Byzantine-Resilient Locally Optimum Detection

Using Collaborative Autonomous Networks

Bhavya Kailkhura, Priyadip Ray, Deepak Rajan, Anton Yen, Peter Barnes, Ryan Goldhahn
Lawrence Livermore National Laboratory

Abstract—In this paper, we propose a locally optimum detec- based on consensus algorithms has been explored in [5]–[10].
tion (LOD) scheme for detecting a weak radioactive source buried However, all these approaches assume a clairvoyant detection
in background clutter. We develop a decentralized algorithm, where all the parameters of the detection system and signal
based on alternating direction method of multipliers (ADMM), model are completely known. Note that, for our application
for implementing the proposed scheme in autonomous sensor of interest (i.e., nuclear radiation detection) the location of
networks. Results show that algorithm performance approaches
the centralized clairvoyant detection algorithm in the low SNR
the radiation sources is rarely known. Centralized approaches
regime, and exhibits excellent convergence rate and scaling manage this challenge by employing composite hypothesis
behavior (w.r.t. number of nodes). We also devise a low-overhead, testing frameworks such as the generalized likelihood ratio test
robust ADMM algorithm for Byzantine-resilient detection, and (GLRT). In GLRT, the detection procedure replaces unknown
demonstrate its robustness to data falsification attacks. parameters in the detection algorithm with their maximum
likelihood estimates, which need multiple sensing intervals for
Keywords—locally optimum detection, data falsification, Byzan- a reasonably accurate parameter estimate. This overhead and
tines, autonomous networks, ADMM
delay is not desirable in nuclear radiation detection problems,
especially under weak signal models. Secondly, due to the
I. I NTRODUCTION non-linearity introduced by the estimation step in GLRT, a
decentralized implementation of GLRT is non-trivial. Finally,
Autonomous vehicles provide sensing platforms which are the implementation of non-linear detectors on low cost UAVs
small, low cost, and maneuverable. Because of size, weight is difficult in practice. Thus, a decentralized solution with a
and power restrictions, the sensors onboard are of limited simple implementation for the radiation detection problem with
performance. Signal processing and data fusion techniques are unknown source location is of utmost interest.
thus needed to approach the performance of a more capable
sensor with a large number of adaptively re-configurable low Autonomous detection schemes are quite vulnerable to
cost sensors. This work provides a computationally tractable different types of attacks. One typical attack on such networks
scheme for autonomous detection, applied to the problem of is a Byzantine (or data falsification) attack [12]–[18]. Few
detecting a radioactive source. attempts have been made to address the Byzantine attacks
in conventional consensus-based detection schemes in recent
Detection of radioactive sources using sensor networks research [19]–[25]. There exist several methods for decentral-
has received significant attention in the literature. In [1], the ized consensus optimization, including distributed subgradient
authors examine the gain in signal-to-noise ratio obtained descent algorithms [26], dual averaging methods [27], and
by a simple combination of data from networked sensors the alternating direction method of multipliers (ADMM) [28].
compared to a single sensor for radioactive source detection. Among these, the ADMM has drawn significant attention,
The costs and benefits of using a network of radiation detectors as it is well suited for distributed convex optimization and
for radioactive source detection are analyzed and evaluated demonstrates fast convergence in many applications. However,
in [2]. In [3], the authors derived a test for the fusion of the performance analysis of ADMM in the presence of data
correlated decisions and obtained optimal sensor thresholds for falsifying Byzantine attacks has thus far not been addressed in
two sensor case. In [4], the authors considered the problem the literature.
of detecting a time-inhomogeneous Poisson process buried
in the recorded background radiation using sensor networks. To overcome the aforementioned challenges, in this paper
However, all these works assume existence of a centralized we propose a simple to implement locally optimum detection
fusion center (FC) to fuse the data from multiple sensors and algorithm to detect radioactive source signal buried in noise.
to make a global decision. We also devise a robust variant of ADMM algorithm to im-
plement this detection scheme in autonomous networks in the
In many scenarios, a centralized FC may not be available. presence of Byzantine attacks. To the best of our knowledge,
Furthermore, due to the distributed nature of future com- there have been no existing results on the Byzantine-resilient
munication networks and various practical constraints (e.g., locally optimum detection in collaborative autonomous sensor
absence of the FC, transmit power or hardware constraints, networks.
and dynamic characteristic of wireless communications), it
may be desirable to achieve collaborative decision making by II. S YSTEM M ODEL
employing peer-to-peer local information exchange to reach a
global decision. Recently, collaborative autonomous detection A. Signal Model

