This repository contains an enhanced implementation of BatchCrypt based on the ATC'20 paper:
"BatchCrypt: Efficient Homomorphic Encryption for Cross-Silo Federated Learning"
Original authors: ATC'20 Paper
Enhanced by Geonha Kim, Hankuk University of Foreign Studies.
- Implements a novel "BatchCrypt Zero-Skipping" optimization that skips encryption of all-zero gradient blocks.
- Demonstrates significant reductions in training time with minimal impact on accuracy.
- Includes clipped quantization, batching, and homomorphic encryption based on the Paillier cryptosystem.
detail features in: BatchCrypt/paper/Contribution_Overview.pdf
- Python 3.7+
- TensorFlow 2.x
This repository contains the original implementation of the BatchCrypt Zero-Skipping method that served as the foundation for the following research paper:
Optimizing Homomorphic Encryption in Federated Learning with Zero-Skipping
Accepted at [IEEE ICCE 2026]
Authors: Yoo-Bin Tae, Su-Jeong Park, Geon-Ha Kim, Seung-Ho Lim
The core implementation of BatchCrypt Zero-Skipping was developed by Kunha Kim, including:
- Implementation of zero-skipping homomorphic encryption
- Block-level gradient sparsity detection
- Integration with BatchCrypt batching framework
- Federated learning experiment pipeline
The accepted paper extends this implementation by introducing skip-threshold–based experimentation and additional empirical analysis.
paper in: BatchCrypt/paper/research_paper.pdf