Federated learning with adaptively sparsified and recoverable perturbation for internet of things

Abstract The Internet of Things (IoT) is transforming organizations by improving efficiency and productivity through real-time data analysis, which enables them to optimize their operations. It also facilitates predictive maintenance by detecting potential equipment failures, reducing costs and downtime. Federated learning empowers IoT by enabling intelligent and real-time decision-making while keeping sensitive industrial data localized and sharing only model parameters. However, susceptibility to adversarial attacks, the need for efficient model updates, and maintaining model performance across diverse environments remain critical challenges. In this paper, we introduce a clustered federated learning model that enhances privacy and prediction performance by applying recoverable sparsified perturbation to fragmented parameters. Each edge device splits the parameters into multiple fragments and applies a separate sparsified perturbation to each fragment to improve privacy. The server receives the fragmented perturbed parameters, clusters each fragment according to its distribution, and performs aggregation for each cluster. Once edge devices receive the fragmented aggregated parameters, a noise recovery approach is utilized to improve prediction performance. Following this, each fragment is merged, and the next round of training continues. The experimental evaluation conducted with real-world testbed datasets proves the effectiveness of the proposed approach over existing state-of-the-art methods.