Distributed Detection and Estimation in Wireless Sensor Networks: Resource Allocation, Fusion Rules, and Network Security

Distributed Detection and Estimation in Wireless Sensor Networks: Resource Allocation, Fusion Rules, and Network Security Edmond Nurellari The University of Leeds, UK School of Electronic and Electrical Engineering In accordance with the requirements for the degree of Doctor of Philosophy June 6, 2017 Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 1 / 42

Overview 1 Introduction Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 2 / 42

Overview 1 Introduction 2 Optimal Quantization and Power Allocation Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 2 / 42

Overview 1 Introduction 2 Optimal Quantization and Power Allocation 3 Centralized Quantized Fusion Rules Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 2 / 42

Overview 1 Introduction 2 Optimal Quantization and Power Allocation 3 Centralized Quantized Fusion Rules 4 Distributed Two-Step Quantized Fusion Rules Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 2 / 42

Overview 1 Introduction 2 Optimal Quantization and Power Allocation 3 Centralized Quantized Fusion Rules 4 Distributed Two-Step Quantized Fusion Rules 5 Sensor Detection in the Presence of Falsified Observations Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 2 / 42

Overview 1 Introduction 2 Optimal Quantization and Power Allocation 3 Centralized Quantized Fusion Rules 4 Distributed Two-Step Quantized Fusion Rules 5 Sensor Detection in the Presence of Falsified Observations 6 Summary Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 2 / 42

Overview 1 Introduction 2 Optimal Quantization and Power Allocation 3 Centralized Quantized Fusion Rules 4 Distributed Two-Step Quantized Fusion Rules 5 Sensor Detection in the Presence of Falsified Observations 6 Summary 7 Key Conclusions Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 2 / 42

Overview 1 Introduction 2 Optimal Quantization and Power Allocation 3 Centralized Quantized Fusion Rules 4 Distributed Two-Step Quantized Fusion Rules 5 Sensor Detection in the Presence of Falsified Observations 6 Summary 7 Key Conclusions Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 3 / 42

1. Introduction Motivation WSNs spatially deployed over a field can be designed to collect information and monitor many phenomena of interest. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 4 / 42

1. Introduction Motivation WSNs spatially deployed over a field can be designed to collect information and monitor many phenomena of interest. Important role in several daily application scenarios such as health-care monitoring, home applications, smart farming, environment monitoring, and military. 1.2. Design Challenges in WSNs Nature... Attacker 1 FUSION CENTER T2 SN2 SN5 T5 SN9 T1 SN1 T4 SN4 SN7 SN8 SN3 SN6 T6 T3 SN10 Figure 1: (left) A WSN architecture. (right) Smart city infrastructure. SN1 active sensor node, manipulated by the attacker Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 4 / 42

1. Introduction Design Challenges in WSNs Low Power Hardware: Clearly, the biggest design constraint in WSNs still remains the power consumption. Even-though the SNs are being designed using low-power micro controllers, their power dissipation is still orders of magnitude too high. Resource Constraints: Battery operated devices with limited on-board energy, both the system lifetime and communication bandwidth (BW) are restricted. Both the signal processing and communication should be carefully designed to consume minimal energy in order to extend the lifetime and improve the overall reliability of the WSN. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 5 / 42

1. Introduction Design Challenges in WSNs Low Power Hardware: Clearly, the biggest design constraint in WSNs still remains the power consumption. Even-though the SNs are being designed using low-power micro controllers, their power dissipation is still orders of magnitude too high. Resource Constraints: Battery operated devices with limited on-board energy, both the system lifetime and communication bandwidth (BW) are restricted. Both the signal processing and communication should be carefully designed to consume minimal energy in order to extend the lifetime and improve the overall reliability of the WSN. Network Security:Usually unattended (geographically dispersed) and this makes them vulnerable to attacks. The overall detection and estimation strongly depends on the reliability of these SNs. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 5 / 42

