Robust Multi‐Agent Reinforcement Learning Against Adversarial Attacks for Cooperative Self‐Driving Vehicles

ABSTRACT Multi‐agent deep reinforcement learning (MARL) for self‐driving vehicles aims to address the complex challenge of coordinating multiple autonomous agents in shared road environments. MARL creates a more stable system and improves vehicle performance in typical traffic scenarios compared to single‐agent DRL systems. However, despite its sophisticated cooperative training, MARL remains vulnerable to unforeseen adversarial attacks. Perturbed observation states can lead one or more vehicles to make critical errors in decision‐making, triggering chain reactions that often result in severe collisions and accidents. To ensure the safety and reliability of multi‐agent autonomous driving systems, this paper proposes a robust constrained cooperative multi‐agent reinforcement learning (R‐CCMARL) algorithm for self‐driving vehicles, enabling robust driving policy to handle strong and unpredictable adversarial attacks. Unlike most existing works, our R‐CCMARL framework employs a universal policy for each agent, achieving a more practical, nontask‐oriented driving agent for real‐world applications. In this way, it enables us to integrate shared observations with Mean‐Field theory to model interactions within the MARL system. A risk formulation and a risk estimation network are developed to minimise the defined long‐term risks. To further enhance robustness, this risk estimator is then used to construct a constrained optimisation objective function with a regulariser to maximise long‐term rewards in worst‐case scenarios. Experiments conducted in the CARLA simulator in intersection scenarios demonstrate that our method remains robust against adversarial state perturbations while maintaining high performance, both with and without attacks.