HalfFedLearn: A secure federated learning with local data partitioning and homomorphic encryption

Abstract

Federated Learning (FL) is an emerging technology in collaborative machine learning, where multiple data owners train a unified model by exchanging model parameters instead of private data. Despite providing data privacy and a wide range of applications, FL faces several challenges, such as slow convergence, high computation and communication costs, and security in parameter sharing. In this paper, we propose HalfFedLearn to address these challenges using Homomorphic Encryption (HE) and local horizontal data partitioning. We leverage the inherent distribution of the dataset to use horizontal data partitioning based on data sensitivity and enforce selective security on private data samples using HE. HalfFedLearn minimizes the data volume per client, which reduces local training time and computation. Also, the number of communication rounds decreases due to the reduction in local dataset size. We experimented on MNIST, CIFAR-10, and FMNIST datasets with varying clients and number of rounds. The results demonstrate that HalfFedLearn increases accuracy by 3%–6% and achieves an average reduction of 29.33% in training rounds, with a maximum training time reduction of 9.94% per round. Additionally, we performed a comparative analysis of the computation, communication cost, and security that shows the efficacy of HalfFedLearn.