisioned by Ironic as Baremetal, Docker 1.4.1 as a container technology and

KVM/QEMU 2.0.0 as a hypervisor. Ironic, Docker and KVM are de facto standard

software in OpenStack community. Server instances are Ubuntu 14.04 Linux

servers with Apache2 web servers from 10GB image file and we request 3 types

instances provisioning to a same physical server using OpenStack compute

component Nova.© IEICE 2015DOI: 10.1587/comex.4.228Received June 2, 2015Accepted June 30, 2015Published July 28, 2015

229


3.1 Performance measurement items

– Measured servers: Baremetal provisioned by Ironic, Containers based on Docker,

Virtual machines on KVM

– Virtual server number: 1, 2, 3, 4

Only 1 for Baremetal case, 1–4 containers for Docker case and 1–4 virtual

machines for KVM case. When there are plural virtual servers, all physical

resources are equally separated to these plural servers.

– Performance measurement

UnixBench [4] is conducted to acquire UnixBench performance indexes. Note

that UnixBench is a major system performance benchmark.

– Start up time measurement

A time from Nova server instance creation API call to each Linux and Apache2

server start up is measured. For Baremetal case, we measure not only total time but

also each processing time of start up and we also measure the 1st time boot and the

2nd time boot.

3.2 Performance measurement environment

For a performance measurement environment, we prepared 1 physical server on

which 3 types servers were provisioned and 1 physical server which had OpenStack

components (Nova, Ironic, PXE server for Ironic PXE boot and so on). These

servers were connected with Gigabit Ethernet and Layer 2 switch. Fig. 1 shows

each server specification.

4 Performances of Baremetal, Docker and KVM

4.1 UnixBench performance

Fig. 2 shows a performance comparison of 3 types servers. Vertical axis shows

UnixBench performance index value and horizon axis shows each server with

virtual server number changing.

Based on Fig. 2 results, it is clear that Docker containers performance degra-

dation is about 75% performance compared to Baremetal performance. And it is

also said that Docker performance is degraded when we change virtual server

number but it is not inverse proportion. Almost all performances of Docker are

Fig. 1. Performance measurement environment servers

© IEICE 2015DOI: 10.1587/comex.4.228Received June 2, 2015Accepted June 30, 2015Published July 28, 2015

230


better than KVM but file copy performances are worse than KVM, therefore the

total index value is not much higher than KVM. Meanwhile, virtual machines on

KVM performance degradation is more larger and only 60% performance com-

pared to Baremetal performance and KVM performance degradation tendency with

virtual server number change is as same as Docker.

4.2 Start up time

Fig. 3(a) shows start up time of 3 type servers. When virtual servers are plural,

average start up time is showed. Fig. 3(b) shows each processing time of Baremetal

start up for the 1st time boot and the 2nd time boot. From Fig. 3(a), Baremetal start

up takes much long time than KVM and Docker. This is because Baremetal start up

needs image writing for PXE boot for the 1st time boot and it takes long time. For

the 2nd time boot, it does not need image writing and total start up time is about

only 200 sec (see, Fig. 3(b)).

Comparing Docker and KVM, Docker containers start up are shorter than

KVM virtual machines and are less than 15 sec. This is because a virtual machine

start up needs OS boot but a container creation only needs a process invocation.

Precisely, Docker instance creation only takes several hundred msec but OpenStack

processing such as API check, port creation and IP address setting take about 5 sec.

Fig. 2. UnixBench performance index score comparison

Fig. 3. Start-up time comparison. (a) Baremetal, Docker and KVMstart up time. (b) Each processing time of Baremetal start up.


231


4.3 Discussion

Here, we discuss appropriate usages of IaaS servers based on quantitative data.

Because Baremetal shows better performances than other 2 types servers, it is

suitable to use large scale DB processing or real time processing which have

performance problems when we use virtual machines. Containers lack flexibility of

kernel but performance degradation is small and start up time is short. Thus, it is

suitable for auto scaling for existing servers or shared usages of basic services such

as Web or mail. Hypervisors are suitable to use for areas which need system

flexibility such as business applications on specific OS.

5 Conclusion

This paper measured performances and start up time of 3 types IaaS servers;

Baremetal, Docker and KVM with virtual server number changing and showed

quantatitive data. We also studied application areas of each type based on the

rusults. In the future, we plan to enhance IaaS services line up for appropriate use of

3 types servers.


