Stealthy Black-Box Attack With Dynamic Threshold Against MARL-Based Traffic Signal Control System

IEEE Transactions on Industrial Informatics (TII), 2024

Yan Ren1 Heng Zhang1 Linkang Du2 Zhikun Zhang3 Jian Zhang4 Hongran Li4

1. College of Electronic Engineering, Jiangsu Ocean University, Lianyungang, China 2. College of Control Science and Engineering, Zhejiang University, Hangzhou, China 3. College of Computer Science and Technology, Zhejiang University, Hangzhou, China 4. College of Computer Engineering, Jiangsu Ocean University, Lianyungang, China

Abstract


Deep reinforcement learning has shown promise for adaptive traffic signal control, but existing attacks on DRL-based systems are mostly static and easier to detect. We propose a stealthy black-box attack with a dynamic threshold against multi-agent reinforcement learning (MARL)-based traffic signal control. The attacker adaptively triggers perturbations according to traffic-state dynamics to degrade control performance while reducing detectability. Experiments on benchmark traffic simulators show significant congestion amplification and strong stealth characteristics compared with prior attacks.

Resources


Citation 

 @inproceedings{RZDZZL24,
    author = {Yan Ren and Heng Zhang and Linkang Du and Zhikun Zhang and Jian Zhang and Hongran Li},
    title = {{Stealthy Black-Box Attack With Dynamic Threshold Against MARL-Based Traffic Signal Control System}},
    booktitle = {{Transactions on Industrial Informatics}},
    publisher = {IEEE},
    year = {2024},
}