VLMGuard framework uses unlabeled data to detect malicious VLM prompts

Researchers have developed VLMGuard, a new framework designed to detect malicious prompts targeting vision-language models (VLMs). This system addresses the challenge of limited labeled data by utilizing unlabeled user prompts collected in the wild. VLMGuard automatically estimates the maliciousness of prompts, enabling the training of a classifier without human annotation. Experiments indicate that VLMGuard outperforms existing methods, improving detection accuracy. AI

IMPACT Enhances the security and reliability of vision-language models by providing a method to detect harmful inputs without extensive manual labeling.

RANK_REASON The cluster contains a research paper detailing a new framework for detecting malicious prompts in vision-language models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• CatalyzeX
• DagsHub
• Gotit.pub
• Hugging Face
• IArxiv
• ScienceCast
• vision-language model
• VLMGuard
• Xuefeng Duan

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →