Reproduce “Defending LVLMs Against Vision Attacks through Partial-Perception Supervision” 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
Qi Zhou, Tianlin Li, Qing Guo, Dongxia Wang, Yun Lin, Yang Liu, Jin Song Dong
DPS defends large vision-language models from misleading image attacks by using partial-image observations to guide a new answer with the full image visible. Partial-view agents describe crops, whose text guides a full-view model answering the original question. On conditioned samples, Qwen-VL-Plus averaged 0.31 ASR with DPS versus 0.48 with random-crop Standard across three datasets; this does not isolate causation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.