Who Can You Trust on the Road? Adversarial Intentions in Cooperative Multiagent Reinforcement Learning

Abstract

To enhance safety, comfort, and efficiency in intelligent transportation systems—especially in complex, mixed-autonomy traffic—connected and autonomous vehicles (CAVs) must interact seamlessly with both other CAVs and human-driven vehicles (HVs), despite the inherent unpredictability and potential deceptiveness of human behavior in high-stakes scenarios such as onramp merging. We propose Trust-MARL, a trust-based multiagent reinforcement learning framework that integrates dynamic trust modeling with game-theoretic decision-making to guide cooperative lane-changing behaviors. At the macro level, Trust-MARL fosters emergent group coordination to improve overall traffic efficiency, while at the micro level, it adapts cooperation strategies based on real-time and historical interactions with neighboring agents. To improve robustness under noncooperative or adversarial human behaviors, we augment the framework with an Adversarial Theory of Mind (Adversarial ToM) module that introduces belief-level perturbations during training, enabling CAVs to learn resilient policies against misleading or strategically deceptive intent signals. Extensive experiments and ablation studies across a range of traffic densities and CAV penetration rates show that our approach significantly improves safety, ride comfort, and throughput, outperforming conventional MARL and rule-based baselines even under adversarial perturbations.