Random projection defense against sensitive information extraction from RAG retrieval corpora
Start here
A paper released on October 16, 2026 proposes using random projection to protect RAG knowledge base privacy: applying the same projection to knowledge base embeddings and query embeddings, the authors claim it can provably preserve semantic similarity between queries and retrieved items while significantly mitigating data extraction attacks. Across multiple RAG architectures and datasets, the method outperforms existing approaches that rely on formal differential privacy (DP) in retrieval and generation performance, with comparable empirical privacy under strong attack models. The authors claim this result is the first to show random projection can serve as a practical baseline for privacy-preserving RAG.
Generated by AI from reports · Updated 27 days ago
Attention and developments
Attention, hourly
Now 29, 0% in 24 hours·Peak 29 (Nov 6 03:00)
Related research
This event's paper
- Private Retrieval Augmented Generation via Random Projection
arXiv · 2026-10-15 · Deep reading
Progress timeline
- Paper TrackerPrivacy-preserving retrieval-augmented generation via random projection
The paper proposes a method to protect RAG knowledge base privacy using random projection: applying the same projection to both knowledge base embeddings and query embeddings. The authors claim it provably preserves semantic similarity between queries and retrieved items while significantly mitigating data extraction attacks. Across various RAG architectures and datasets, the method achieves better retrieval and generation performance than existing methods relying on formal differential privacy (DP), with comparable empirical privacy under strong attack models. The authors claim this is the first result showing random projection can serve as a practical baseline for privacy-preserving RAG.