232


Index-based maximumlikelihood RaptorQ codedecoder

Yi-Pin Lua) and Yi-Yao LanGraduate Institute of Electronics Engineering, National Taiwan University,

Taipei City, Taiwan


Abstract: Aiming at the next generation forward error correction code

(FEC) in the application layer, a RaptorQ code decoding algorithm is

proposed in this paper. The proposed index-based decoder significantly

reduces the decoding complexity by tabulates the indices of the non-zero

entries in the sparse code generator matrix. As the number of tabulated

indices is much less than the dimension of the code generator matrix, the

computational complexity is up to ten times lower than direct implementa-

tion of the Raptor code decoder, the previous version of RaptorQ code.

Finally, saving of up to two orders of magnitude in the required memory is

also achieved by the proposed solution.

Keywords: error correction code, fountain code, Raptor code, RaptorQ

code

Classification: Fundamental Theories for Communications

References

[1] M. Luby, A. Shokrollahi, M. Watson, T. Stockhammer, and L. Minder, “RaptorQforward error correction scheme for object delivery,” Internet Engineering TaskForce, RFC6330, http//:tools.ietf.org/html/rfc6330, accessed Jun. 11, 2015.

[2] P. Cataldi, M. P. Shatarski, M. Grangetto, and E. Magli, “Implementation andperformance evaluation of LT and Raptor codes for multimedia applications,”Proc. IEEE Int. Conf. on Intelligent Information Hiding and Multimedia SignalProcessing (IIH-MSP), Pasadena, California, USA, pp. 263–266, Dec. 2006.DOI:10.1109/IIH-MSP.2006.264994

[3] T. Mladenov, S. Nooshabadi, and K. Kim, “Implementation and evaluation ofRaptor codes on embedded systems,” IEEE Trans. Comput., vol. 60, no. 12,pp. 1678–1691, Dec. 2011. DOI:10.1109/TC.2010.210

[4] S. Kim, S. Lee, and S. Y. Chung, “An efficient algorithm for ML decoding ofRaptor codes over the binary erasure channel,” IEEE Commun. Lett., vol. 12,no. 8, pp. 578–580, Aug. 2008. DOI:10.1109/LCOMM.2008.080599

[5] A. Shokrollahi, “Raptor codes,” IEEE Trans. Inf. Theory, vol. 52, no. 6,pp. 2551–2567, June 2006. DOI:10.1109/TIT.2006.874390


233


http://http//:tools.ietf.org/html/rfc6330



http://dx.doi.org/10.1109/IIH-MSP.2006.264994




http://dx.doi.org/10.1109/TC.2010.210

http://dx.doi.org/10.1109/TC.2010.210

http://dx.doi.org/10.1109/TC.2010.210

http://dx.doi.org/10.1109/TC.2010.210

http://dx.doi.org/10.1109/LCOMM.2008.080599




http://dx.doi.org/10.1109/TIT.2006.874390




1 Introduction

RaptorQ code [1], the first standardized forward error correction code (FEC) in the

application layer, has recently drawn much attention from researchers. While

conventional FEC systems need to retransmit the complete coded sequence if the

receiver fails to decode the information sequence, the RaptorQ code decoder retains

the partial coded sequences that are correctly received, until enough newly-

generated coded symbols for reconstructing the information sequence are received.

Once the accumulated correctly-received coded sequence length exceeds the

threshold, satisfactory error rate performance can be guaranteed [1]. However,

the RaptorQ decoder is too complicated owing to the requirement of inversion of a

huge matrix [2]. A matrix inverse with dimension up to 216 is required, accounting

for 92% of the total decoding complexity [3]. Kim et al. [4] reduces the complexity

of such matrix inversion for the Raptor code [5], the previous version of the

RaptorQ code. The Raptor and RaptorQ codes are nearly identical except that some

entries in the RaptorQ code generator matrix are octet, while the Raptor code only

uses binary values to construct its code generator matrix.

In this paper, we propose an index-based maximum likelihood (ML) RaptorQ

code decoder based on the two-stage decoder structure in [4]. However, since the

RaptorQ code generator matrix contains entries with octet values, the complexity of

the RaptorQ decoder is significantly higher than that of the Raptor code. To

mitigate this issue, we first partition these two stages properly so that most

computations remain binary. Then, instead of storing the complete code generator

matrix [3], we propose to tabulate the indices of the nonzero binary entries and then

perform all the corresponding binary operations based on these indices. Owing to

the sparse property of the code generator matrix, the number of the tabulated

indices is much less than the dimension of the code generator matrix, leading to

significant complexity reduction.

