Worst-case analysis of joint attack detection and resilient state estimation

This work investigates the effects of signal attacks possibly combined with network deception attacks injecting fake measurements on stochastic cyber-physical systems. The goal of the attacker is to maximize the estimation error based on the information available about the system and the measurement models, preferably without being detected. This problem is formulated following a worst-case approach characterizing the maximum degradation the attacker can induce at each time instant when a Bayesian filter developed within the random finite set (RFS) framework is employed for simultaneous attack detection and resilient state estimation. A novel concept of error which captures the switching (Bernoulli) nature of the signal attack is proposed as an appropriate distance measure for joint detection-estimation. Furthermore, the notion of stealthiness is introduced in order to derive attack policies useful to synthesize undetectable perturbations that can deceive a Maximum Aposteriori Probability (MAP) detector implemented for security.