Privacy-Preserving Vector Search for Retrieval-Augmented Generation

Picture of mengxuan-zhang.md Mengxuan Zhang

15 Jun 2026

### About the project

Vector databases are now the backbone of large-scale similarity search and Retrieval-Augmented Generation (RAG). But the same indexes that make billion-scale search fast also leak information: recent work shows that an attacker can tell whether a specific document was in a retrieval corpus (*membership inference*), or even reconstruct the original text from stored embeddings (*embedding inversion*).

This project asks a concrete question: can we give a vector database a rigorous differential privacy (DP) guarantee without destroying query throughput and recall? Naively adding noise to high-dimensional embeddings collapses search quality. The interesting research lives in mechanisms that respect the geometry of the index.

The student will join an active line of work spanning high-throughput ANN search and the privacy/security of RAG, with a clear path from a well-defined problem to a publishable result.

### What you will work on

You will help design and evaluate a DP-aware vector index. A typical trajectory:

- Establish a threat model and attack baseline using membership inference against a RAG retrieval corpus.

- Build a differentially private quantized index that keeps the raw vectors private while preserving recall.

- Measure the three-way trade-off — recall@k × query throughput × attack success rate vs. privacy budget (ε) — on real RAG datasets.

- (Stretch) Quantify the “free” privacy contributed by random projection and tune it to a target guarantee.

### Who we’re looking for

We welcome applicants who have, or are close to completing, a strong undergraduate or coursework degree in Computer Science, Data Science, Mathematics, or a related field, and who have:

- solid programming skills (C++ a plus) and comfort with linear algebra and probability;

- interest in one or more of: approximate nearest neighbor search, vector databases, differential privacy, or the security of machine learning / LLM systems;

- the curiosity and persistence to read research papers and turn ideas into careful experiments.

Prior research experience or coursework in privacy, information retrieval, or high-dimensional data is a bonus but not required — we will help you get up to speed.

### Why join

You’ll be co-supervised across two complementary groups: one focused on high-throughput and high-recall vector search, and one focused on the privacy and security of RAG systems. The topic sits in a genuinely open research gap, giving a motivated student a real shot at a first-author publication.

### How to apply

Email a short expression of interest to Mengxuan.Zhang@anu.edu.au with:

  1. a brief statement on why this project interests you;

  2. your CV and academic transcript;

  3. any code, projects, or writing you’re proud of (optional but welcome).

### Related papers reading:

Privacy-Preserving Approximate Nearest Neighbor Search on High-Dimensional Data

Differentially Private In-Context Learning with Nearest Neighbor Search

RAGLeak: Membership Inference Attacks on RAG-Based Large Language Models

### This one-year project is under the co-supervision of Dr. Mengxuan Zhang (Australian National University) and Dr. Yanjun Zhang (Griffith University). Informal enquiries are warmly encouraged before applying. We’re happy to chat about whether the project is a good fit.

arrow-left bars magnifying-glass xmark