Security Control Against FDI Attacks via Adaptive Off-Policy Value Iteration Q-Learning Approach

This article addresses the security control problem of cyber-physical systems with partially unknown system dynamics under false data injection (FDI) attacks. An adaptive off-policy value iteration Q-learning (VIQL)-based security control scheme is proposed, where the control input and attack signal are modeled as two players in a zero-sum game. By computing the optimal control input, this approach effectively mitigates the impact of FDI attacks. To accelerate convergence, a relaxation factor is introduced, and the convergence of the scheme is analyzed under different relaxation factors. Furthermore, an unknown security control framework is developed using offline data and a critic-only neural network to implement this scheme while ensuring interpretability and implementability. Simulation results further demonstrate the effectiveness of the proposed approach.