A Model-Free Stealthy Attack for Cyber-Physical Systems Based on Deep Reinforcement Learning

This article, from the attacker's standpoint, develops a model-free stealthy attack that can steer the system state to the predefined target value and evade detection, without prior knowledge of the system dynamics. A constrained Markov decision process (CMDP) is first modeled to characterize the objective of the stealthy attack. On the basis of the established CMDP, an actor-critic reinforcement learning algorithm is proposed to train the attacker's policy. Furthermore, by introducing a Lyapunov function constructed from the action value function to the algorithm, convergence of the attacked system's state to the target is theoretically guaranteed. Differing from existing model-free stealthy attacks which are only suitable for linear systems, the proposed approach guarantees the applicability to nonlinear systems. A linear numerical example and a nonlinear example of flotation industrial system are provided to validate the effectiveness of our proposed stealthy attack.