New BASIS defense reduces LLM over-refusal from prompt injection attacks

Researchers have developed BASIS, a novel defense system designed to mitigate prompt injection attacks in large language models. Unlike previous methods that broadly reject potentially compromised inputs, BASIS employs a two-stage approach using attention competition ratios to predict if an injection will actually compromise the model. This allows BASIS to selectively refuse harmful injections while permitting safe ones, thereby reducing unnecessary over-refusal and improving the usability of LLM applications. AI

IMPACT This defense mechanism could improve the reliability and usability of LLM applications by preventing unnecessary rejections of user inputs.

RANK_REASON The cluster contains a research paper detailing a new method for LLM security. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• Attention Competition Ratio
• BASIS
• CatalyzeX
• DagsHub
• Gotit.pub
• Hugging Face
• large language model
• prompt injection
• ScienceCast

• paper
• safety

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