PDP-RAG: Privacy-Preserving Retrieval-Augmented Generation Based on Personalized Differential Privacy

Current large language models (LLMs) demonstrate remarkable capabilities by unifying various natural language processing tasks into a comprehensive text generation framework. Compared to fine-tuned models, retrieval augmented generation (RAG) enhances the quality of model output at a lower update cost by retrieving relevant information from external knowledge sources as context for the LLM. However, RAG introduces a significant risk of privacy leakage. Users can design prompts to extract these context information. If the context information contains sensitive data, it may result in privacy leakage. To address this issue, we propose a RAG data source privacy protection framework based on personalized differential privacy, named PDP-RAG. Experimental results show that our method achieves a good trade-off between data utility and privacy. We also incorporate personalized differential privacy to accommodate varying privacy requirements among data owners. In addition, the proposed method can generate a representative and synthetic dataset, which is smaller in scale than the original dataset and reduces the amount of external knowledge required for storage, without compromising the data utility. Compared to the existing synthetic text generation method SAGE, the dataset size can be reduced by 50%.