Carousel scores a billion encrypted vectors without client index state
Original title: CAROUSEL: GPU-Accelerated Private Vector Search via Homomorphic Sketching
On BIGANN SIFT1B it reports recall@100 of 0.9715; 100 million vectors take 3.735 seconds on one GPU.
Carousel is a homomorphic private vector-search design that scores the full corpus under encryption and compresses results for client-side ranking. The server raises encrypted scores to a power that preserves order and suppresses most values, then packs the sparse result with homomorphic additions. The paper says clients hold no database-dependent state, so inserts and deletes apply after one vector write, and reports no recall loss from discarding candidates.
- It searches 1 billion BIGANN SIFT1B vectors across multiple GPUs.
- Recall@100 is 0.9715 on that billion-vector corpus.
- At 100 million SIFT vectors, recall@100 is 0.9845.
- That smaller search takes 3.735 seconds on a single GPU.
- Compression uses homomorphic additions into a logarithmic number of ciphertexts.
From the original article
1 min read at arxiv.orgVector search on untrusted infrastructure exposes the \emph{query embedding}, a faithful summary of the user's intent. Fully homomorphic encryption can hide the query, but existing approaches still require a full scan followed by ranking of encrypted scores.
By Sohaib · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.