FedWDP: A Wasserstein-distance-based federated learning for privacy and heterogeneous data in IoT

Abstract

As the need for interconnected devices and data exchange grows in the Internet of Things (IoT), traditional centralized data processing methods increasingly struggle to maintain privacy and adapt to the diverse and dispersed nature of IoT devices. Federated learning, a decentralized approach to machine learning, presents a viable solution to these challenges. Yet, the varied nature of IoT data and stringent privacy requirements introduce unique obstacles for federated learning. This paper introduces FedWDP, a federated learning method specifically designed for IoT privacy needs and heterogeneous data. FedWDP uses the Wasserstein distance to quantify the gap between local and global parameters, integrating this measure as a regularization term in the loss function to reduce model discrepancies and improve accuracy. To further balance privacy and usability, an exponential decay strategy is implemented, allowing for adaptive distribution of differential privacy noise. For better performance on high-dimensional data, PCA-FedWDP is proposed, which combines principal component analysis (PCA) with differentially private federated learning to perform dimensionality reduction. Experimental results on non-IID datasets reveal that this approach significantly enhances both accuracy and availability for heterogeneous data while safeguarding user privacy. This study thus provides a valuable framework for applying federated learning in IoT settings, contributing to the secure and intelligent use of IoT data in both theoretical and practical contexts.