From One Query to Many: A Critical Survey of Privacy Accounting in RAG
This survey examines how privacy accounting changes when moving from generating a single response to running many queries in Retrieval-Augmented Generation (RAG) systems. It categorizes existing methods into three types: query-time privatization, multi-query accounting, and one-time corpus privatization. The authors propose a new framework for evaluating DP guarantees based on privacy granularity, accounting scope, and release boundaries.