“Confidential Vector Search: Knowledgebase Homomorphic Encryption” introduces a practical path to RAG systems that can search sensitive embeddings without ever revealing them. Building on the SIAM study “Maturing Homomorphic Encryption (HE) to Enable Privacy Preserving Vector Search,” Sulimon Sattari will unpack how techniques like dimensional scrambling, noise injection, CKKS, and chaotic mapping can be combined with new schemes such as DIEHARD and ROME to preserve inner products while keeping queries and documents encrypted.
The session will translate the paper’s math into system design: how to encrypt a knowledge base on an untrusted host, run similarity search directly in ciphertext space, and still achieve usable performance for real‑time applications. Sattari will walk through threat models (Bob–Eve–Alice), benchmark results comparing different HE approaches, and what it would take to integrate homomorphic vector search into Kwaai-style personal AI and healthcare scenarios where confidentiality is non‑negotiable.
The presentation will take place in Ballroom F on Thursday, March 5, 2026 - 14:40 to 15:00