Contribution-Publications List 1 E. Nurellari, D. McLernon, and M. Ghogho A Secure Optimum Distributed Detection Scheme in Under-Attack Wireless Sensor Networks, in IEEE Trans. on Signal and Information Processing over Networks (TSIPN), April 2017. 2 E. Nurellari, D. McLernon, and M. Ghogho, Distributed Binary Event Detection Under Data-Falsification and Energy-Bandwidth Limitation, in IEEE Sensors Journal, vol. 16, no. 16, pp. 6298-6309, Aug. 15, 2016. 3 E. Nurellari, D. McLernon, and M. Ghogho, Distributed Two-Step Quantized Fusion Rules via Consensus Algorithm for Distributed Detection in Wireless Sensor Networks, in IEEE Transactions on Signal and Information Processing over Networks (TSIPN), vol. 2, no. 3, pp. 321-335, Sept. 2016. 4 S. Aldalahmeh, M. Ghogho, D. McLernon, and E. Nurellari, Optimal fusion rule for distributed detection in clustered wireless sensor networks, EURASIP Journal on Advances in Signal Process., 2016:5, Jan. 2016. 5 E. Nurellari, D. McLernon, and M. Ghogho, Distributed detection in practical wireless sensor networks via a two step consensus algorithm, in Proc. IET Int. conf. on Intelligent Signal Process. (ISP), London, United Kingdom, 1-2 Dec. 2015. 6 E. Nurellari, D. McLernon, M. Ghogho and S. A. R. Zaidi, Distributed Optimal Quantization and Power Allocation for Sensor Detection Via Consensus, Proc. IEEE VTC Spring, Glasgow, U.K., 11-14 May 2015. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 6 / 42

Contribution-Publications List 1 E. Nurellari, D. McLernon, and M. Ghogho A Secure Optimum Distributed Detection Scheme in Under-Attack Wireless Sensor Networks, in IEEE Trans. on Signal and Information Processing over Networks (TSIPN), April 2017. 2 E. Nurellari, D. McLernon, and M. Ghogho, Distributed Binary Event Detection Under Data-Falsification and Energy-Bandwidth Limitation, in IEEE Sensors Journal, vol. 16, no. 16, pp. 6298-6309, Aug. 15, 2016. 3 E. Nurellari, D. McLernon, and M. Ghogho, Distributed Two-Step Quantized Fusion Rules via Consensus Algorithm for Distributed Detection in Wireless Sensor Networks, in IEEE Transactions on Signal and Information Processing over Networks (TSIPN), vol. 2, no. 3, pp. 321-335, Sept. 2016. 4 S. Aldalahmeh, M. Ghogho, D. McLernon, and E. Nurellari, Optimal fusion rule for distributed detection in clustered wireless sensor networks, EURASIP Journal on Advances in Signal Process., 2016:5, Jan. 2016. 5 E. Nurellari, D. McLernon, and M. Ghogho, Distributed detection in practical wireless sensor networks via a two step consensus algorithm, in Proc. IET Int. conf. on Intelligent Signal Process. (ISP), London, United Kingdom, 1-2 Dec. 2015. 6 E. Nurellari, D. McLernon, M. Ghogho and S. A. R. Zaidi, Distributed Optimal Quantization and Power Allocation for Sensor Detection Via Consensus, Proc. IEEE VTC Spring, Glasgow, U.K., 11-14 May 2015. 7 E. Nurellari, S. Aldalahmeh, M. Ghogho, and D. McLernon, Quantized Fusion Rules for Energy-Based Distributed Detection in Wireless Sensor Networks, Proc. IEEE SSPD, Edinburgh, Scotland, 8-9 Sep. 2014. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 6 / 42

Contribution-Publications List 1 E. Nurellari, D. McLernon, and M. Ghogho A Secure Optimum Distributed Detection Scheme in Under-Attack Wireless Sensor Networks, in IEEE Trans. on Signal and Information Processing over Networks (TSIPN), April 2017. 2 E. Nurellari, D. McLernon, and M. Ghogho, Distributed Binary Event Detection Under Data-Falsification and Energy-Bandwidth Limitation, in IEEE Sensors Journal, vol. 16, no. 16, pp. 6298-6309, Aug. 15, 2016. 3 E. Nurellari, D. McLernon, and M. Ghogho, Distributed Two-Step Quantized Fusion Rules via Consensus Algorithm for Distributed Detection in Wireless Sensor Networks, in IEEE Transactions on Signal and Information Processing over Networks (TSIPN), vol. 2, no. 3, pp. 321-335, Sept. 2016. 4 S. Aldalahmeh, M. Ghogho, D. McLernon, and E. Nurellari, Optimal fusion rule for distributed detection in clustered wireless sensor networks, EURASIP Journal on Advances in Signal Process., 2016:5, Jan. 2016. 5 E. Nurellari, D. McLernon, and M. Ghogho, Distributed detection in practical wireless sensor networks via a two step consensus algorithm, in Proc. IET Int. conf. on Intelligent Signal Process. (ISP), London, United Kingdom, 1-2 Dec. 2015. 6 E. Nurellari, D. McLernon, M. Ghogho and S. A. R. Zaidi, Distributed Optimal Quantization and Power Allocation for Sensor Detection Via Consensus, Proc. IEEE VTC Spring, Glasgow, U.K., 11-14 May 2015. 7 E. Nurellari, S. Aldalahmeh, M. Ghogho, and D. McLernon, Quantized Fusion Rules for Energy-Based Distributed Detection in Wireless Sensor Networks, Proc. IEEE SSPD, Edinburgh, Scotland, 8-9 Sep. 2014. 8 E. Nurellari, D. McLernon, M. Ghogho and S. Aldalahmeh, Optimal quantization and power allocation for energy-based distributed sensor detection, Proc. IEEE EUSIPCO, Lisbon, Portugal, 1-5 Sept. 2014. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 6 / 42

