Reproduce “SilentDrift: Exploiting Action Chunking for Stealthy Backdoor Attacks on Vision-Language-Action Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Bingxin Xu, Yuzhang Shang, Binghui Wang, Emilio Ferrara
SILENTDRIFT uses a triggered gradual offset during robot action chunks, when visual correction is delayed. Its Smootherstep ramp adds drift smoothly during a precision-critical approach, while selected near-target frames are poisoned. On the four-suite LIBERO benchmark, VLA-Adapter reports 93.2% average attack success and 95.3% clean-task success at 2% poisoning, versus 96.6% baseline success; these are benchmark results, not real-robot evidence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.