This work was performed under the auspices of the U.S. Department of
Consider two hypotheses H0 (radioactive source is absent)
Energy by Lawrence Livermore National Laboratory under Contract DE- and H1 (radioactive source is present). Also, consider a
AC52-07NA27344. LLNL-CONF-731964 network of N autonomous nodes which must determine which

408
of the two hypotheses is true. The observations received by the most popular tests is the Generalized Likelihood Ratio Testing
node i for i = 1, · · · , N under both hypotheses are as follows. (GLRT). The GLRT has an estimation procedure built into it,
H0 : zi = bi + wi where the underlying parameter estimates are used as a plug-in
H1 : zi = ci + bi + wi (1) estimate for the test statistic. More specifically, the GLRT test
where bi , ci and wi are the background radiation count, source statistic is as follows:
N  
f1 (zi ; λci ) H1
radiation count and measurement noise respectively, at node
X
max log ≷ log λ. (3)
i located at {Xi , Yi }1 . The background radiation count is λci
i=1
f0 (zi ) H0
assumed to be Poisson distributed with known rate parameter Next, we show that in the low signal to noise ratio (SNR)
λb . The source radiation count at node i is assumed to be regime, there exist a locally optimum detection scheme which
Poisson distributed with rate parameter λci . We assume an alleviates the difficulties (e.g., delay, overhead and non-
isotropic behavior of radiation in the presence of the source; linearity) in implementing GLRT in an autonomous setting.
the rate λci is a function of the source intensity Is and distance
of the ith sensor from the source, given by III. C OLLABORATIVE AUTONOMOUS L OCALLY O PTIMUM
Is D ETECTION (CA-LOD)
λci = , (2)
(Xi − Xs ) + (Yi − Ys )2
2
where {Xs , Ys } represent the source coordinates. The mea- For ease of exposition, we first derive the new locally
surement noise wi is Gaussian distributed with a known vari- optimum detection scheme for a centralized scenario. Then,
ance σw 2
. The background radiation count bi and measurement we present an approach to implement the proposed detection
noise wi are assumed to be independent. We also assume that scheme in a decentralized setting.
the observations at any node are conditionally independent
and identically distributed given the hypothesis. It is well A. Locally Optimum Centralized Detection
known that the above signal model can be approximated by Theorem 1: The locally optimal test statistic is given by
the Gaussian distribution [3]. Thus, under H0 , we have XN XN
(zi − λb )2 H1
2 (zi − λb ) + ≷ γ,
f0 (zi ) = N (λb , λb + σw ). 2
(4)
2(λb + σw ) H0
Similarly, under the H1 hypothesis, i=1 i=1
f1 (zi ) = N (λci + λb , λci + λb + σw2
), where γ is chosen such that the probability of false alarm is
where λci is a function of node i’s position relative to source. constrained below a pre-specified level δ.
Proof: The LRT for known λci is given by
N
B. Collaborative Autonomous Detection: Clairvoyant Case
  H1
X f1 (zi ; λci )
log ≷ log λ (5)
For ease of exposition, we first consider the clairvoyant i=1
f0 (zi ) H0

case, i.e., the values of source intensity Is and source co- XN N

X H1
ordinates {xs , ys } are assumed to be known. In our setting, ⇔ log f1 (zi ; λci ) − log f0 (zi ) ≷ log λ (6)
H0
however, the source location is unknown, which is addressed i=1 i=1

