State Backdoor: Towards Stealthy Real-world Poisoning Attack on Vision-Language-Action Model in State Space
Researchers have developed a new type of backdoor attack targeting Vision-Language-Action (VLA) models, which are crucial for embodied AI applications like robotics. Unlike previous methods that rely on visible visual triggers, this novel "State Backdoor" utilizes the initial state of a robot arm as the trigger. A Preference-guided Genetic Algorithm was employed to find minimal yet effective state-based triggers, achieving over 90% attack success without degrading performance on normal tasks. AI
IMPACT Reveals a new vulnerability in embodied AI, potentially requiring new security measures for robotic systems.