Multi-Agent and Group Aware Reasoning for Robust Traffic Signal Control Through Jacobian Norm Based on Adversarial Defense

Abstract Traffic Signal Control (TSC) is a critical component of intelligent transportation systems, directly influencing urban mobility, congestion levels, and environmental sustainability. While Multi-Agent Reinforcement Learning (MARL) has shown significant promise in TSC through decentralized decision-making, it remains vulnerable to adversarial attacks that manipulate learned policies, leading to inefficient traffic management. To overcome this issue, this study proposes a novel Multi-Agent and Group Aware Reasoning Framework with Jacobian Norm-Based Adversarial Defense for robust traffic signal control. Our approach leverages graph-based coordination to enhance multi-agent cooperation while incorporating Jacobian norm regularization to mitigate adversarial perturbations, ensuring both efficiency and resilience in traffic flow optimization. To evaluate the effectiveness of our method, we conducted extensive simulations using real-world traffic datasets from New York City, Jinan, and Hangzhou under varying congestion levels. The results demonstrate that our approach consistently outperforms state-of-the-art traffic control models in both standard and adversarial conditions. Compared to the strongest baseline state-of-the-art models, our method achieves a 9.2% reduction in Average Travel Time (ATT) in Jinan (from 199.24s to 190.04s), a 9.3% reduction in New York City (from 748.32s to 704.30s), and a 4.5% reduction in New York City (from 1039.15s to 992.26s). Under adversarial perturbations, our Jacobian norm-based model maintains robust performance, achieving an ATT of 711.61s, and 997.57s in complex scenarios, significantly outperforming Deep Perturbation and Adversarial Bayesian Neural Networks. This study sets a new benchmark for intelligent transportation systems with the lowest ATT, shortest queue lengths, and highest resilience to adversarial conditions.