in detail in subsequent sections. The collaborative autonomous However, since we are considering a weak signal scenario, λci
tends to zero, and hence linearizing the LRT around λci = 0
detection scheme usually contains three phases: 1) sensing, results in,
2) collaboration, and 3) decision making. In the sensing N
H1
phase, each node acquires the summary statistic about the
X d
(λci − 0) log f1 (zi ; λci )|λci =0 ≷ log λ
phenomenon of interest. Next, in the collaboration phase, each i=1
dλci H0

node communicates with its neighbors to update/improve their XN

d  1 2
state values (summary statistic) and continues with this process ⇔ λci − log(2π(λci + λb + σw ))
dλ ci 2
until the whole network converges to a steady state which is i=1

the global test statistic. Finally, in the decision making phase, (zi − λci − λb )2  H1
− 2
|λci =0 ≷ log λ
nodes make their own decisions about the presence of the 2(λci + λb + σw ) H0
phenomenon using this global test statistic. N N
X X (zi − λb )2 H1
2 N
⇔ (zi − λb ) + ≷ (λb + σw ) log λ + .
The clairvoyant detector is easy to implement in a decen- i=1 i=1
2(λb + σw2)
H0 2
trlized setup using a consensus based approaches [29] and is
the log likelihood ratio test (LRT) given by:
N   The resulting test statistic is independent of the unknown
f1 (zi ) H1
parameter λci , and is the uniformly most powerful (UMP) test
X
log ≷ log λ,
i=1
f0 (zi ) H0 for weak signals.
where λ is chosen such that the probability of false alarm is
constrained below a pre-specified level δ. B. Collaborative Autonomous Detection Using ADMM
The LOD test statistic derived in the previous section is of
C. Detection with Unknown Source Location: GLRT the form below:
N
1 X H1 γ
In many practical scenarios, including the focus of this f (zi ) ≷
work, the location of the radioactive source is not known and N i=1 H0 N
the LRT cannot be implemented. In such scenarios, one of the (zi − λb )2
where f (zi ) = (zi − λb ) + 2)
. The LOD statistic is
1 Note that, the proposed scheme can easily be extended to a three- 2(λb + σw
dimensional setting separable and the function f (zi ) is strongly convex. Next, we

409
 (a) (b) (c) (d)

Fig. 1. (a) Convergence of state values of a network with 10 nodes using ADMM based CA-LOD scheme. (b) Performance comparison of CA-LOD
with clairvoyant LRT. (c) Convergence of vanilla ADMM based CA-LOD in the presence of Byzantines. Blue curve represents Byzantine’s state values. (d)
Susceptibility of CA-LOD to Byzantine attack in terms of ROC.

