WiP: An On-device LLM-based Approach to Query Privacy Protection

Privacy leakage from user queries is a widely-concerned issue in search engines and chatbot services. Existing solutions based on privacy information removal, obfuscation, and encryption may inevitably hurt service quality or require full trust of the service provider. Inspired by the remarking language understanding and generation abilities of large language models (LLMs), we introduce LLM-QueryGuard, an LLM-based tool designed to mitigate privacy leakage in continuous user queries. The core of LLM-QueryGuard is an on-device LLM that automatically understands the private properties contained in the user queries and generates false queries to obfuscate the private properties. By making the generated queries indistinguishable and mixing them into the real queries, our approach can be seamlessly integrated into any query-driven applications (e.g. search engine and ChatGPT), with some cost of additional queries.