Document Type : Original Article
Authors
1 Department of Computer Engineering, Faculty of Engineering, University of Guilan, Rasht, Iran.
2 Dep. Of Computer Engineering, University Of Guilan, P.O. Box 3756
Abstract
The adoption of deep learning in healthcare for disease diagnosis is hindered by limited data availability due to strict patient privacy regulations. Federated learning (FL) enables collaborative training without raw data sharing, yet remains vulnerable to reconstruction attacks such as GAN-based inversion, especially from honest-but-curious central servers. Existing defenses like differential privacy trade model utility for privacy, while encryption adds heavy overhead. This paper proposes FedPrEx, a decentralized FL framework with pruning and exchange that eliminates the central aggregator. FedPrEx introduces three innovations: (1) randomized pruning where each client randomly drops 20% of model parameters before transmission to disrupt gradient inversion; (2) secure peer-to-peer exchange among edge servers using symmetric encryption and one time random pairing; and (3) hierarchical decentralized aggregation that prevents any single entity from accessing consecutive updates. We evaluate FedPrEx on MNIST and eICU datasets. Compared to the state of the art Fed Select framework, FedPrEx achieves higher accuracy (98.3% vs. 96.1% on MNIST) and more stable convergence without oscillations. Unlike differential privacy, FedPrEx imposes low computational overhead via pruning instead of noise addition, while resisting GAN based and collusion attacks through randomized exchange and periodic central model evaluation. Security analysis shows that the decentralized architecture with encrypted peer exchange and dynamic participation effectively mitigates internal and external privacy threats. FedPrEx offers a practical, scalable, privacy preserving solution for training deep learning models on distributed healthcare data without compromising model quality.
Keywords
)