Deep Reinforcement Learning Control Approach to Mitigating Attacks

In this chapter, the deep reinforcement learning based secure control problem of CPSs under actuator attacks is first investigated. A reinforcement learning algorithm is proposed to learn the secure policy for CPSs, based on which deep neural networks are constructed and offline trained. In the inference, trained deep neural networks are deployed to output the secure control signal. The main contributions of this chapter can be summarized as follows: It is the first time to develop a deep reinforcement learning secure control algorithm for CPSs under actuator attacks. CPSs under attacks is converted into an MDP. In this way, the physical model can be nonlinear, and uncertainties, disturbance can be included in the model. Compared with existing results, a more general system model is used.