ComFLEX: A Communication efficient FL training framework for Edge computing environments
Satya Srinivasa Arun Kumar Chavali, Akarsh K. Nair, Jayakrushna Sahoo · 2025
The increasing usage of Internet of Things devices generates vast amounts of data, raising significant concerns over data privacy and communication constraints, particularly in decentralized machine learning systems. Federated learning has emerged as a privacy-preserving solution, enabling collaborative model training without centralized data aggregation. However, traditional FL methods face challenges in handling non-independent and identically distributed data, high client heterogeneity, and significant communication overhead. To address these limitations, we propose the ComFLEX framework integrating concepts of gradient sparsification, an adaptive learning rate mechanism, and a step-ahead error feedback mechanism. These innovations significantly enhances convergence speed and communication efficiency. Through extensive experiments on the MNIST and CIFAR-10 datasets, under both IID and non-IID client distributions, our framework achieves up to 1.2 X faster convergence and only requiring 35% of communication cost compared to conventional FL methods. Additionally, the Com-FLEX framework demonstrates robustness to varying client participation rates and high data divergence. Ablation studies also prove the viability and practical applicability, achieving notable improvements in training time and communication efficiency. These results establish the ComFLEX framework as a promising solution for optimizing FL in heterogeneous and resource-constrained environments.