Reproduce “Toward Universal and Transferable Jailbreak Attacks on 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

Kaiyuan Cui, Yige Li, Yutao Wu, Xingjun Ma, Sarah Erfani, Christopher Leckie, Hanxun Huang

The UltraBreak framework develops a novel visual jailbreak attack for Vision-Language Models (VLMs) that achieves both universality across malicious targets and transferability to unseen black-box models. It demonstrates superior attack success rates on open-source and proprietary VLMs compared to previous methods by combining semantic textual objectives with robust visual input constraints.

Key insights

  • Previous gradient-based VLM jailbreak attacks fail to transfer due to the combination of cross-entropy loss and image-space constraints creating spiky loss landscapes that lead to surrogate overfitting.
  • Achieving strong generalization and transferability in VLM jailbreaks requires smoothing the loss landscape through semantically guided objectives and promoting robust, model-invariant visual features.
  • Effective transferability to unseen black-box VLMs can be achieved using a single surrogate model by focusing on generating robust, text-like visual patterns and optimizing with a stable, semantically-aware loss function.

Summary by alphaXiv.

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