InfoResearchPreprint
RECLAIM: Robot-Vision-Informed Adversarial Patch Defense via Embodied Attention Feedback
- Published
- Record updated
Summary
RECLAIM is a training-free, inference-time defense for vision-language-action robot policies against adversarial patches. It uses a frozen VLA's own attention response to local visual edits to guide where and how to purify the scene during closed-loop manipulation. Under patch attacks from four methods, it succeeds in at least one rollout on 95% of evaluated tasks, raising LIBERO-Object success from 0% to 70.0% at 548.9 ms per control step.
Mitigation
RECLAIM, a training-free, inference-time defense that uses the frozen VLA's attention response to local visual interventions to guide purification, combined with robot geometry to refine patch localization and revalidate treatments as the scene changes.
Topics
Related items
- MediumGHSA-25rm-9wvm-m38v: falcon-rust: Discrete-Gaussian sampler runs at 2^-33 precision, far below Falcon's requirement (reference KAT vectors no longer pass)Similar attack · GitHub Advisory Database
- LowExecGap: Exploiting Command-Execution Gap for Patch Attacks on Vision-Language-Action ModelsSimilar attack · Arxiv (cs.RO + cs.CV security)
- InfoConfGuard: a physics-grounded reference monitor for predictive-maintenance large-language-model agentsSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)
- MediumGemini AI agents repeatedly bypassed safeguards to attack real websiteSimilar attack · BleepingComputer
- InfoWatermark Evidence under Open-Weight LLM Modification: Exact Bounds for Low-Rank Output WatermarksSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)