Title: SilentDrift: Leveraging Action Chunking for Undetectable Backdoor Attacks on Vision-Language-Action Models

Abstract:

As Vision-Language-Action (VLA) models become more prevalent in robotic systems where safety is paramount, their security weaknesses have not been sufficiently investigated. This study reveals a critical security defect in contemporary VLA architectures: the interplay between action chunking and delta pose representations generates an intra-chunk visual open-loop. This structural characteristic compels the robot to carry out sequences of K-step actions, thereby permitting perturbations at each step to accumulate via integration.

To capitalize on this vulnerability, we introduce SILENTDRIFT, a stealthy black-box backdoor attack. The proposed technique utilizes the Smootherstep function to generate perturbations that ensure C2 continuity. This guarantees that both velocity and acceleration are zero at the boundaries of the trajectory, a requirement for strict kinematic consistency. Additionally, our approach employs a keyframe-based strategy that targets only the essential approach phase for poisoning. This selective method amplifies the attack’s effectiveness while reducing the visibility of the trigger. Consequently, the corrupted trajectories remain visually indistinguishable from legitimate successful demonstrations. In evaluations conducted on the LIBERO benchmark, SILENTDRIFT attained an Attack Success Rate of 93.2% with a poisoning rate below 2%, while preserving a Clean Task Success Rate of 95.3%.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC