Secret Vector Search: Secure and Private Content Querying with Semantic Embeddings
Contributor(s)
Madelyne Xiao, Sarah Scheffler, Micha Gorelick
Session
Cryptography and Privacy for the People
Abstract
We present SecretVectorSearch (SVS), a system that permits privacy-preserving calculation of summary statistics and trends over consensually-donated text messages from any source, including end-to-end encrypted messages. Our method protects the identities of individual message senders and the contents of individual messages with differential privacy guarantees as well as a query infrastructure utilizing homomorphic encryption that only allows queriers to receive counts of messages sufficiently similar to their query.