Adaptive Batch Homomorphic Encryption for Joint Federated Learning in Cross-Device Scenarios

Cross-device federated learning (FL) enables the privacy-preserving and collaborative training of machine learning models across heterogeneous clients. To prevent gradient information leakage, homomorphic encryption (HE) has been widely utilized due to its strong protection without losing accuracy. However, our experiments demonstrate that for clients with heterogeneous data and system capabilities, previous plain HE methods (i.e., encryption applied per gradient) and batch HE methods (i.e., encryption applied per batch of gradients) either significantly prolong training time or suffer from accuracy loss due to gradient quantization. We propose an adaptive batch HE framework for cross-device FL, which determines cost-efficient and sufficiently secure encryption strategies for clients with heterogeneous data and system capabilities. By leveraging the sparsity of convolutional neural networks for privacy-preserving similarity measurement of clients’ data, we first split the clients with similar data into their respective clusters. Then, we develop a fuzzy logic-based method to determine a cost-efficient and sufficiently secure HE key size for each client corresponding to its system capability, and arrange the clients with identical key size to the same groups. Finally, we design an efficient batch encryption approach for accuracy-lossless model aggregation. Extensive experiments on multiple heterogeneity scenarios demonstrate that our framework achieves comparable accuracy to plain HE, while reducing training time by $3\times $ – $31\times $ , and communication cost by $45\times $ – $66\times $ .