New Framework Protects Sensitive Data in RAG Systems.

Retrieval-Augmented Generation (RAG) systems are widely used to enhance Large Language Model (LLM) performance in answering user queries. While existing privacy research for RAG focuses on access control, a significant overlooked issue is that external LLM generators often receive queries and retrieved documents containing confidential information, posing risks of misuse or unintended access. To address this, the Sensitive Entity Alias Generator (SEAG) framework has been developed. SEAG is a lightweight model designed to identify sensitive entities, generate corresponding aliases, and create an entity replacement table. This table is then used to substitute sensitive words in both the user's query and the retrieved documents before they are sent to an external generator. Experimental results, using specially constructed datasets, demonstrate SEAG's success: it achieved over 80% accuracy in providing correct user responses while effectively concealing sensitive information from the external LLM. Further analysis showed models like Qwen-3, LLaMA-3.2, and Phi-4 achieved total accuracies of 77.83%, 76.73%, and 74.91% respectively in hiding all sensitive entities within documents.