WIP: A Benchmark for Security Evaluation of Adversarial Manoeuvre Attacks in Autonomous Driving

This paper presents a simulation-only benchmark for security evaluation of autonomous driving agents under adversarial manoeuvre attacks. The benchmark is implemented in MetaDrive and supports training attacker policies, evaluating victim policies, and comparing attack outcomes under controlled driving settings. The attacker is black-box with respect to the victim and operates only through physically valid driving actions; sensor spoofing, communication tampering, model-internal access, and code tampering are excluded.

The benchmark records training logs, per-episode outcomes, aggregate safety metrics, trained policies, and evaluation logs for reproducible analysis. Preliminary results show that the benign driving performance of victim policies varies substantially across policy types and driving settings. In addition, highway experiments show that a trained PPO attacker increases the victim mission-failure rate by 10 percentage points. The benchmark provides an open-source platform for training attackers, testing victim policies, and studying adversarial manoeuvre attacks under controlled simulation conditions.