Data Transmission Resilience to Cyber-attacks on Heterogeneous Multi-agent Deep Reinforcement Learning Systems

This paper investigates the data transmission resilience between agents of a cluster-based, heterogeneous, multi-agent deep reinforcement learning (MADRL) system under gradient-based adversarial attacks. We propose an algorithm using a deep Q-network (DQN) approach and a proportional feedback controller to defend against the fast gradient sign method (FGSM) attack and improve the DQN agent performance. The feedback control system is an auxiliary tool that helps the DQN algorithm reduce system deficiencies. In accordance with the achieved results and under FGSM adversarial attack, the resilience of the developed system is evaluated in three different ways termed robust, semi-robust, and non-robust based on average reward and DQN loss. The data transfer is carried out between agents of a MADRL system in timely and time-delayed manners, for both leaderless and leader-follower scenarios. Simulation results are included to verify the presented results.