New Audit Method Tracks Content Origin in RAG Systems

Researchers have developed a new method called Source-aware Membership Audit (SMA) to track the origin of content generated by retrieval-augmented generation (RAG) and multimodal RAG systems. Unlike previous methods that focused on whether data was memorized, SMA pinpoints whether generated content comes from pre-training data, external retrieval, or user input. This is achieved through an attribution estimation mechanism using zero-order optimization and a cross-modal technique that leverages multimodal large language models (MLLMs) to analyze image inputs. AI

IMPACT This new audit method could enhance privacy and accountability in generative AI systems by providing fine-grained source attribution for generated content.

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

Read on arXiv cs.AI →

• large-language models
• MLLMs
• Multimodal Retrieval-Augmented Generation
• retrieval-augmented generation
• Shixuan Sun
• Source-aware Membership Audit

• paper
• safety

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