Reproduce “Differentially Private Retrieval-Augmented Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tingting Tang, James Flemings, Yongqin Wang, Murali Annavaram
Researchers at the University of Southern California developed DP-KSA, a differentially private retrieval-augmented generation (RAG) algorithm that extracts keywords from sensitive data using a propose-test-release mechanism. This framework enabled RAG systems to leverage external knowledge with formal privacy guarantees, outperforming non-RAG baselines by up to ~3.5 F1 points while achieving privacy budgets (ε) as low as 2.
Key insights
- Many question-answering queries can be sufficiently answered by a limited set of keywords, allowing for a lower-dimensional semantic space where DP can be applied more effectively.
- The Propose-Test-Release (PTR) mechanism provides a practical way to privately select top-k items in high-dimensional spaces, releasing information only when robustly stable rankings can be maintained with minimal noise.
- Aggregating keyword frequencies from an ensemble of independently generated responses enhances the signal-to-noise ratio for relevant keywords, improving robustness to DP noise.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.