Compass: Encrypted Semantic Search with High Accuracy
summary
Compass implements encrypted semantic search by running HNSW-style graph traversal on encrypted embeddings with a client-side retriever and ORAM-backed storage, claiming plaintext-matching accuracy on datasets up to 9M items while keeping queries, data, and access patterns hidden. It combines Directional Neighbor Filtering (≈10× network bandwidth reduction), Speculative Neighbor Prefetch (2–16× fewer round trips), and a Graph-Traversal Tailored ORAM white-box design to reach 0.6–1.3s end-to-end latencies and orders-of-magnitude speedups versus secure baselines.
AI-written summary · about
chapters (13)
- 0:00Compass introduction and goals
- 2:05Need for encrypted search
- 3:53Prior encrypted search limits
- 4:56HNSW for efficient search
- 5:58Compute-storage with ORAM
- 7:24HNSW over ORAM challenges
- 8:47Directional neighbor filtering
- 10:33Speculative neighbor prefetch
- 12:18ORAM for graph traversal
- 12:59Evaluation setup
- 13:42Baseline accuracy results
- 14:46Latency and efficiency
- 15:29Conclusion
related talks
18mSimEnc: A High-Performance Similarity-Preserving Encryption Approach for...
37mFrontiers in Cryptography
43mFederated Vector Retrieval Across Organizations With Milvus, OPA...
31mFast, highly effective database security using searchable encryption (without the yak shave)
01h 11mPrivate Information Retrieval and Oblivious RAM
21mMapSafe: A complete tool for achieving geospatial data sovereignty
01h 14mPrivate Information Retrieval and Oblivious RAM
36mThe Forbidden Index: How We Turned OpenSearch Into a Privacy-Preserving Superhero