Authors: Zhenxing Niu, Haodong Ren, Xinbo Gao, Gang Hua, Rong Jin

Published on: February 04, 2024

Impact Score: 8.22

Arxiv code: Arxiv:2402.02309

Summary

• What is new: A new maximum likelihood-based algorithm to create image Jailbreaking Prompts (imgJP) that works across different models and prompts.
• Why this is important: The issue of multi-modal large language models (MLLMs) producing objectionable responses to specific user queries.
• What the research proposes: Developing an algorithm that finds image jailbreaking prompts to prevent MLLMs from generating harmful responses.
• Results: The imgJP exhibits strong model-transferability and efficiency in preventing unwanted content generation compared to existing methods.

Technical Details

Technological frameworks used: nan

Models used: MiniGPT-v2, LLaVA, InstructBLIP, mPLUG-Owl2

Data used: nan

Potential Impact

Companies leveraging MLLMs for content generation, moderation, and interaction with users could benefit from these insights to enhance model safety and reliability.

Want to implement this idea in a business?

We have generated a startup concept here: ShieldAI.