Stop-and-Go: Exploring Backdoor Attacks on Deep Reinforcement Learning-Based Traffic Congestion Control Systems

Recent work has shown that the introduction of autonomous vehicles (AVs) in traffic could help reduce jams. Deep reinforcement learning methods demonstrate good performance complex control problems, including vehicle control, and have been used state-of-the-art AV controllers. However, deep neural networks (DNNs) render automated driving vulnerable to machine learning-based attacks. In this work, we explore backdooring/trojanning DRL-based We develop a trigger design methodology is based on well-established principles physics. The malicious actions include deceleration acceleration cause stop-and-go waves emerge (congestion attacks) or resulting crashing into front (insurance attack). test our attack single-lane two-lane circuits. Our experimental results show backdoored model does not compromise normal operation performance, with maximum decrease cumulative rewards being 1%. Still, it can be maliciously activated crash congestion when corresponding triggers appear.