A Unified Framework for Communication-Efficient and Privacy-Preserving Federated Learning Using Adaptive Differential Privacy and Sparse Model Aggregation

Abstract

Federated Learning (FL) has emerged as a promising distributed learning paradigm that enables collaborative model training without sharing raw data. However, conventional FL frameworks suffer from significant communication overhead and privacy leakage through exchanged model updates, particularly in heterogeneous client environments. To address these challenges, this paper proposes ADP-SFed, a unified communication-efficient and privacy-preserving federated learning framework that integrates Adaptive Differential Privacy (ADP) with Sparse Model Aggregation (SMA). The proposed framework dynamically adjusts the privacy budget across communication rounds to achieve an effective privacy–utility trade-off while transmitting only the most significant model parameters to reduce communication cost. Experiments were conducted on the MedMNIST and CIFAR-10 benchmark datasets under federated learning settings and compared with state-of-the-art methods, including FedAvg, FedProx, DP-FedAvg, Sparse-FedAvg, and FedAdam. Experimental results demonstrate that ADP-SFed achieved the highest classification accuracies of 95.82% on MedMNIST and 91.43% on CIFAR-10, with corresponding AUC values of 0.987 and 0.973, respectively. Furthermore, the proposed framework reduced communication cost by 67.86%, converged within 65 communication rounds, and achieved the lowest training time (128 min) and inference time (14.1 ms) among the compared methods. The adaptive privacy scheduling mechanism effectively preserved data privacy while maintaining high predictive performance, and sparse aggregation significantly improved communication efficiency. These results demonstrate that ADP-SFed provides a scalable, efficient, and privacy-preserving federated learning solution suitable for distributed medical image analysis, computer vision, and resource-constrained edge computing applications.