WIP: Learning Adversarial Attacks on Adaptive Traffic Signal Control Systems Under Cooperative Perception

Significant advancements in traffic control systems, such as integration with sensing and communication technologies, have led to increased system complexity. While these developments offer substantial benefits, they also introduce heightened vulnerabilities in cyberspace. This paper presents a security analysis of adaptive traffic control systems operating under cooperative perception environments with connected and automated vehicles (CAVs). To explore system vulnerabilities, we propose a novel reinforcement learning-based black-box adversarial attack framework, which demonstrates effectiveness against state-of-the-art adaptive traffic control systems. Specifically, the multi-action proximal policy optimization (multi-PPO) algorithm is employed to train the attacker agent capable of generating a fake CAV along with its "detected" vehicles. Experimental results indicate that the fake CAV can fool a learning-based traffic control system by injecting falsified detection data, leading to a 62.5% increase in average vehicle delay.