Federated Learning over Non-IID Edge Devices

Abstract

Federated Learning (FL) enables decentralized model training across edge devices while preserving data privacy. However, the inherent non-Independent and Identically Distributed (non-IID) nature of data on these devices poses significant challenges to model convergence and performance. This paper reviews strategies developed before 2018 to address these challenges, focusing on data-sharing techniques, model aggregation methods, and communication-efficient algorithms. We analyze the effectiveness of these approaches in improving model accuracy and convergence speed under non-IID conditions. Our review highlights the trade-offs between privacy preservation and model performance, providing insights into the evolution of FL methodologies.