Overview 1 Introduction 2 Optimal Quantization and Power Allocation 3 Centralized Quantized Fusion Rules 4 Distributed Two-Step Quantized Fusion Rules 5 Sensor Detection in the Presence of Falsified Observations 6 Summary 7 Key Conclusions Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 7 / 42

2. Optimal Quantization and Power Allocation System Architecture Target SN1 T1 [T1]Q T2 SN2 [T2]Q [T3]Q Fusion Center [T4]Q [T5]Q T3 SN3 T4 SN4 T q f = M α i [T i ] Q i=1 SN5 T5 Figure 2: Communication architecture between peripheral SNs and the FC. Each SN generates a test statistic by observing the target and can communicate with the FC only over an energy-constrained/bandwidth-constrained link. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 8 / 42

2. Simulation Results 1/2 h 2 i 1.5 1 0.5 0 4 Equal weighting in (3.3.4) Optimum weighting in (3.3.4) 1 2 3 4 5 6 7 8 9 10 sensor i p i 2 0 4 1 2 3 4 5 6 7 8 9 10 sensor i L i 2 0 1 2 3 4 5 6 7 8 9 10 sensor i Figure 3: Equal weight (α i = 1 M, i) and optimal weight combining (α = α opt ) transmit power and channel quantization bits allocation for P fa = 0.1, P t = 10, U = 0.1, and M = 10. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 9 / 42

Figure 4: Receiver operating characteristic with P t = 10, U = 0.1 and M = 10 for two different weighting schemes. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 10 / 42 2. Simulation Results 2/2 1 Probability of detection, P d 0.8 0.6 0.4 Optimal weight, N=100 samples 0.2 Optimal weight, N=300 samples Equal weight, N=100 samples Equal weight, N=300 samples 0 0 0.2 0.4 0.6 0.8 Probability of false alarm, P fa 1

Overview 1 Introduction 2 Optimal Quantization and Power Allocation 3 Centralized Quantized Fusion Rules 4 Distributed Two-Step Quantized Fusion Rules 5 Sensor Detection in the Presence of Falsified Observations 6 Summary 7 Key Conclusions Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 11 / 42

Figure 5: Probability of detection (P d ) versus the signal to noise ratio (ξ a ) for M = 20, N = 10, P t = 10, P fa = 0.1 and B = 0.5. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 12 / 42 3. Simulation Results 1/3 Probability of detection. P d 1 0.8 0.6 0.4 0.2 Opt LRT-based LRT-based in (4.4.8) Opt lin comb in (4.4.9) Eq LRT-based Linear combi in (4.3.9) Eq lin combining 0-14 -12-10 -8-6 -4 a (db)

Figure 6: Probability of detection (P d ) versus the number of samples (N) for M = 10 sensors, P fa = 0.1, ξ a = 8.5 db and B = 1. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 13 / 42 3. Simulation Results 2/3 1 Probability of detection, P d 0.8 0.6 0.4 0.2 Opt fusion rule, P t =10 2 Opt linear combining, P t =10 2 Opt fusion rule, P t =10-1 Opt linear combining, P t =10-1 0 0 20 40 60 80 100 120 number of samples, N

Figure 7: Probability of detection (P d ) versus number of sensors (M) for N = 10, P t = 10, P fa = 0.1, ξ a = 8.5 db and B = 0.5. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 14 / 42 3. Simulation Results 3/3 1 Probability of detection, P d 0.9 0.8 0.7 0.6 0.5 0.4 Optimum fusion rule LRT-based LRT-based with weights in (4.4.8) Optimum linear combining in (4.4.9) Equal weight LRT-based Linear combining with weights in (4.3.9) Equal weight linear combining 20 40 60 80 100 number of sensors, M

Overview 1 Introduction 2 Optimal Quantization and Power Allocation 3 Centralized Quantized Fusion Rules 4 Distributed Two-Step Quantized Fusion Rules 5 Sensor Detection in the Presence of Falsified Observations 6 Summary 7 Key Conclusions Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 15 / 42

