Skip to content

RECLAIM: robot vision-guided adversarial patch defense via embodied attention feedback

RECLAIM: Robot-Vision-Informed Adversarial Patch Defense via Embodied Attention Feedback

arXiv: 其他 AI 安全论文· arXiv:2611.24494· Jiayou Zhu, Peng Guo, Litian Zhang, Xinze Shi, Yuting Zhang, Zelong Yu, Yize Tian, Jirong Liu, Zifeng Kang·Indexed by this site · 13 hr ago· Originally publishedDefenses· Its place in the scan record →
AI summary

Researchers propose RECLAIM, a training-free, inference-time defense that uses a frozen vision-language-action (VLA) policy's own attention responses to local visual interventions to guide purification in closed-loop embodied manipulation, defending against adversarial patches that corrupt action prediction. Such patches alter subsequent observations; unlike existing defenses that rely on policy adaptation or auxiliary models and require additional training or inference, RECLAIM uses attention changes induced by local edits as feedback, combines robot geometry to refine patch localization and select a local processing method, and re-verifies as the agent's actions change the observed scene. Across multiple manipulation task suites under frozen VLA control, the method defends against patches generated by four attack methods, achieving at least one successful rollout on 95% of evaluated tasks, raising LIBERO-Object rollout success from 0% without defense to 70.0%, with an average decision time of only 548.9 milliseconds per control step.

Read originalarxiv.org