show that the LOD statistic can be implemented in a distributed autonomous manner. This implies that with these updates, each
manner using ADMM. To apply ADMM, we first formulate a node can learn the global LOD test statistic only by local
convex optimization problem information exchanges.
N
X (x̂ − f (zi ))2 Next, to gain insight into the solution, we present illus-
x∗ = arg min (7)
x̂
i=1
2 trative examples that corroborate our results. We consider a
where the data average is the solution to a least-squares 10 node network employing the ADMM updates as given
minimization problem. Next, we reformulate (7) in the ADMM in (12) to determine the presence (or absence) of a radioactive
amenable form as below source. Source and node locations and adjacency matrix were
XN
(xi − f (zi ))2 generated randomly in a region of interest of dimension
minimize{xi },{yij } (8) 3.0 × 3.0 units. The ADMM parameter ρ was set to 1.0. We
i=1
2 assume a mean background radiation with count λb = 0.5 and
2
subject to xi = yij , xj = yij , ∀(i, j) ∈ A (9) measurement noise with σw = 0.5. We further assume that the
where A is the adjacency matrix, xi is the local copy of prior probability of hypothesis is P0 = P1 = 0.5 and detection
the common optimization variable x̂ at node i and yij is performance is empirically found by performing 1000 Monte-
an auxiliary variable imposing the consensus constraint on Carlo runs.
neighboring nodes i and j. In the matrix form, let us denote
F (x) = 12 kx − f (z)k22 , then, the optimization problem is 1) Convergence Analysis: To better understand the conver-
minimizex,y F (x) + G(y) gence properties of the proposed approach, we next present
subject to Ax + By = 0 (10) an instance of ADMM based CA-LOD in Fig. 1(a). We
where G(y) = 0. Here B = [−I|A| ; −I|A| ] and A = [A1 ; A2 ] assume that each node starts with its local LOD statistic and
collaborate with its neighbors to improve its performance. We
with Ak ∈ R2E×N . If (i, j) ∈ A and yij is the qth entry of y,
plot the updated state values (LOD statistic) at each node as
then the (q, i)th entry of A1 and the (q, j)th entry of A2 are
a function of information exchange iterations. Fig. 1(a) shows
1; otherwise the corresponding entries are 0. The augmented
the state values of each node as a function of the number of
Lagrangian of (10) is given by
ρ iterations. We see that the state values converges to the global
Lρ (x, y, λ) = F (x) + hλ, Ax + Byi + kAx + Byk22 , statistic within 20 iterations using local interactions.
2
where λ = [β1 ; β2 ] with β1 , β2 ∈ R2E is the Lagrange 2) Detection Performance Analysis: Next, we analyze the
multiplier and ρ is a positive algorithm parameter. The updates detection performance of the proposed scheme. In Fig. 1(b),
for ADMM are
we plot steady state receiver operating characteristic (ROC)
x-update : ∇F (xk+1 ) + AT λk + ρAT (Axk+1 + Byk ) = 0,
curves for the proposed CA-LOD approach for different source
y-update : BT λk + ρBT (Axk+1 + Byk+1 ) = 0, intensities Is . We compare the performance of the proposed
λ-update : λk+1 − λk − ρ(Axk+1 + Byk+1 ) = 0, (11) approach with clairvoyant LRT based approach which has
where ∇F (xk+1 ) = xk+1 − f (z) is the gradient of F (.) knowledge of the true source location. For both Is = 0.1 and
at xk+1 . The global convergence of ADMM was established Is = 0.5, the proposed CA-LOD approach performs almost as
in [28]. Since our objective function F (x) is strongly convex good as the clairvoyant LRT based approach.
in x, we obtain x∗ equal to the global test statistic as given
in (4) as the unique solution. IV. C OLLABORATIVE AUTONOMOUS D ETECTION IN THE
The updates in (11) 
can be further simplified to [30],  P RESENCE OF B YZANTINE ATTACKS
1 X k A. Byzantine Attack Model: Modus Operandi
xk+1
i = ρ|Ni |xki + ρ xj − αik + f (zi ) ,
1 + 2ρ|Ni | j∈Ni Note that, the ADMM update at node i at iteration k is
  a function of its neighbors’ parameters {xkj }j∈Ni . Instead of
broadcasting the true parameters {xkj }, some nodes (referred to
X
αik+1 = αik + ρ |Ni |xk+1
i − xk+1
j
 (12)
j∈Ni as Byzantines) can deviate from the prescribed strategies. More
at node i where Ni denotes the set of neighbors of node i. specifically, we assume that the Byzantine node j falsifies its
Note that, the updates in (12) only depend on the data from data at ADMM iteration k as follows:
the neighbors of the node i and can be implemented in a fully xkj = xkj + δjx

410
 (a) (b) (c) (d)

Fig. 2. (a) Convergence of proposed algorithm based CA-LOD in the presence of Byzantine attack. Blue curve represents Byzantines state values. (b) Detection
performance of CA-LOD in the presence of Byzantine attacks. (c) Scaling behavior of the proposed algorithm for bounded neighborhood size. (d) Overhead
comparison in the absence of Byzantine attacks.