From our analysis, the computational complexity of the proposed index-based

ML decoder is only quadratic to the dimension of the code generator matrix, while

the implementation of [4] for RaptorQ code needs cubic complexity. Simulation

results demonstrate that when the code generator matrix is large, only about one-

tenth computational complexity and one-hundredth memory are respectively re-

quired when compared with [4].

2 Preliminary knowledge of the RaptorQ decoder

Before introducing the proposed index-based ML RaptorQ decoder, this section

reviews the concept of ML decoding algorithm for the Raptor code. ML decoding,

as known as full rank decoding, is performed by solving a set of linear equation,

since each coded symbol y is a linear combination of the M � 1 source symbols x

by using

y ¼ A1

A2

" #x; ð1Þ

where A1 is an M �M code generator matrix for x. Then, y is infinitely generated

by the code generator matrix ½A1>A2

>�>, where A2 is generated by a designated© IEICE 2015DOI: 10.1587/comex.4.233Received June 12, 2015Accepted June 30, 2015Published July 28, 2015

234


random generator. By removing the erasures from the coded symbols, the received

symbols are denoted as �y. When the numbers of received symbols reaches M, the

receiver acknowledge the transmitter to stop sending new coded symbols. The

decoder then recover the source symbols by computing the inverse of the corre-

sponding code generator matrix D

x ¼ D�1 �y; ð2Þwhere it will be successful if and only if D is full rank. Since the positions and the

number of erasures are random and depend on transmission, D�1 has to be

computed on-line. For more details, interested readers are referred to [1].

2.1 Gaussian Elimination (GE)

The equation in (2) is commonly solved by GE, a recursive algorithm comprising

the row/column exchanges and row operations. The step-by-step operation of GE

in the ith iteration step is given by

• Pivot row identification: The pivot row possessed the fewest non-zero entries

is identified and exchanged with the ith row vector,

• Pivot column identification: Denote the first non-zero entry in the pivot row

vector (the ith row currently) as the pivot entry and the corresponding column

as the pivot column. Put the pivot entry to the diagonal line by the column

exchange. Normalize the entry values in the pivot row vector by the value of

the pivot entry,

• Forward elimination: Use the pivot row vector (the ith row currently) to null

the pth entries in the pivot column (the ith column currently) by the row

operation, ði þ 1Þ � p � M,

• Backward elimination: Use the pivot row vector to null the qth entries in

the pivot column by the row operation, 1 � q � ði � 1Þ. Increase the value of iby 1.

• If i � M, repeat steps from 1) to 4); otherwise, the GE procedure is terminated.

With the GE procedure, D is diagonalized, leading to the solution of D�1. Sincethe complexity of each iteration in GE is OðM2Þ due to the pivot row identification

and the forward/backward eliminations, the complexity of the GE including M

iterations is OðM3Þ.

2.2 Two-stage Raptor ML decoding algorithm

Instead of using GE, the low-complexity Raptor ML decoder in [4] splits the

diagonalization of D into two stages that respectively adopted a modified GE

(MGE) and GE. MGE is first proposed in the 3GPP standard [1] and tailored for the

Raptor code generator matrix.

MGE is different from GE in that the fourth step backward elimination is

revised to

• Column inactivation: Except the pivot column, all columns with non-zero

entries in the ith position are moved to the rightmost part of the matrix.

Thus, the MGE is simpler than the GE since all the row operations involving

additions and subtractions in backward elimination are substituted by simple

column exchanges. However, please note that instead of diagonalization, MGE© IEICE 2015DOI: 10.1587/comex.4.233Received June 12, 2015Accepted June 30, 2015Published July 28, 2015

235


generates a matrix with non-zero entries located in the diagonal line and the

rightmost columns.

The two-stage Raptor decoder operates as follows: In the first stage, MGE

upper triangluarizes matrix D. The columns shifted to the rightmost part are

grouped and defined as U, whose column number increases as MGE operates, as

shown in Fig. 1(a). The matrix D after the first stage is illustrated in Fig. 1(b),

where the left-upper and left-bottom parts respectively become a diagonal matrix

and a null matrix, while the rightmost part is a dense matrix since many non-zero

entries are included in U by the column inactivation. Then, the dense matrix U is

processed by GE in the second stage so as to fully diagonalize the matrix. By using

the low-complexity MGE to process most of the columns in the first stage, more

than 90% of the overall complexity is reduced [4].

