Securing DRL-Based Traffic Signal Control Against Experience Replay Manipulation Attacks

The integration of Deep Reinforcement Learning (DRL) into Adaptive Traffic Control Systems (ATCS) has significantly improved traffic management in smart cities. However, DRL-based ATCS are particularly vulnerable to Experience Replay Manipulation Attacks (ERMA), where adversarial entities subtly manipulate stored experiences in the replay buffer to mislead the learning process and degrade system performance. This attack exploits weaknesses in Traffic Signal Controllers (TSCs) to induce artificial congestion or disrupt optimal traffic flow. This paper addresses the security risks posed by ERMA and proposes a defense framework to enhance the robustness of DRL-based TSCs. Our approach integrates anomaly detection, probabilistic validation, and adaptive input correction to safeguard the integrity of traffic signal decisions. Experimental validation in a simulated traffic environment demonstrates the effectiveness of our method in detecting and mitigating ERMA while maintaining optimal traffic flow.