Preserving User Privacy in Retrieval Augmented Generation: A Novel Approach Using Local Placeholder Tagging

Abstract

Retrieval Augmented Generation (RAG) is a popular approach that enhances the accuracy of Large Language Models (LLMs) by leveraging a knowledge base. It is rapidly becoming integral tools across various applications. However, as the use of RAG continues to expand, so do the challenges associated with their deployment, particularly in terms of data privacy. As a part of RAG pipeline, user query and all retrieved documents should be sent as a prompt to the LLM providers, leaving them open to privacy hazards such data leaks or illegal access. This study presents RLPT, a framework designed to enhance user privacy in RAG. It achieves this by identifying and eliminating sensitive information from user inputs before sending them to the LLM. The RLPT framework utilizes a local LLM to rapidly identify sensitive information in user input and subsequently replaces it with distinctive placeholders. These placeholders are used to indicate and hide the actual sensitive data, ensuring that the LLM does not capture the original sensitive information during prompt processing. The framework is evaluated using a dataset consisting of 4000 synthesized context documents. The results indicate that it is capable of accurately detecting and filtering privacy and sensitive information, achieving a high accuracy rate of 88,7%.