3 Index-based ML RaptorQ decoder

The major difference of the Raptor and RaptorQ codes lies in the fact that some

entries in the RaptorQ code generator matrix are octet. Based on the Raptor decoder

introduced in the previous section, our proposed RaptorQ decoder also consists of

two stages. In the first stage, the row vectors with octet entries are excluded from

the candidates of the pivot rows, in order to make most computations of the MGE

remain binary.

An index table is then constructed to store all non-zero binary entries in D. For

instance, for the row vector [1 1 0 1 …], the indices stored in the table are

[1, 2, 4, …]. Since the column/row exchange and the row operation for binary

entries in the first stage are translated into the update of the index table, the

computations in the first stages are greatly simplified. Specifically, considering the

exchange of the mth and the nth columns, all the binary non-zero entries in the row

vectors can be quickly identified by the index table. Then, only the row vector with

different values of the mth and the nth entries need to update the index in the table

by changing m to n, and vice versa. When the values of the mth and the nth entries

are identical, either zero or one, the index table remains the same. For the exchange

of the row vectors, we only need to interchange their indices stored in the table.

Last, the row operation can be done by first checking which row vector has non-

zero value in the same position as the position of the pivot entry, i.e., the index i in

the ith iteration. Then, these row vectors are updated by removing those indices that

also exist in the pivot row, and adding new indices that only exist in the pivot row.

For example, for the pivot row with tabulated indices [2 4 7] which 2 is the

Fig. 1. A set of two subfigures: (a) Matrix D during the first stage.(b) Matrix D after the first stage.


236


position of the pivot entry, the row vector with tabulated indices [2 4 6] is updated

to [6 7] after the row operation.

Last, due to the sparsity of D, the size of the index table is much smaller than

M2, and the computations row/column exchange and row operation are thus

reduced from processing M entries to only a few indices. Consequently, the

computational complexity of the MGE is reduced from OðM3Þ to OðM2Þ.

4 Simulation results

The computational complexity comparisons are depicted in Fig. 2 by measuring the

average runtime of the Matlab code with the Intel(R) i7 CPU @ 2.93GHz. The

dashed lines are the linear regressions of the numerical results of the decoding

algorithms. We can see that while the two reference algorithms have complexity of

OðM3Þ, the proposed algorithm shows a complexity OðM2Þ. Thus, the complexity

saving of the index-based ML decoder is more pronounced for large M. Only when

M < 1200, the index-based ML decoder is inferior since each row vector stores

different numbers of indices, leading to the overhead of the irregular memory

configuration and access. Nevertheless, this overhead is negligible for large M

where up to ten times computational complexity saving is achieved.

Memory is precious resource in a complicated decoder. We compare the

memory requirement by the three decoder algorithms and show the comparison

results in Fig. 3. Unlike the reference algorithms that use one bit and eight bits to

store all entry values of the code generator matrix, the proposed algorithm keeps

only the indices of nonzero binary entries and thus its memory requirement is

drastically reduced. More than hundred times saving is achieved for the cases with

large sequence length.

Fig. 2. Average runtime vs. sequence length for various decodingalgorithms. The numbers in the boxes are the slopes of theregression lines for various algorithms.


237


5 Conclusion

In this paper, we develop the first decoder for the RaptorQ code with the octet code

generator matrix. By properly partitioning the decoding process, most operations

are performed with binary values. Together with the proposed index table that saves

the positions of the binary entries, our proposed index-based ML RaptorQ decoder

consumes only one-tenth computational complexity and one-hundredth memory,

when compared with the decoders that directly extend the algorithms for Raptor

code decoders.

Acknowledgments

This work was supported in part by Ministry of Science and Technology, Taiwan

(R.O.C.) under Grant no. NSC MOST 103-2221-E-002-088.

Fig. 3. Memory requirement vs. sequence length for various decodingalgorithms.


