A Cluster-Driven Adaptive Training Approach for Federated Learning.

Area of Science:

• Edge Computing
• Machine Learning
• Data Privacy

Background:

• Federated learning (FL) offers privacy-preserving collaborative training in edge computing, but practical issues like non-IID data and stragglers hinder widespread adoption.
• Existing FL methods struggle with statistical heterogeneity and varying client performance, limiting real-world applicability.

Purpose of the Study:

• To propose a novel cluster-driven adaptive training approach (CATA-Fed) for enhancing federated learning performance in practical edge environments.
• To address challenges of non-IID data distribution and straggler clients within federated learning systems.

Main Methods:

• CATA-Fed implements adaptive local model training to improve efficiency and mitigate straggler impact.
• Client clustering based on data size and proportional fair scheduling are used to reduce gradient magnitude differences and balance client loads.
• The approach specifically targets statistical heterogeneity, such as non-IID data, for improved global model updates.

Main Results:

• Extensive experiments on MNIST, Fashion-MNIST, and CIFAR-10 datasets demonstrate CATA-Fed's superiority over FedAVG, FedProx, and TiFL.
• CATA-Fed significantly improves both training speed and test accuracy under diverse federated learning conditions.
• The proposed methods effectively handle non-IID data and mitigate issues caused by straggler clients.

Conclusions:

• CATA-Fed provides an effective solution for practical federated learning implementation by addressing key challenges.
• The cluster-driven adaptive training approach enhances efficiency, accuracy, and robustness in edge computing environments.
• CATA-Fed represents a significant advancement for deploying federated learning in real-world applications.