Apple's Deployment of Homomorphic Encryption at Scale
summary
Deploys homomorphic encryption plus differential privacy and Oblivious HTTP to power real features: private Live Caller ID (iOS 18), photo landmark tagging and Mail business lookups, shipping at scale across a billion devices. Uses BFV (RNS) with SIMD packing and eval-key optimizations, achieves about 25,000 QPS and ~500 KB per query after a 10x response reduction, binds DP with epsilon=0.8 using 1–2 fake queries per client and runs as an auditable Swift/CoreCrypto stack.
AI-written summary · about
chapters (10)
- 0:00Homomorphic encryption intro
- 1:03Privacy tech need
- 3:52Visual Search overview
- 7:26Risks of direct embeddings
- 8:52Limitations of direct homomorphic encryption
- 13:06Differential privacy + HE
- 21:05HE optimizations for size
- 24:14Scale: 25K QPS
- 25:42Privacy tech summary
- 26:44Q&A
related talks
20m10 Weird Applications of Public Key Cryptography
23mBatchCrypt: Efficient Homomorphic Encryption for Cross-Silo Federated Learning
29mProtecting data vs systems: practicality, performance, and problems...
40mTrust No One
40mIt’s a Kind Of Magic: Computation on Ciphertexts
50mBuilding Production-Grade End to End Encrypted App...
41mCryptographic speed, reliability & security
52mMachine Learning over Encrypted Data with Fully Homomorphic Encryption