Publication Date

5-2017

Date of Final Oral Examination (Defense)

4-7-2017

Type of Culminating Activity

Dissertation

Degree Title

Doctor of Philosophy in Electrical and Computer Engineering

Department

Electrical and Computer Engineering

Supervisory Committee Chair

Hao Chen, Ph.D.

Supervisory Committee Member

John N. Chiasson, Ph.D.

Supervisory Committee Member

Leming Qu, Ph.D.

Supervisory Committee Member

Qi Cheng, Ph.D., External Examiner

Abstract

Comprised of a large number of low-cost, low-power, mobile and miniature sensors, wireless sensor networks are widely employed in many applications, such as environmental monitoring, health-care, and diagnostics of complex systems. In wireless sensor networks, the sensor outputs are transmitted across a wireless communication network to legitimate users such as fusion centers for final decision-making.

Because of the wireless links across the network, the data are vulnerable to security breaches. For many applications, the data collected by local sensors are extremely sensitive, and care must be taken to prevent that information from being leaked to any malicious third parties, e.g., eavesdroppers. Eavesdropping is one of the most significant threats to wireless sensor networks, where local sensors are tapped by an eavesdropper in order to intercept information.

I considered distributed inference in the presence of a global, greedy and informed eavesdropper who has access to all local node outputs rather than access. My goal is to develop secured distributed systems against eavesdropping attacks using a physical-layer security approach instead of cryptography techniques because of the stringent constraints on sensor networks energy and computational capability. The physical-layer security approach utilizes the characteristics of the physical layer, including transmission channels noises, and the information of the source. Additionally, physical-layer security for distributed inference is scalable due to the low computational complexity.

I first investigate secrecy constrained distributed detection under both Neyman-Pearson and Bayesian frameworks. I analyze the asymptotic detection performance and proposed a novel way of analyzing the maximum performance trade-off using Kullback-Leibler divergence ratio between the fusion center and eavesdropper. Under the Neyman-Pearson framework, I show that the eavesdropper's detection performance can be limited such that her decision-making is no better than random guessing; meanwhile, the detection performance at the fusion center is guaranteed at the prespecified level. Similar analyses and proofs are provided under the Bayesian framework, where it was shown that an eavesdropper can be constrained to an error probability level equal to her prior information. Additionally, I derive the asymptotic error exponent and show that asymptotic perfect secrecy and asymptotic perfect detection are possible by increasing the number of sensors under both frameworks if the fusion center has noiseless channels to the sensors.

For secrecy constrained distributed estimation, I conducted similar analysis under both a classical setting and Bayesian setting. I derived the maximum achievable secrecy performance and show that under the condition that the eavesdropper has noisy channels and the fusion center has noiseless channels, both asymptotic perfect secrecy and asymptotic perfect estimation can be achieved under a classical setting. Similarly, under a Bayesian setting, I derived the performance trade-off using Fisher information ratio and show that the fusion center outperforms the eavesdropper significantly in the simulation section.

Secrecy constrained in distributed inference with Rayleigh fading binary symmetric channel is considered as well. Similarly, I derive the maximum achievable secrecy performance ratio for both detection and estimation.

The maximum achievable trade-off turns out to be almost the same in distributed estimation as in distributed detection. This suggests that a universal framework for generally structured inference problems are feasible. Further investigations are needed to justify this conjecture for more general applications.

DOI

https://doi.org/10.18122/B2WM5S

Download

Share

COinS