Figure 8: A distributed communication architecture among peripheral SNs. The SNs have partial connectivity (thin lines) among themselves (i.e., not a complete graph). Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 16 / 42 4. Distributed Two-Step Quantized Fusion Rules Communication Architecture Target SN5 T5 T2 SN2 T1 SN1 T4 SN4 SN3 SN6 T6 T3

4. Quantized Distributed Soft Decision Fusion Rule Proposition Here we propose a scheme, where SN i encodes the data (using a simple uniform quantizer with q i bits) prior to information exchange. 1 We also propose to establish a link between any two SNs i and j based on the (known) SNR at node j, i.e. } if SNR ij < Υ, e ij = e ji = 0 if SNR ij Υ, e ij = e ji = 1. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 17 / 42

4. Quantized Distributed Soft Decision Fusion Rule Proposition Here we propose a scheme, where SN i encodes the data (using a simple uniform quantizer with q i bits) prior to information exchange. 1 We also propose to establish a link between any two SNs i and j based on the (known) SNR at node j, i.e. } if SNR ij < Υ, e ij = e ji = 0 if SNR ij Υ, e ij = e ji = 1. 2 Υ is a SNR threshold parameter and SNR ij defined as: SNR ij = pt ij h2 ij ζ 0 d γ. ij Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 17 / 42

4. Quantized Distributed Soft Decision Fusion Rule Proposition We propose to quantize with q i bits at SN i before transmitting to SN j : q i 1 2 log 2 (1 + Υ) bits/sample Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 18 / 42

4. Quantized Distributed Soft Decision Fusion Rule Proposition We propose to quantize with q i bits at SN i before transmitting to SN j : q i 1 2 log 2 (1 + Υ) bits/sample A large Υ means: 1 Fewer communication links and so slower information diffusion across the network. 2 An increase in the number of bits that each SN can transmit to its neighbors. A small Υ means: 1 Establishes a more connected graph and dictates a faster information diffusion across the network. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 18 / 42

4. Quantized Distributed Soft Decision Fusion Rule Proposition We propose to quantize with q i bits at SN i before transmitting to SN j : q i 1 2 log 2 (1 + Υ) bits/sample A large Υ means: 1 Fewer communication links and so slower information diffusion across the network. 2 An increase in the number of bits that each SN can transmit to its neighbors. A small Υ means: 1 Establishes a more connected graph and dictates a faster information diffusion across the network. 2 Allows less transmission bits resulting in an increase in the quantization noise variance. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 18 / 42

4. Simulation Results 1/6 $ Norm. E # PT P $ d 1 0.5 Proposed two-step Conventional cons. 0 0 10 20 30 40 50 60 70 80 ( 0.74 0.72 0.7 0.68 0 10 20 30 40 50 60 70 80 1 ( ; 0.5 0 0 10 20 30 40 50 60 70 80 ( Figure 9: Normalized average power consumption (E [ ] P T ), achievable 8 probability of detection (Pd ) and the average communication link density (ρ) versus Υ, with σe 2 h = 0, decision fusion in (5.4.16), P g fa = 0.1, U = 3, N = 20, M = 17 and with α i (scaled by M). Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 19 / 42

4. Simulation Results 2/6 Global prob. of detection, P g d 1 0.9 0.8 0.7 0.6 0.5 Upper bound K 1 = 800 K 1 = 500 K 1 = 350 K 1 = 200 K 1 = 100 0.4 Centralized optimum linear rule (5.3.12) 0.3 Proposed weighted two-step, K 1 = 100 Proposed weighted two-step, K 1 = 200 0.2 Proposed weighted two-step, K 1 = 350 Proposed weighted two-step, K 1 = 500 0.1 Proposed weighted two-step, K 1 = 800 Centralized LRT-based [36] 0 0 0.2 0.4 0.6 0.8 1 Global prob. of false alarm, P g fa Figure 10: Averaged (over 500 h 2 ij realizations) ROC for the proposed two-step weighted algorithm with decision fusion in (40), U = 3, N = 20, M = 17, K 2 = 3, Υ = 30, σ 2 e h = 0 and with α i (scaled by M) in (5.3.9). Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 20 / 42