238


Two-stage SPIT detectionscheme with betweennesscentrality and social trust

Miho Kurataa), Kentaroh Toyodaa), and Iwao SasaseDept. of Information and Computer Science, Keio University,

3–14–1 Hiyoshi, Kohoku, Yokohama, Kanagawa 223–8522, Japan


Abstract: Detecting SPIT (Spam over Internet Telephony) is an urgent

demand in voice communication services. In this paper, we propose a two-

stage SPIT detection scheme using BC (Betweenness Centrality) and social

trust to decrease misdetection of a call from low-frequent users as SPIT. BC

indicates user’s centrality in the entire network and the BC against legitimate

users gradually increases with time even if users seldom call. We first use

BC to identify a call request from a low-frequent user then judge the call

legitimacy by using social trust. By the computer simulation, we show that

our scheme improves the detection accuracy.

Keywords: SPIT detection, VoIP, security, social trust

Classification: Internet

References

[1] A. D. Keromytis, “A comprehensive survey of voice over IP security research,”IEEE Comm. Surv. and Tutor., vol. 14, no. 2, pp. 514–537, 2012. DOI:10.1109/SURV.2011.031611.00112

[2] J. Seedorf, N. D’Heureuse, S. Niccolini, and M. Cornolti, “Detecting trustworthyreal-time communications using a Web-of-Trust,” IEEE GLOBECOM, pp. 1–8,2009. DOI:10.1109/GLOCOM.2009.5425529

[3] T. Kusumoto, E. Y. Chen, and M. Itoh, “Using call patterns to detect unwantedcommunication callers,” IEEE/IPSJ International Symposium on Applicationsand the Internet (SAINT), pp. 64–70, 2009. DOI:10.1109/SAINT.2009.19

[4] M. A. Azad and R. Morla, “Caller-REP: Detecting unwanted calls with callersocial strength,” Comput. Secur., vol. 39, Part B, pp. 219–236, 2013. DOI:10.1016/j.cose.2013.07.006

[5] N. Chaisamran, T. Okuda, and S. Yamaguchi, “Trust-based VoIP spam detectionbased on calling behaviors and human relationships,” J. Inf. Process., vol. 21,no. 2, pp. 188–197, 2013. DOI:10.2197/ipsjjip.21.188

[6] U. Brandes, “A faster algorithm for betweenness centrality,” J. Math. Sociol.,vol. 25, no. 2, pp. 163–177, 2001. DOI:10.1080/0022250X.2001.9990249

1 Introduction

Various voice communication services are getting popular with the growing

smartphone market. However, it is reported that malicious users or companies

© IEICE 2015DOI: 10.1587/comex.4.239Received June 7, 2015Accepted July 2, 2015Published July 31, 2015

239


http://dx.doi.org/10.1109/SURV.2011.031611.00112






http://dx.doi.org/10.1109/GLOCOM.2009.5425529




http://dx.doi.org/10.1109/SAINT.2009.19




http://dx.doi.org/10.1016/j.cose.2013.07.006






http://dx.doi.org/10.2197/ipsjjip.21.188




http://dx.doi.org/10.1080/0022250X.2001.9990249

http://dx.doi.org/10.1080/0022250X.2001.9990249

http://dx.doi.org/10.1080/0022250X.2001.9990249

http://dx.doi.org/10.1080/0022250X.2001.9990249

may abuse them for advertisement or fraud, which is called SPIT (Spam over

Internet Telephony) [1]. Hence, detecting SPIT calls or spammers in voice com-

munication services is an urgent demand for the service providers. In particular, a

social trust-based approach is receiving much attention due to growing SNS-based

voice communication services. A social trust-based approach judges the legitimacy

of a call with a trust value calculated from caller-callee relationships [2, 3, 4, 5]. We

especially pay attention to the scheme [5] since it can correctly classify unknown

users. This scheme uses the call duration as the trust value and the longer a user

calls to a callee, the higher trust value the callee gives to the caller. However we

notice that the scheme [5] raises false alarms for low-frequent legitimate users as

time goes on. That is, the trust value of low-frequent users gradually decreases

since they seldom receive calls.

In this paper, we propose a two-stage SPIT detection scheme with BC

(Betweenness Centrality) and social trust. We use BC to allow a call request from

a low-frequent legitimate user at the first stage. And then we judge the legitimacy

of a call by using the conventional scheme [5]. BC indicates how much a user is

gone through shortest paths between paths among other users. The intuition behind

utilizing BC is that spammers call towards users while they seldom receive calls

and thus spammers tend to be ‘isolated’ at the edge of the entire network. On the

other hand, the value of BC against legitimate users gradually increases with time

even if users seldom call. By the computer simulation, we show the legitimacy of

introducing BC and also clarify that our scheme improves the detection accuracy.

