AD² analyzes vulnerabilities in end-to-end driving agents like Transfuser to physics, EMI, and digital attacks in CARLA. Driving scores drop up to 99% under threats. Proposes lightweight attention-based detector for spatial-temporal consistency.
Key Points
- 1.Exposes black-box attack vectors on visual perception
- 2.Up to 99% score drop in agents
- 3.Superior detection efficiency across multi-cameras
Impact Analysis
Highlights safety gaps in autonomous driving, enabling robust mitigation strategies.
Technical Details
Closed-loop eval in CARLA sim. Attention mechanisms capture inconsistencies.