SilentDrift: Stealthy Backdoor Attacks on Vision-Language-Action Models via Action Chunking

The researchers propose SilentDrift, a black-box backdoor attack against vision-language-action (VLA) models that exploits the intra-chunk visual open loop formed by action chunking and delta pose representations, allowing step-by-step perturbations to accumulate. The method uses the Smootherstep function to construct perturbations satisfying C2 continuity, making velocity and acceleration zero at trajectory boundaries, and poisons only the approach phase of keyframes. On LIBERO, it achieves a 93.2% attack success rate with a poisoning rate below 2%, while maintaining a 95.3% clean task success rate.