MetAdv is a novel platform for assessing the adversarial robustness of autonomous driving systems. It integrates virtual simulations with real-world vehicle feedback to enable realistic and dynamic interactive evaluations. Through a three-tiered, closed-loop test environment, it performs end-to-end adversarial evaluations, from high-level integrated adversarial attack generation to mid-level simulation-based interactions and low-level real-world vehicle execution. It supports a variety of autonomous driving tasks and algorithmic paradigms (e.g., modular deep learning pipelines, end-to-end learning, and vision-language models), and is compatible with commercial platforms such as Apollo and Tesla. Its human-involvement capabilities provide flexibility in configurations and real-time collection of driver physiological signals and behavioral feedback, providing new insights into human-machine trust in adversarial environments.
