Private Retrieval Augmented Generation via Random Projection
The authors report better retrieval and generation than differential privacy methods, with comparable privacy in attack tests.
Their method applies the same random projection to datastore embeddings and query embeddings. They prove that it preserves semantic similarity and report substantially reduced data extraction across multiple RAG architectures and datasets. The privacy comparison is empirical; the abstract does not claim the formal differential privacy guarantees of the competing methods.
ArXiv · AI/CL/LG's note
Their method applies the same random projection to datastore embeddings and query embeddings. They prove that it preserves semantic similarity and report substantially reduced data extraction across multiple RAG architectures and datasets. The privacy comparison is empirical; the abstract does not claim the formal differential privacy guarantees of the competing methods.
ArXiv · AI/CL/LG's note
score 4