where δjx ∼ N (µx , σx2 ). The strength of the attack is charac- 1) Robustness Analysis: In Fig. 2(a), we plot the conver-
terized by (µx , σx2 ). gence of the proposed R-ADMM algorithm with updates as
given in (13). We assume the Byzantine’s parameters to be
1) Performance Analysis of CA-LOD with Byzantines: In µx = 1.5 and σx2 = 0.1. It can be seen that, as opposed to
this section, we study the susceptibility of CA-LOD in the Fig. 1(c), the state values of the honest nodes converge close
presence of Byzantine attacks. We assume that there is only 1 to the global LOD statistic despite the presence of Byzantine
Byzantine in the network which is chosen randomly. attack.
In Fig. 1(c), we plot the convergence of the ADMM
algorithm with updates as given in (12). We assume the Next, in Fig. 2(b), we compare the steady state ROC
Byzantine’s parameters to be µx = 1.5 and σx2 = 0.1. It can for CA-LOD of vanilla ADMM based approach with the R-
be seen that the Byzantine attack can severely degrade the ADMM based approach. We assume attack parameters to be
convergence performance. More specifically, it can be seen µx = 2.5 and σx2 = 0.1. It can be seen that the R-ADMM
from Fig. 1(c) that a single Byzantine can make the rest of based Byzantine-resilient CA-LOD approach performs signif-
the network converge to a state value which is significantly icantly better compare to the vanilla ADMM based approach,
different from the global LOD statistic. which breaks down in the the presence of the Byzantine attack.
Next, in Fig. 1(d), we plot the steady state ROC for 2) Scaling Analysis: In Fig. 2(c), we plot the convergence
different values of attack strength µx keeping σx2 fixed to 0.1. behavior of R-ADMM based CA-LOD as network grows
Observe that, as the attack strength increases, the detection larger. We consider a practical scenario where we fix the
performance degrades severely and an adversary can make the number of nodes (or neighbors) each node can talk to to
steady state statistic (or data) non-informative. In other words, be 10. We plot relative convergence rates defined as T ∗ /N
the optimal detection scheme at each node performs no better where T ∗ is the number of iterations needed to reach within
than a coin flip detector. 95% of the global LOD statistic. Note that, the convergence
rate T ∗ increases as number of nodes N increases in the
B. Robust Collaborative Autonomous Detection using network, however, the relative convergence rate decreases. This
Byzantine-Resilient ADMM implies that the proposed approach retains the excellent scaling
Our approach draws inspiration from robust statistic for properties of ADMM and is amenable for large scale networks.
anomaly detection to make ADMM resilient to Byzantine
attacks. More specifically, we propose the following robust In Fig. 2(d), we compare the overhead caused by the
ADMM algorithm to tolerate at most p Byzantines R-ADMM based CA-LOD scheme. We consider the case
where there is no Byzantine in the network and compare the
1   performance of ADMM based CA-LOD and R-ADMM based
k+1 k k k
xi = ρ|Ni |xi + ρΓp ({xj }j∈Ni ) − αi + f (zi ) , CA-LOD in terms of relative convergence rate. It can be seen
1 + 2ρ|Ni |
  that the overhead caused by the R-ADMM based CA-LOD
αik+1 = αik + ρ |Ni |xk+1
i − Γp ({xk+1
j }j∈Ni ) (13) scheme is very small. In practice, this overhead is dominated
by the sorting step in R-ADMM algorithm and is a constant
where the sum over neighbors’ data in (12) has been replaced for a bounded neighborhood.
by a robust function Γp ({xkj }j∈Ni ) which operates as follows:
Operation of Γp (.): First, sort the elements in S = {xkj }j∈Ni V. C ONCLUSION AND F UTURE W ORK
in a non-decreasing order (breaking ties arbitrarily), and
replace the smallest p values and the largest p values with In this paper, we proposed a decentralized locally optimum
mean of remaining (|Ni | − 2p) values.2 Next, return the sum detection scheme for radioactive source detection. We also de-
of the elements in the new set. vised a robust version of the ADMM algorithm for Byzantine-
Next, we analyze the performance of the proposed resilient detection and demonstrated its robustness to data
Byzantine-resilient autonomous detection scheme in the pres- falsification attacks. There are still many interesting questions
ence of Byzantine attacks. We assume p = 1. that remain to be explored in the future work such as analysis
and extension of the problem with more realistic signal and
2 We assume that |Ni | > 2p, ∀i. communications models and collaborative Byzantine attacks.

