Reproduce “Dataset Poisoning Attacks on Behavioral Cloning Policies” 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
Akansha Kalra, Soumil Datta, Ethan Gilmore, Duc La, Guanhong Tao, Daniel S. Brown
The authors test whether poisoning behavioral-cloning demonstrations creates a hidden visual trigger in CarRacing. They add a 3×3 red patch to gas-labeled observations without changing actions. Poisoning 5% of gas examples, or 2.31% of the dataset, makes the policy choose gas on nearly every triggered observation while ordinary reward stays broadly stable. This finding is limited to the tested simulation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.