Private Retrieval Augmented Generation via Random Projection

Key points

• Retrieval-Augmented Generation (RAG) enhances the capabilities of large language models (LLMs) by querying external structured knowledge.1
• However, it can also introduce privacy risks by leaking sensitive information from the retrieval datastore.1
• We propose a simple method to preserve datastore privacy in RAG systems via random projection.1
• By applying the same projection to both datastore embeddings and query embeddings, our method provably preserves semantic similarity between queries and retrieved items while substantially mitigating data extraction attacks.1

Sources (1)

• [1]Private Retrieval Augmented Generation via Random Projection
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 05:45 PM

  Across multiple RAG architectures and datasets, we show that this lightweight approach achieves superior retrieval and generation performance compared to prior methods which leverage formal differential privacy (DP) guarantees, while exhibiting comparable empirical privacy under strong attack models.

  Retrieval-Augmented Generation (RAG) enhances the capabilities of large language models (LLMs) by querying external structured knowledge.

Extractive summary: sentences quoted from the sources.