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

Adversarial manoeuvre attacks examine how an Autonomous Driving policy responds when another road user performs physically valid but adversarial vehicle manoeuvres. Studying these attacks requires changing road geometry, traffic density, attacker behaviour, victim policy, and evaluation metrics under controlled conditions.

The demonstration presents an open-source MetaDrive-based benchmarking platform for configuring these conditions, training driving policies, training adversarial attackers, and evaluating attack outcomes through replays, logs, and metrics. The live demonstration uses a highway road geometry and reinforcement-learning policies trained with PPO, showing the complete workflow from victim training to attacker training and final evaluation.