FreezeVLA: Action Freezing Attacks against Vision-Language-Action Models
FreezeVLA: Action-Freezing Attacks against Vision-Language-Action Models
The researchers propose FreezeVLA, an attack framework targeting vision-language-action (VLA) models, in which adversarial images can cause VLA models to "freeze" and ignore subsequent instructions, achieving an average attack success rate of 76.2% across three state-of-the-art VLA models and four robotic benchmarks, significantly outperforming existing methods. The framework generates and evaluates action-freezing attacks through min-max bilevel optimization; the generated adversarial images exhibit strong transferability, and a single image can reliably paralyze the model under multiple language prompts. The authors note that such attacks disconnect the robot's "digital brain" from physical actions and may lead to inaction at critical moments of intervention, highlighting the urgent need for robust defense mechanisms.
Related events and news
1 events · 1 news mentions
News that mentions this paper
VicOne测试:被操纵输入可改变机器人行为
- IEEE Spectrum AIcites the paper's id or linkVicOne LAB R7:机器人安全功能有效,输入被操纵会怎样?
FreezeVLA found that an adversarial image could cause tested models to ignore later instructions.