Private High-Dimensional Similarity Search for Cloud-Based Retrieval Augmented Generation

Retrieval augmented generation (RAG) has emerged as a robust technique that combines the strengths of information retrieval and generative models to produce accurate and context-rich responses for applications powered by large language models (LLMs). Due to the high cost and complexity of maintaining external data sources, there is a growing demand for cloud-based RAG solutions that provide external data retrieval
through similarity search as a service. However, this approach raises significant concerns regarding the potential exposure of sensitive query information, which jeopardizes user privacy.

Traditional methods for private information retrieval often rely on complex cryptographic techniques, which can be computationally demanding and negatively impact retrieval and generation performance. Moreover, current research on securityprovable private vector similarity search is primarily limited to low-dimensional data, resulting in poor performance when applied to high-dimensional embedding vectors associated with RAG. Conversely, existing studies on high-dimensional vector data lack efficient algorithms that ensure both query accuracy and provable security. To address these challenges, this research proposal seeks to develop novel privacy-preserving similarity search techniques based on secure hardware for cloud-based RAG, ensuring user query privacy while maintaining the efficacy and accuracy of RAG tasks. More specifically, we will undertake four key research tasks. First, we plan to design TEEbased fully oblivious high-dimensional similarity search methods that leverage stateof-the-art graph-based indexing for cloud-based RAG. Second, to further enhance query efficiency, we intend to devise differentially oblivious versions of secure highdimensional similarity search methods utilizing differential privacy. Third, we will extend the supported scenario to multiple RAG service providers, enabling efficient similarity search in a federated setting while preserving query privacy. Finally, thorough theoretical analysis and empirical studies will be conducted to evaluate the proposed methods and techniques.

The anticipated outcomes of this research include the development of a theoretical foundation for private high-dimensional similarity search, the design of practical algorithms that can be deployed in real-world applications, and comprehensive evaluations to demonstrate the effectiveness and security of the proposed solutions. By addressing the privacy concerns associated with cloud-based RAG, this research aims to enable broader adoption of RAG techniques in various industries where user privacy is of paramount importance.