FLaTEC: An efficient federated learning scheme across the Thing-Edge-Cloud environment

Abstract

Federated Learning (FL) has become a cornerstone for enabling decentralized model training in mobile edge computing and Internet of Things (IoT) environments and maintaining data privacy by keeping data local to devices. However, the exponential growth in the number of participating devices and the increasing size and complexity of Machine Learning (ML) models amplify FL’s challenges, including high communication overhead, significant computational and energy constraints on edge devices, and the issue of heterogeneous data distribution, i.e., non-Independent and Identically Distributed (non-IID) data across clients. To address these challenges, we propose FLaTEC, a novel FL system tailored for the Thing-Edge-Cloud (TEC) architecture. First, FLaTEC introduces a split-training architecture that divides the global model into three components: a lightweight base model trained on resource-constrained edge devices, a computationally intensive core model trained on edge servers, and a simplified core model designed for on-device training and inference. Second, FLaTEC adopts a separate training strategy in which feature data is uploaded periodically from devices to edge servers to train the core model, reducing frequent data exchanges and mitigating the non-IID problem in FL. Third, to enhance the performance of the simplified model used for on-device training, FLaTEC applies knowledge distillation from the core model trained at the edge. A cloud server orchestrates the entire system by aggregating the base, core, and simplified core models using a federated averaging algorithm, ensuring consistency and coordination across devices and edge servers. Extensive experiments conducted across multiple datasets and diverse ML tasks validate FLaTEC’s superior performance, demonstrating its ability to achieve high accuracy, reduced communication overhead, and resilience to data heterogeneity compared to state-of-the-art methods.