Reproduce “Exploiting Vision Encoder Vulnerabilities for Universal Adversarial Perturbations on Large Vision-Language Models” 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
Hee-Seon Kim, Minbeom Kim, Seokil Ham, Changick Kim
VEV-UAP targets vulnerable middle-layer value vectors to learn a bounded universal image perturbation, selected using hidden-state tests. It optimizes across images without prompts or language-model parameters. On LLaVA-1.5 with CLIP-336, it reaches 96.1% attack-success rate on ImageNet classification at a response-similarity threshold of 0.9, versus 29.9% for the strongest baseline; transfer is demonstrated only between models sharing an encoder.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.