2 System model

We define a voice-based spammer as the attacker model and the aims of spammers

are advertisement, voice phishing, and illegal sales. Spammers call towards ran-

domly chosen users but they seldom receive calls from others. The call frequency is

higher than that of legitimate users. Since the contents of SPIT seems to irritate

ordinary users, the call duration tends to be much shorter than that of legitimate

users.

We assume that a SPIT detection system is deployed in a voice communication

service provider and its task is to judge whether a call request should be established

or not when receiving a call request from a user. We assume that as many as Nuser

users (including both legitimate users and spammers) in the service provider and

the system can access to users’ CDR (Call Detail Records) and buddy lists (friend

lists) for the inspection.

3 Conventional scheme

Chaisamran et al. propose a voice-based SPIT detection scheme with a social trust

[5]. This scheme always allows a call from a user in the callee’s buddy list.

Otherwise, i.e., if a call is from an unknown user, the system judges the legitimacy

of call with an inferred trust value calculated from trust values of other users. Since

multiple paths between an unknown caller u and a callee v may exist, the system

chooses the maximum inferred trust Tu!v as shown in Eq. (1).© IEICE 2015DOI: 10.1587/comex.4.239Received June 7, 2015Accepted July 2, 2015Published July 31, 2015

240


Tu!v ¼ maxp2Pu!v

Tpathpu!v

� �; ð1Þ

where Pu!v denotes a set of paths between u and v and Tpathpu!v indicates an inferred

trust value calculated between users in a path p and is represented as Eq. (2).

Tpathpu!v ¼

Yi2p

TiðtÞ: ð2Þ

In order to make a trust value reliable, each user is assigned a trust value from his/

her friend depending on the cumulative call duration. This will give a low trust

value for spammers since they seldom receive calls. More specifically, a user i in

the path p has its own trust value at time t as Eq. (3).

TiðtÞ ¼ �RiðtÞ þ ð1 � �ÞTiðt�1Þ; ð3Þwhere α denotes a weight variable ð� 2 ½0; 1�Þ and a raw trust value RiðtÞ at time t is

represented as Eq. (4).

RiðtÞ ¼ CvðtÞYn

j¼1 CjðtÞ� �1

n

; ð4Þ

where CjðtÞ denotes the cumulative call duration that a user j calls to user v and n

denotes the number of v’s friends, respectively.

Finally, the system compares the inferred trust value Tu!v and the pre-defined

threshold Tth. If Tu!v > Tth, the system establishes a call request from a user u to v.

Otherwise, it rejects a call request.

3.1 Shortcomings of the conventional scheme

We argue that the scheme [5] raises false alarms for low-frequent legitimate users as

time goes on. That is, the trust value of low-frequent users gradually decreases

since they seldom receive calls.

4 Proposed scheme

Here, we propose a two-stage SPIT detection scheme with BC and social trust in

order for the system to correctly identify a call request from low-frequent legitimate

users as a legitimate one. We use BC as a feature to allow a call from low-frequent

users at the first stage. After that, we judge the legitimacy of a call by using the

social trust-based approach [5].

4.1 Introduction of BC

In graph theory, BC indicates a user’s centrality in the social network [6]. Formally,

BC is defined as the ratio of the number of shortest paths from all users to all others

that pass through that user. Let �st denote the number of shortest paths from s 2 U

to t 2 U, where U denotes a set of users in the entire network. Let �stðuÞ denote thenumber of shortest paths from s 2 U to t 2 U that pass through u 2 U. By using

�st and �stðuÞ, BCðuÞ, which is the BC of a user u, can be represented as Eq. (5).

BCðuÞ ¼X

s≠u≠t2U

�stðuÞ�st

: ð5Þ© IEICE 2015DOI: 10.1587/comex.4.239Received June 7, 2015Accepted July 2, 2015Published July 31, 2015

241


The above description is rigid and difficult to understand why Eq. (5) indicates

the centrality. Hence we calculate BC with an SNS that consists of five users with

Fig. 1. In this example, we calculate two users’ BC, which are u1 and u3 and they

are represented as BCðu1Þ ¼ �2;3ðu1Þ�2;3

þ �2;4ðu1Þ�2;4

þ �2;5ðu1Þ�2;5

þ �3;4ðu1Þ�3;4

þ �3;5ðu1Þ�3;5

þ �4;5ðu1Þ�4;5

¼01þ 0

1þ 0

1þ 0

1þ 0

1þ 0

1¼ 0 and BCðu3Þ ¼ 0

1þ 1

1þ 1

1þ 1

1þ 1

1þ 0

1¼ 4. Therefore,

BCðu3Þ > BCðu1Þ and thus the user u3 is located more central than the user u1.

