A Lightweight and Communication-Efficient Protection Framework for RAG Retrieval via Heterogeneous Local Differential Privacy

Abstract

In order to solve the risk of privacy leakage in the Large Language Model (LLM) Search Augmented Generation (RAG), and the problems of “dimension disaster” and high communication cost faced by traditional local differential privacy (LDP) in high-dimensional space, this paper proposes a lightweight protection framework: reduced-dimension heterogeneous local differential privacy (DRH-LDP). In this framework, principal component analysis (PCA) is used to extract the core manifold of query vectors, and an innovative heterogeneous privacy budget allocation mechanism based on eigenvalue ratio is introduced to maximize the signal-to-noise ratio of core semantics. In addition, the fixed-scale 8-bit truncated quantization technique is used to compress the ciphertext and filter out the heavy-tailed noise. Experiments on real data sets show that under strict (, δ)-LDP constraints, the proposed method reduces the network traffic to 1/48 of the original size, and improves the retrieval recall rate by more than 5 times under privacy constraints. This research provides an efficient engineering paradigm for the safe deployment of RAG system in data-sensitive industries.