Summary of Jailbreaking Attack Against Multimodal Large Language Model, by Zhenxing Niu and Haodong Ren and Xinbo Gao and Gang Hua and Rong Jin
Jailbreaking Attack against Multimodal Large Language Model
by Zhenxing Niu, Haodong Ren, Xinbo Gao, Gang Hua, Rong Jin
First submitted to arxiv on: 4 Feb 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Computation and Language (cs.CL); Cryptography and Security (cs.CR); Computer Vision and Pattern Recognition (cs.CV)
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 The proposed algorithm targets jailbreaking attacks on multi-modal large language models (MLLMs), aiming to elicit objectionable responses from MLLMs when faced with harmful user queries. The method, based on maximum likelihood, generates an image Jailbreaking Prompt (imgJP) that can be applied across various unseen prompts and images. This approach demonstrates strong model-transferability, allowing the generated imgJP to jailbreak different models, including MiniGPT-v2, LLaVA, InstructBLIP, and mPLUG-Owl2, in a black-box manner. Furthermore, the connection between MLLM-jailbreaks and LLM-jailbreaks is revealed, leading to an efficient construction-based method for LLM-jailbreaks that surpasses current state-of-the-art methods. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper explores ways to “trick” big language models into saying things we don’t want them to. The goal is to make these models behave badly when given bad inputs. Researchers came up with a way to create prompts that can make different models say something undesirable. This approach works well across many models and even helps with another type of problem where the models are made to act poorly. The code for this project is available online. |
Keywords
* Artificial intelligence * Likelihood * Multi modal * Prompt * Transferability