This matches the fact that the user u1 is located at the edge of the entire network

while the user u3 is located in the center of the network in Fig. 1.

We argue that BC for spammers does not increase. Since spammers call but

seldom receive calls, they tend to be ‘isolated’ at the edge of the entire network and

hardly go through the shortest paths between users. Hence the numerator of Eq. (5)

for spammers does not increase well. On the other hand, BC for legitimate users

gradually increases with time even if users seldom call. This is because legitimate

users gradually make connection with legitimate users and the number of paths that

go through legitimate users increases. Therefore, the numerator in Eq. (5) for

legitimate users gradually increases. Spammers may collude each other to increase

the value of their BC, which is so-called Sybil attack. Although they can “locally”

increase their BC, it is difficult to “globally” increase it. Hence spammers should

account for large part of entire user to succeed the Sybil attack. In reality, spammers

would be less compared to the legitimate users and thus BC for spammers is still

small even if they collude.

4.2 Algorithm

When receiving a call establish request from a caller u, the server first checks

whether the caller u is in the callee’s buddy list. If the caller is in callee’s buddy list,

they are assumed to be friends and thus the system establishes a call request.

Otherwise, the system proceeds to the first detection stage that checks whether his/

her BCðuÞ is bigger than a pre-defined threshold BCth. If BCðuÞ > BCth, the system

judges that the call is legitimate and establishes the call. Otherwise, the system

calculates the inferred trust value Tu!v by using Eqs. (1)–(4) and checks whether

the caller’s inferred social trust from the callee v Tu!v is bigger than a pre-defined

threshold Tth. If Tu!v > Tth, the system judges that the call is legitimate and

establishes a call. Otherwise, the system rejects the call establishment request.

Fig. 1. Toy example of SNS that consists of five users.


242


5 Simulation results

We evaluate the characteristics of BC and the detection accuracy by the computer

simulation. We use the simulation parameters specified in [5] and set the other

undefined parameters as Nuser ¼ 1;000 and BCth ¼ 50.

5.1 Characteristics of BC

Fig. 2(a) shows the boxplot of an average BC values against legitimate callers

and spammers when five months passed. As we can see from Fig. 2(a), BC for

spammers is concentrated around 0 while BC for legitimate callers mostly ranges

from 50 to 200. Fig. 2(b) shows the average BC for legitimate users with time. As

we can see from Fig. 2(b), the BC for legitimate users gradually increases with

time.

5.2 Detection accuracy

Fig. 3(a) and Fig. 3(b) show the true positive rate and false positive rate versus

elapsed time. The true positive rate denotes the ratio of correctly identified calls

from spammers while the false positive rate is the ratio of mistakenly identified

calls from legitimate users, respectively. We first discuss the true positive rate. From

Fig. 3(a), we confirm that our scheme does not degrade the true positive rate. We

then discuss the false positive rate. In Fig. 3(b), the false positive rate against

the conventional scheme is getting worse with time. On the other hand, the false

positive rate of the proposed scheme is within 2% and does not degrade with time.

Fig. 2. Characteristics of BC.

Fig. 3. Detection accuracy.


243


The false positive rate seems to be irrespective of the ratio of spammers. This is

because the scheme judges whether each ‘call request’ (not ‘caller’) is legitimate or

not. From this result, we can say that the false positive rate can be remedied by

using BC.

6 Conclusion

We have pointed out that calls from low-frequent users are gradually identified as

SPIT in the conventional scheme. To remedy this issue, we have proposed a two-

stage SPIT detection scheme with BC and social trust. By the computer simulation,

it is shown that our scheme achieves low false positive rate (< 2%) without

lowering true positive rate.

Acknowledgments

This work is partly supported by the Grant in Aid for Scientific Research

(No. 26420369) from Ministry of Education, Sport, Science and Technology,

Japan.


244


ieice communications express, vol.4, no.7, 223 227 ... · 4 conclusion a theoretical method for...

Documents