4. Simulation Results 3/6 Global prob. of detection, P g d 1 0.9 0.8 0.7 0.6 0.5 0.4 Upper bound K 1 = 50 K 1 = 150 K 1 = 300 0.3 Proposed weighted two-step, K 1 = 50 0.2 Proposed weighted two-step, K 1 = 150 Proposed weighted two-step, K 1 = 300 0.1 Centralized LRT-based [36] Centralized optimum linear rule (5.3.12) 0 0 0.2 0.4 0.6 0.8 1 Global prob. of false alarm, P g fa Figure 11: Averaged (over 500 h 2 ij realizations) ROC against first step iterations number (K 1 ), with decision fusion in (41), K 2 = 2, U = 3, N = 20, M = 17, Υ = 10, σ 2 e h = 0 and with α i (scaled by M) in (5.3.9). Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 21 / 42

4. Simulation Results 4/6 Global prob. of detection, P g d 0.9 0.8 0.7 Proposed with (5.4.16) 0.6 0.5 0.4 Centr opt linear rule 0.3 0.2 0.1 Proposed with (5.4.15) LRT-based Centr. LRT-based in [36] Centr. opt linear rule (5.3.12) Proposed two-step with (5.4.16) Proposed two-step with (5.4.15) -25-20 -15-10 -5 0 (db) a Global prob. of detection, P g d 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 < 2 e h = 4 < 2 e h = 0 < 2 e h = 1 Centr. LRT-based in [36] Proposed two-step with (5.4.16) Proposed two-step with (5.4.16) Proposed two-step with (5.4.16) -20-15 -10-5 0 (db) a Figure 12: Averaged (over 500 hij 2 realizations) probability of detection (P g d ) against the signal to noise ratio (ξ a) with P g fa = 0.1, U = 3, N = 20, M = 17, K1 = 320, Υ = 20, ξ i = ξ, i in (4) and with α i (scaled by M) in (5.3.9): (left) ideal, σe 2 h = 0; (right) non-ideal, σe 2 h 0. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 22 / 42

Figure 13: Averaged (over 500 h 2 ij realizations) ROC for the proposed (quantized) two-step weighted fusion rule with U = 3, N = 20, Υ = 20, M = 17 and with α i (scaled by M) in (5.3.9). Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 23 / 42 4. Simulation Results 5/6 Global prob. of detection, P g d 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 Centr. opt. K 1 = 320 Centr. eq. comb. K 1 = 320 K 1 = 400 Proposed weighted two-step with (5.4.15) Unquantized eq. comb. (, i = 1) in (5.3.14) Proposed eq. comb. (, i = 1) two-step with (5.4.16) 0.2 Proposed eq. comb. (, i = 1) two-step with (5.4.15) 0.1 Proposed weighted two-step with (5.4.16) Centr. opt. linear rule (5.3.12) 0 0 0.2 0.4 0.6 0.8 1 Global prob. of false alarm, P g fa

4. Simulation Results 6/6 Global prob of detection, P g d 1 0.8 0.6 0.4 Centralized detector K 1 = 10 K 1 = 20 K 1 = 10 0.2 Unquantized eq comb in (5.3.14) Proposed eq comb two-step, K 1 = 10 Proposed eq omb two-step, K 1 = 20 0-14 SN 3 eq comb -rst step, K 1 = 10-12 -10-8 -6-4 9 a (db) -2 Figure 14: Probability of detection (P g d ) versus the signal to noise ratio (ξa) for M = 13, Υ = 72, U = 2, N = 20, P g fa = 0.1 and ξ i = ξ. i in (3.2.4) and α i = 1, i in (5.4.4). The topology used is given in right of Fig. 5.5. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 24 / 42

Overview 1 Introduction 2 Optimal Quantization and Power Allocation 3 Centralized Quantized Fusion Rules 4 Distributed Two-Step Quantized Fusion Rules 5 Sensor Detection in the Presence of Falsified Observations 6 Summary 7 Key Conclusions Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 25 / 42

5. Sensor Detection in the Presence of Falsified Observations Motivation 1 Geographically dispersed to cover large areas, the SNs are constrained in both bandwidth and power. Usually unattended and this makes them vulnerable to different attacks. Contributions Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 26 / 42

5. Sensor Detection in the Presence of Falsified Observations Motivation 1 Geographically dispersed to cover large areas, the SNs are constrained in both bandwidth and power. Usually unattended and this makes them vulnerable to different attacks. 2 The overall detection performance strongly depends on the reliability of these SNs in the network. 3 While fusing the data received by the spatially deployed SNs allows the FC to make a reliable decision, it is possible that one or more SNs (compromised by an attacker) deliberately falsify their local observations. Contributions 1 The problem of centralized detection in the presence of compromised SNs is investigated. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 26 / 42