411
 R EFERENCES [16] B. Kailkhura, S. Brahma, Y. S. Han, and P. K. Varshney, “Distributed
Detection in Tree Topologies With Byzantines,” IEEE Trans. Signal
[1] S. M. Brennan, A. M. Mielke, and D. C. Torney, “Radioactive source Process., vol. 62, pp. 3208–3219, June 2014.
detection by sensor networks,” IEEE Transactions on Nuclear Science,
vol. 52, no. 3, pp. 813–819, 2005. [17] ——, “Optimal Distributed Detection in the Presence of Byzantines,”
in Proc. The 38th International Conference on Acoustics, Speech, and
[2] D. L. Stephens and A. J. Peurrung, “Detection of moving radioactive
Signal Processing (ICASSP 2013), Vancouver, Canada, May 2013.
sources using sensor networks,” IEEE Transactions on Nuclear Science,
vol. 51, no. 5, pp. 2273–2278, 2004. [18] A. Vempaty, K. Agrawal, H. Chen, and P. K. Varshney, “Adaptive
[3] A. Sundaresan, P. K. Varshney, and N. S. V. Rao, “Distributed detection learning of Byzantines’ behavior in cooperative spectrum sensing,” in
of a nuclear radioactive source using fusion of correlated decisions,” in Proc. IEEE Wireless Comm. and Networking Conf. (WCNC), March
2007 10th International Conference on Information Fusion, July 2007, 2011, pp. 1310 –1315.
pp. 1–7. [19] B. Kailkhura, S. Brahma, and P. K. Varshney, “Data falsification attacks
[4] C. D. Pahlajani, I. Poulakakis, and H. G. Tanner, “Networked decision on consensus-based detection systems,” IEEE Transactions on Signal
making for poisson processes with applications to nuclear detection,” and Information Processing over Networks, vol. 3, no. 1, pp. 145–158,
IEEE Transactions on Automatic Control, vol. 59, no. 1, pp. 193–198, March 2017.
Jan 2014. [20] H. Tang, F. Yu, M. Huang, and Z. Li, “Distributed consensus-based
security mechanisms in cognitive radio mobile ad hoc networks,”
[5] W. Zhang, Z. Wang, Y. Guo, H. Liu, Y. Chen, and J. Mitola, “Distributed Communications, IET, vol. 6, no. 8, pp. 974–983, May 2012.
Cooperative Spectrum Sensing Based on Weighted Average Consensus,”
in Global Telecommunications Conference (GLOBECOM 2011), 2011 [21] F. Yu, H. Tang, M. Huang, Z. Li, and P. Mason, “Defense against
IEEE, Dec 2011, pp. 1–6. spectrum sensing data falsification attacks in mobile ad hoc networks
with cognitive radios,” in Military Communications Conference, 2009.
[6] F. Pasqualetti, A. Bicchi, and F. Bullo, “Consensus Computation in Un-
MILCOM 2009. IEEE, Oct 2009, pp. 1–7.
reliable Networks: A System Theoretic Approach,” Automatic Control,
IEEE Transactions on, vol. 57, no. 1, pp. 90–104, Jan 2012. [22] F. Yu, M. Huang, and H. Tang, “Biologically inspired consensus-based
[7] G. Xiong and S. Kishore, “Consensus-based distributed detection al- spectrum sensing in mobile Ad Hoc networks with cognitive radios,”
gorithm in wireless ad hoc networks,” in Signal Processing and Com- Network, IEEE, vol. 24, no. 3, pp. 26–30, May 2010.
munication Systems, 2009. ICSPCS 2009. 3rd International Conference [23] S. Liu, H. Zhu, S. Li, X. Li, C. Chen, and X. Guan, “An adaptive
on, Sept 2009, pp. 1–6. deviation-tolerant secure scheme for distributed cooperative spectrum
