Summary of White-box Multimodal Jailbreaks Against Large Vision-language Models, by Ruofan Wang et al.
White-box Multimodal Jailbreaks Against Large Vision-Language Models
by Ruofan Wang, Xingjun Ma, Hanxu Zhou, Chuanjun Ji, Guangnan Ye, Yu-Gang Jiang
First submitted to arxiv on: 28 May 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Artificial Intelligence (cs.AI)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary A recent study has explored the robustness of Large Vision-Language Models (VLMs) in multimodal tasks. The research focuses on the adversarial attacks that can compromise the integrity of these models, particularly their ability to generate affirmative responses with high toxicity. The proposed attack method involves a dual optimization objective that jointly targets both text and image modalities to exploit vulnerabilities within VLMs. This approach is more comprehensive than existing methods, which primarily assess robustness through unimodal attacks on images alone. The experiment demonstrates the effectiveness of this universal attack strategy in jailbreaking MiniGPT-4 with a 96% success rate. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Large Vision-Language Models are very powerful tools that can do many things, like understand and generate text and pictures. But they have a weakness – they can be tricked into saying or showing things that might not be nice. The researchers who wrote this paper found a way to make the models say yes to bad instructions most of the time. This is important because it means we need to find new ways to keep these models from doing harm. |
Keywords
* Artificial intelligence * Optimization