5. Sensor Detection in the Presence of Falsified Observations Motivation 1 Geographically dispersed to cover large areas, the SNs are constrained in both bandwidth and power. Usually unattended and this makes them vulnerable to different attacks. 2 The overall detection performance strongly depends on the reliability of these SNs in the network. 3 While fusing the data received by the spatially deployed SNs allows the FC to make a reliable decision, it is possible that one or more SNs (compromised by an attacker) deliberately falsify their local observations. Contributions 1 The problem of centralized detection in the presence of compromised SNs is investigated. 2 Attacker-based and FC-based parameter optimization are considered and some expressions have been derived. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 26 / 42

5. Sensor Detection in the Presence of Falsified Observations Motivation 1 Geographically dispersed to cover large areas, the SNs are constrained in both bandwidth and power. Usually unattended and this makes them vulnerable to different attacks. 2 The overall detection performance strongly depends on the reliability of these SNs in the network. 3 While fusing the data received by the spatially deployed SNs allows the FC to make a reliable decision, it is possible that one or more SNs (compromised by an attacker) deliberately falsify their local observations. Contributions 1 The problem of centralized detection in the presence of compromised SNs is investigated. 2 Attacker-based and FC-based parameter optimization are considered and some expressions have been derived. 3 A reputation based scheme to identify the compromised SNs in the network and control their influence to the global FC decision is also proposed. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 26 / 42

5. Sensor Detection in the Presence of Falsified Observations Communication Architecture Target SN1 T1 Tf = M αit q i i=1 SN2 T2 T q 3 T q 2 T q 1 Fusion Center T3 fal SN3 T4 SN4 T q 4 T q 5 T q 6 Attacker SN5 SN6 T6 T5 fal Figure 15: Under attack communication architecture between peripheral SNs and the FC. While the honest SNs test statistics remain unchanged, the compromised SNs falsify their test statistics before transmitting to the FC. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 27 / 42

5. Simulation Results 1/4 2 h i 2 1 0 20 1 2 3 4 5 6 7 8 9 10 11 12 SNs p o i 10 L i 0 5 0 1 2 3 4 5 6 7 8 9 10 11 12 SNs 1 2 3 4 5 6 7 8 9 10 11 12 SNs C=5 C=0.5 C=0 Figure 16: SN optimal transmit power (p o i ) and channel bit allocation (L i) with P t = 60, U = 3, ξ a = 10.5 db, N = 20, β = 0.1 and σ 2 e h = 0. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 28 / 42

5. Simulation Results 2/4 Probability of detection, Pd 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 =0.5 =0.1 AF in [35] opt. in (6.2.22),-= 0:1 OAFBB,-= 0:1 WAFBB,-= 0:1 opt. in (6.2.22),-= 0:5 OAFBB,-= 0:5 WAFBB,-= 0:5 OAFBB,-= 1 WAFBB,-= 1 =1 0 0 0.2 0.4 0.6 0.8 1 Probability of false alarm, P fa Figure 17: Probability of detection (P d ) versus probability of false alarm (P fa ) with U = 3, P t = 60, M = 12, N = 20, C i = 0.9, i and σ 2 e h = 0. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 29 / 42

5. Simulation Results 3/4 Probability of detection, Pd 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 C=1.4 C=1.1 C=0.9 C=0.2 C=0.45 0.98 0.96 0.94 0.92 0.9 0.88 0.86 C=0 C=0.6 C=0 C=0.2 C=0.4 C=0.45 C=0.6 C=0.9 C=1.1 C=1.4 Nash Equilibrium, C=0.4 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1 0 0 0.2 0.4 0.6 0.8 1 Probability of false alarm, P fa Figure 18: Probability of detection (P d ) versus probability of false alarm (P fa ), with U = 3, ξ a = 10.5 db, P t = 60, M = 12, N = 20, β = 0.2, σ 2 e h = 0 and with optimum weights in (6.2.22). Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 30 / 42

5. Simulation Results 4/4 Modi-ed de.ection coe/cient, ~ d 2 3 2.5 2 1.5 1 0.5 0-0.5 Optimum, in (6.2.22) Non-optimum, First derivative -1 0 2 4 6 8 10 Attacker strength, C Figure 19: Modified deflection coefficient ( d 2 ) versus the attacker strength (C) with U = 3, ξ a = 10 db, s i = 0.1, i, P t = 60, M = 12, N = 20, β = 0.1 and σ 2 e h = 0. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 31 / 42

