Skip to content
InfoResearchPreprint

When Backdoors Can Be Switched Off: Suppression-Aware Backdoor Attacks on Vision-Language-Action Models

Published
Record updated
View JSON

Summary

Researchers study suppression-aware backdoors in Vision-Language-Action (VLA) models, which combine conjunctive trigger activation with instruction-conditioned suppression. Three variants are tested on SmolVLA using two SO-100 manipulation datasets against four baselines. Self-Suppressing and Multi-Trigger cut Isolation Forest detection from 96--98% to 1--18% and raise attack success to 63--71% with perfect clean-task performance, and experiments on OpenVLA with LIBERO show transfer to a 7B-parameter policy that evades STRIP and action-distribution-shift detection.