Secure Consensus for Multiagent Systems Under False Data Injection Attacks: A Hybrid Reinforcement Learning Scheme

Abstract

This work addresses the problem of secure consensus in heterogeneous multiagent systems (MASs) under false data injection attacks (FDIAs). To balance the impact of malicious attacks against system performance, an $H_{\infty }$ consensus control scheme is developed, which treats attacks as worst case disturbances, attenuates their effects, and guarantees consensus of system trajectories. Moreover, a novel hybrid iteration (HI) algorithm based on reinforcement learning (RL) is developed to address the graphical game algebraic Riccati equations (GAREs) without relying on prior information of complete system dynamics. By combining the merits of policy iteration (PI) and value iteration (VI), the proposed HI algorithm can be applied to cases without an initial stable control policy (ISCP) while ensuring a fast iteration speed. Finally, the effectiveness and superiority of the proposed control method are validated through numerical examples and comparative cases.