Figure 20: Under attack schematic communication architecture between peripheral SNs and the fusion center (FC). While the i th (i = {1, 2, 4, 6}) honest SN indicator (test statistic) remains unchanged (i.e., Ĩi = I i ), the j th (j = {3, 5}) compromised SN falsify its indicator (test statistic) as in (6.3.7) before transmitting to the FC. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 32 / 42 5. A Secure Sub-optimum Detection Scheme in Under-Attack WSNs Target SN1 I 1 T f = M α i Ĩ i i=1 SN2 I 2 Ĩ 2 Ĩ 3 Ĩ 1 Fusion Center I C 3 SN3 I 4 SN4 Ĩ 4 Ĩ 5 Ĩ 6 Attacker SN5 SN6 I 6 I C 5

5. A Secure Sub-optimum Detection Scheme in Under-Attack WSNs FC Optimum Weighting ( )( 1 β p αopt i i = d pfa) i ( + β p i,c fa p i,c )( d 2P fal C ( )( ( )) 1) 1 β p i d 1 p i d +β (P flip C + ( pi,c d 1 2P flip) )( C 1 P flip C + pi,c d ( (1) 2P flip C 1)). Depends upon the local pfa i and the pi d as well as on the β (fraction of compromised SNs) and the probability of flipping the local decisions by the attacker. The FC cannot implement the optimum weight combining fusion rule Attacker Flipping Probability Optimisation Lemma 6.3.2: The optimum flipping probability ( PC,opt) flip which minimizes the modified deflection coefficient is: ( M ( P flip C,opt = β 1 α i p i d p i ) fa i=1 2β M i=1 α i ( p i,c fa p i,c ) d Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 33 / 42 ) + 1 2 (2)

5. Simulation Results 1/6 Reputation metric, r i 0.09 0.08 0.07 0.06 0.05 0.04 0.03 0.02 SN9 SN10 SN11 SN12 SN13 SN15 Reputation metric, r i 0.35 0.3 0.25 0.2 0.15 0.1 SN21 SN22 SN23 SN24 SN29 SN30 SN32 SN35 SN39 0.01 0.05 0 5 10 15 20 0 5 10 15 20 FC threshold, Λ f FC threshold, Λ f Figure 21: The reliability metric (r i ) versus the FC detection threshold (Λ f ) against the SNs with M = 40, N = 20, β = 0.5, P flip C = 1 and K = 150. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 34 / 42

5. Simulation Results 2/6 37, true d Prob. of det. the compromised SN 37, P 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 K=5 K=10 K=15 K=20 K=30 K=100 K=200 0 0 5 10 15 20 FC detection threshold, f Figure 22: Probability that the (compromised) SN 37 has been truly detected (P 37,true d ) versus the FC detection threshold (Λ f ) with M = 40, N = 20, β = 0.5, P flip = 1 and δ = 0.009. C Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 35 / 42

5. Simulation Results 3/6 Aver. prob. of det. (mis-det.), Ptrue d ( P false d ) 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 β=0.5 β=0.25 β=0.10 P true d, Λ f = 5 P false d, Λ f = 5 P true d, Λ f = 13 P false d, Λ f = 13 P true d, Λ f = 13 P false d, Λ f = 13 5 10 15 20 25 30 35 40 45 50 Time window length, K Figure 23: Average compromised SNs detection probability and honest SNs mis-detection probability versus the time window length (K) and against β with M = 40, N = 20, P flip = 1 and δ = 0.009. C Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 36 / 42

5. Simulation Results 4/6 Prob. of det. minus prob. of false alarm, Pd! Pfa 0.65 0.6 0.55 0.5 0.45 0.4 0.35 f =7 f =9 0.3 Scheme in [73], $ f = 7 Scheme in [73], $ f = 9 0.25 Proposed, $ f = 7 Proposed, $ f = 9 0.2 1 2 3 4 5 6 7 8 9 10 Time window length, K Figure 24: The P d P fa metric versus the time window length (K) against the FC detection threshold (Λ f ) with M = 40, N = 20, β = 0.25, P flip = 0.2, δ = 0.95 and µ = 10. C Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 37 / 42

5. Simulation Results 5/6 1 0.9 0.8 Proposed Prob. of detection, Pd 0.7 0.6 0.5 0.4 0.3 0.2 Upper Bound Equal combining in [35] Perfect SNs iden. and tot. removal Opt. weights (6.3.24), perfect SNs iden. Proposed,7 = 15,/=0:09 Proposed,7 = 15,/=0:12 Proposed,7 = 55,/=0:009 0.1, i =, i AF in (6.3.10), no iden. scheme 0 0 Proposed,7 = 35,/=0:009 No ident. scheme 0.2 0.4 0.6 0.8 Prob. of false alarm, P fa 1 Figure 25: Probability of detection (P d ) versus probability of false alarm (P fa ) with M = 40, N = 20, β = 0.5, P flip C = 1 and K = 5. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 38 / 42