[8] S. Aldosari and J. Moura, “Distributed Detection in Sensor Networks: sensing,” in Global Communications Conference (GLOBECOM), 2012
Connectivity Graph and Small World Networks,” in Signals, Systems IEEE, Dec 2012, pp. 603–608.
and Computers, 2005. Conference Record of the Thirty-Ninth Asilomar [24] Q. Yan, M. Li, T. Jiang, W. Lou, and Y. Hou, “Vulnerability and
Conference on, Oct 2005, pp. 230–234. protection for distributed consensus-based spectrum sensing in cognitive
[9] S. Kar, S. Aldosari, and J. Moura, “Topology for Distributed Inference radio networks,” in INFOCOM, 2012 Proceedings IEEE, March 2012,
on Graphs,” Signal Processing, IEEE Transactions on, vol. 56, no. 6, pp. 900–908.
pp. 2609–2613, June 2008. [25] Z. Li, F. Yu, and M. Huang, “A Distributed Consensus-Based Co-
[10] F. Yu, M. Huang, and H. Tang, “Biologically inspired consensus-based operative Spectrum-Sensing Scheme in Cognitive Radios,” Vehicular
spectrum sensing in mobile Ad Hoc networks with cognitive radios,” Technology, IEEE Transactions on, vol. 59, no. 1, pp. 383–393, Jan
Network, IEEE, vol. 24, no. 3, pp. 26–30, May 2010. 2010.
[11] L. Lamport, R. Shostak, and M. Pease, “The byzantine [26] A. Nedic and A. Ozdaglar, “Distributed subgradient methods for multi-
generals problem,” ACM Trans. Program. Lang. Syst., agent optimization,” IEEE Transactions on Automatic Control, vol. 54,
vol. 4, no. 3, pp. 382–401, Jul. 1982. [Online]. Available: no. 1, pp. 48–61, 2009.
http://doi.acm.org/10.1145/357172.357176
[27] J. C. Duchi, A. Agarwal, and M. J. Wainwright, “Dual averaging for
[12] A. Fragkiadakis, E. Tragos, and I. Askoxylakis, “A survey on security distributed optimization: Convergence analysis and network scaling,”
threats and detection techniques in cognitive radio networks,” IEEE IEEE Transactions on Automatic Control, vol. 57, no. 3, pp. 592–606,
Communications Surveys Tutorials, vol. 15, no. 1, pp. 428–445, 2013. March 2012.
[13] H. Rifà-Pous, M. J. Blasco, and C. Garrigues, “Review of robust
[28] S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, “Distributed
cooperative spectrum sensing techniques for cognitive radio networks,”
optimization and statistical learning via the alternating direction method
Wirel. Pers. Commun., vol. 67, no. 2, pp. 175–198, Nov. 2012.
of multipliers,” Foundations and Trends R in Machine Learning, vol. 3,
[Online]. Available: http://dx.doi.org/10.1007/s11277-011-0372-x
no. 1, pp. 1–122, 2011.
[14] S. Marano, V. Matta, and L. Tong, “Distributed detection in the presence
of byzantine attacks,” IEEE Trans. Signal Process., vol. 57, no. 1, pp. [29] R. Olfati-Saber, J. Fax, and R. Murray, “Consensus and Cooperation in
16 –29, Jan. 2009. Networked Multi-Agent Systems,” Proceedings of the IEEE, vol. 95,
no. 1, pp. 215–233, Jan 2007.
[15] A. Rawat, P. Anand, H. Chen, and P. Varshney, “Collaborative spec-
trum sensing in the presence of byzantine attacks in cognitive radio [30] W. Shi, Q. Ling, K. Yuan, G. Wu, and W. Yin, “On the linear
networks,” IEEE Trans. Signal Process., vol. 59, no. 2, pp. 774 –786, convergence of the admm in decentralized consensus optimization.”
Feb 2011. IEEE Trans. Signal Processing, vol. 62, no. 7, pp. 1750–1761, 2014.

412