Combining Machine Learning and Homomorphic Encryption in the Apple Ecosystem
Summary
Apple describes a privacy-preserving ML stack that uses homomorphic encryption (BFV) to run computations on encrypted data, enabling private server lookups and private nearest neighbor search (PNNS) for on-device ML. The article covers PIR, PNNS, differential privacy (ε=0.8, δ=1e-6) with OHTTP relay, and an open-source library swift-homomorphic-encryption to facilitate adoption. It highlights production considerations such as embedding quantization, data sharding, and end-to-end privacy guarantees.