Thesis Examination Committee
Prof Wan Keung WONG, LIFS/HKUST (Chairperson)
Prof Ling SHI, ECE/HKUST (Thesis Supervisor)
Prof Yang SHI, Department of Mechanical Engineering, University of Victoria (External Examiner)
Prof Wai Ho MOW, ECE/HKUST
Prof Ming LIU, ECE/HKUST
Prof Shuhuai YAO, MAE/HKUST
This thesis investigates security issues for remote state estimation in the context of cyber-physical systems. Cyber threats arising from the interconnection of different technologies often impose challenges on the state estimation performance. To achieve a desirable tradeoff between the estimation quality and the system security, different problems are studied from the perspective of either a malicious attacker or a system designer.
First, from the attacker's perspective, we consider an innovation-based integrity attack which aims to degrade system performance while remaining stealthy to a false-data detector by intercepting and modifying the transmitted data. The recursion of the remote estimation error covariance in the presence of attack is derived and the closed-form expression of the optimal attack is obtained. This problem is generalized from the following two aspects. When Kullback-Leibler divergence is adopted as a stealthiness metric, a two-stage optimization problem is formulated to investigate the optimal attack strategy. The optimal attack is first shown to be Gaussian distributed and then solved using semi-definite programming. On the other hand, innovation-based integrity attacks under different information sets are considered. The worst-case attack consequences are analyzed based on the intercepted data, the sensing data or alternatively the combined information.
Second, from the system's perspective, we consider a secure state estimation problem in a multi-sensor scenario where a subset of the sensors can potentially be compromised by an attacker. To locate the compromised sensors, we propose a Gaussian-mixture-model-based detection algorithm. It is able to cluster the local state estimate autonomously and provide a belief on each sensor, based on which different sensor measurements can be fused accordingly. The performance of the proposed detection algorithm is evaluated by the remote estimation error covariance and the average belief. Furthermore, we evaluate the effectiveness of the proposed detection algorithm under different attack scenarios.