Figure 26: Probability of detection (P d ) versus probability of false alarm (P fa ) against δ and µ with M = 40, N = 20, β = 0.25, and P flip C = 1. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 39 / 42 5. Simulation Results 6/6 Prob. of detection, Pd 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 Eq. comb. No iden. scheme Proposed Upper Bound Opt. weights (6.3.24), perf. SNs iden. Equal combining in [35],-= 0 Proposed,/=0:009, K=80,7 = 6 Proposed,/=0:009, K=40,7 = 6 Proposed,/=0:009, K=5,7 = 10 Proposed,/=0:009, K=5,7 = 14 Proposed,/=0:009, K=5,7 = 1 0.2 With, i =, i AF in (6.3.10), no iden. Scheme in [73] Scheme in [73],/= 1, K = 5 0.1 Equal combining,- = 0:25 0 0 0.2 0.4 0.6 0.8 1 Prob. of false alarm, P fa

Summary We derive the optimum fusion rule and then analyze sub-optimum fusion rules that are realizable and easily implemented in practical WSN deployment scenarios. The effect of fading channels on detection performance is minimized by solving the resource allocation problem. A two-step consensus-based approach with weight combining quantized test statistics exchange is proposed. We relate the communication topology with the number of bits to be shared among SNs. It turns out that there is an optimum topology that maximizes the detection performance. Centralized detection in the presence of compromised SNs is also investigated. Attacker and FC based parameter optimization are considered and some expressions have been derived. A reputation based scheme to identify the compromised SNs in the network and control their influence to the global FC decision is also proposed. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 40 / 42

Key Conclusions Shown that spatially distributed SNs across the field can offer a reliable operation for event detection applications. The system detection performance and the WSN s operating lifetime can be further improved by means of resource allocations, optimisation and signal processing algorithms = complexity to be kept as simple as possible. The data fusion problem: we derive the optimal fusion rules (i.e., for attack-free and under-attack WSN scenarios) and have shown that these fusion rules are not implementable in practice and require complex local signal processing = Derive sub-optimum but simple fusion rules (requiring simple hardware) that offer reliable and good detection performance. A better but more complex approach is to possibly identify these compromised SNs and control their influence on the FC decision = Offers an improved detection performance but requires observing the SN s local reports for a period of time. A larger observation time period (K) may lead to a large detection delay that is critical for most of the event detection applications. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 41 / 42

Key Conclusions Shown that spatially distributed SNs across the field can offer a reliable operation for event detection applications. The system detection performance and the WSN s operating lifetime can be further improved by means of resource allocations, optimisation and signal processing algorithms = complexity to be kept as simple as possible. The data fusion problem: we derive the optimal fusion rules (i.e., for attack-free and under-attack WSN scenarios) and have shown that these fusion rules are not implementable in practice and require complex local signal processing = Derive sub-optimum but simple fusion rules (requiring simple hardware) that offer reliable and good detection performance. A better but more complex approach is to possibly identify these compromised SNs and control their influence on the FC decision = Offers an improved detection performance but requires observing the SN s local reports for a period of time. A larger observation time period (K) may lead to a large detection delay that is critical for most of the event detection applications. We have addressed the fully distributed detection problem and proposed signal processing algorithms for such an approach Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 41 / 42

Key Conclusions Shown that spatially distributed SNs across the field can offer a reliable operation for event detection applications. The system detection performance and the WSN s operating lifetime can be further improved by means of resource allocations, optimisation and signal processing algorithms = complexity to be kept as simple as possible. The data fusion problem: we derive the optimal fusion rules (i.e., for attack-free and under-attack WSN scenarios) and have shown that these fusion rules are not implementable in practice and require complex local signal processing = Derive sub-optimum but simple fusion rules (requiring simple hardware) that offer reliable and good detection performance. A better but more complex approach is to possibly identify these compromised SNs and control their influence on the FC decision = Offers an improved detection performance but requires observing the SN s local reports for a period of time. A larger observation time period (K) may lead to a large detection delay that is critical for most of the event detection applications. We have addressed the fully distributed detection problem and proposed signal processing algorithms for such an approach = Very attractive from both the signal processing perspective and the communication point of view. Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 41 / 42

Questions/Comments Edmond Nurellari (University of Leeds) Distributed Detection and Estimation in WSNs June 6, 2017 42 / 42