SilentDrift: Exploiting Action Chunking for Stealthy Backdoor Attacks on Vision-Language-Action Models
Researchers have developed a novel backdoor attack called SILENTDRIFT that targets Vision-Language-Action (VLA) models used in robotics. The attack exploits a vulnerability in how these models process action sequences, allowing subtle perturbations to accumulate and lead to incorrect execution. SILENTDRIFT achieves a high success rate with minimal poisoning and maintains high performance on clean tasks, making the poisoned trajectories visually indistinguishable from normal operations. AI
IMPACT Highlights critical security vulnerabilities in deployed AI systems, necessitating robust defenses for safety-critical applications.