Reproduce “Dummy Backdoor as a Defense: Removing Unknown Backdoors via Shared Internal Mechanisms for Generative LLMs” 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

Kazuki Iwahana, Masaru Matsubayashi, Takuma Koyama, Toshiki Shibahara, Kenichiro Ominato, Akira Ito

The authors propose using a known dummy trigger to reduce an unknown backdoor targeting the same task, without identifying its trigger. They train the model to respond cleanly to the dummy trigger, then test transfer to the unknown trigger. On Llama-3.1-8B-Instruct’s Style jailbreak, ASR falls from 92.93% to 7.07%; task knowledge remains necessary.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.