Reproduce “Understanding and Defending VLM Jailbreaks via Jailbreak-Related Representation Shift” 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
Zhihua Wei, Qiang Li, Jian Ruan, Zhenxin Qin, Leilei Wen, Ruiyang Qin, Qingzhuo Wang, Dongrui Liu and 1 more
The authors introduce JRS-Rem to reduce harmful answers to image-based jailbreaks in vision-language models. It estimates the image-induced representation shift along a direction from refusal to harmful-answer states, then removes that component while retaining the remainder. On LLaVA-1.5-7B, HADES attack success falls from 84.9% to 12.4% for explicit harmful prompts with SD+TYPO images; the method requires hidden-state access.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.