Reproduce “When Background Matters: Breaking Medical Vision Language Models by Transferable Attack” 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
Akash Ghosh, Subhadip Baidya, Sriparna Saha, Xiuying Chen
MedFocusLeak tests whether a medical vision-language model can be misdirected without changing a scan’s central finding. It coordinates image and text perturbations in non-diagnostic background areas and uses an attention-shift objective. In a six-modality evaluation, it reports MAS 0.670 for InternVL-8B, versus 0.520 for AttackVLM. This is a benchmark score, not a measured clinical misdiagnosis rate.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.