BatchCrypt: Efficient Homomorphic Encryption for Cross-Silo Federated Learning
summary
BatchCrypt encodes batches of quantized gradients into single long integers, uses two's-complement packing plus zeros padding and an analytical clipping method (DACIQ) so additive homomorphic aggregation works obliviously on ciphertexts. Implemented as a FATE plug-in and evaluated on geo-distributed EC2, it reports 23X-93X training speedup, 66X-101X communication reduction, 2048-bit HE compatibility and under 1% accuracy loss while enabling practical cross-silo federated training.
AI-written summary · about
chapters (18)
- 0:00Introduction
- 1:19Target
- 2:00CrossSilo
- 2:43Data Center Distributed Training
- 3:09PrivacyPreserving Techniques
- 7:02BatchCrypt
- 8:20Aggregation
- 8:49Quantization
- 10:47Quantization Requirements
- 15:39Gradient Clipping
- 16:57BatchCrypt Implementation
- 18:03Batch Script Quantization
- 19:04Computation Breakdown
- 19:40Communication Overhead
- 20:12Results
- 20:42Comparison
- 21:22Conclusion
- 22:00Outro
related talks
33mThe Promise of Trustworthy AI
52mMachine Learning over Encrypted Data with Fully Homomorphic Encryption
37mOpen-source Machine Learning on Encrypted Data
18mFLASH: Towards a High-performance Hardware Acceleration Architecture for Cross-silo...
06mLightning Talk: Multi-Party Learning Using Homomorphic Encryption with Kubeflow...
40mIt’s a Kind Of Magic: Computation on Ciphertexts
42mFederated Learning: Unlocking the Values of Discrete Data for Enterprise
30mApple's Deployment of Homomorphic Encryption at Scale