Adversarial Flow Matching for Imperceptible Attacks on End-to-End Autonomous Driving

Researchers introduce Adversarial Flow Matching (AFM), a gray-box attack framework targeting Transformer backbones shared across monolithic VLA and modular E2E autonomous driving architectures. AFM uses flow matching for one-step adversarial perturbation generation, bypassing white-box transparency requirements and black-box query latency bottlenecks while improving cross-model transferability. AD safety engineers and red-teamers validating E2E driving stacks against imperceptible physical-world attacks should prioritize this work immediately.