SEAG: Privacy-Preserving RAG Framework Conceals Sensitive Data from External LLMs

A new academic paper introduces the Sensitive Entity Alias Generator (SEAG), a privacy-preserving framework designed to prevent external large language models (LLMs) from accessing sensitive information during retrieval-augmented generation (RAG). The paper, available on arXiv (ID 2608.12675), highlights a gap in existing RAG privacy research: while prior work focused on preventing unauthorized access to sensitive data, it overlooked the risk that external generators can see the query and retrieved documents, which may contain confidential information. SEAG addresses this by using a lightweight model to identify sensitive entities, generate aliases for them, and construct an entity replacement table. This table replaces sensitive words in the user's query and retrieved documents before they are sent to the external LLM, allowing users to leverage powerful third-party generators without disclosing sensitive information. The framework empowers users to maintain privacy while benefiting from advanced AI capabilities. The paper's announcement type is 'new', indicating a recent addition to the arXiv repository. The research contributes to the field of AI privacy, specifically within the context of RAG systems, and offers a practical solution for organizations and individuals concerned about data confidentiality when using external AI services.