WIP: Learning Adversarial Attacks on Adaptive Traffic Signal Control Systems...
summary
Adaptive traffic controllers under cooperative perception are vulnerable to a reinforcement-learning black-box attack that injects a fake CAV with falsified detections into SDM streams; two implemented methods are a JSMA surrogate (uses policy distillation to craft which approach to add vehicles) and a multi-action PPO attacker trained to place tens of fake vehicles. The attacks reliably fool RL controllers, shifting main-street green times to ~10s or ~40s and producing roughly a 62.5% increase in average vehicle delay in experiments.
AI-written summary · about
chapters (7)
- 0:00Traffic signal hacking intro
- 1:17Adaptive traffic systems background
- 3:22Corporate scenario & RL controllers
- 5:50Threat model & attack assumptions
- 7:16Black-box attack approaches
- 10:03Experimental results & impact
- 12:31Future work & defenses
related talks
50mNavigating Chaos: The Consequences of Compromise in Mixed-Autonomy Freeways
06mWhen Collaboration Betrays: The Evolving Attack Surface of...
44mNeural Network Security: Live Demonstration of an Adversarial Attack
20mSLEIGHT: Real-time Closed-Loop Attack Analytics Framework for Holistic Intrusion...
48mStay Ahead of Adversarial AI in OT/ICS Environments – Mitigating CWE-1039
13mAUTO: Adaptive Congestion Control Based on Multi-Objective Reinforcement Learning...
42mHow to attack AI systems (and how to defend them) !!!!
46mA Short Introduction to Hacking Artificial Intelligence