Learning-based Attack Schedule against Remote State Estimation in Cyber-Physical Systems

Malicious attack against remote state estimation in cyber-physical systems has attracted considerable attention. Nevertheless, most existing works assume that the attacker is powerful and has mastered the communication information when designing the attack strategy. In this paper, we consider that the attacker is energy-limited and has no prior knowledge of transmission pattern. To solve this problem, we introduce a learning-based method for the attacker, which consists of a learning phase and an attack phase, to achieve a smart attack. We first formulate an optimal attack schedule problem, aiming to maximize the estimation error while considering the tradeoff between the learning accuracy and attack efficiency. Since it is hard to solve this problem directly, we split it into two subproblems: i) optimizing the attack pattern; ii) optimizing the eavesdropping times and attack times. Theoretically, we prove that the optimal attack pattern is that the learning phase precedes the grouped attack from the viewpoint of possibility. Furthermore, we propose an algorithm to design the rational learning times and attack times for the attacker. Numerical examples are used to demonstrate the effectiveness of the proposed method.