Onyx Achieves Cost-Efficient Privacy-Preserving Vector Search by Hiding Disk Access Patterns in TEEs

arXiv·high signalXBlueskyLinkedInCopy link

Onyx addresses a critical gap in private AI infrastructure: approximate nearest neighbor (ANN) search on sensitive data in TEEs leaks user queries through SSD disk access patterns even when computation is protected. The system introduces disk-oblivious techniques that hide access patterns on external SSDs while maintaining cost efficiency, unlike prior ORAM-based approaches that are prohibitively expensive. Relevant to any team deploying vector databases for sensitive data (healthcare, finance, legal) on third-party cloud infrastructure.