Benchmarking Federated Learning in Edge Computing Environments: A Systematic Review and Performance Evaluation

A new comprehensive study by researchers Sales Aribe Jr. and Gil Nicholas Cagande provides a crucial performance evaluation of Federated Learning (FL) techniques specifically designed for edge computing environments. Published in the Journal of Advances in Information Technology, the paper systematically reviews and benchmarks state-of-the-art methods, categorizing them across four key dimensions: optimization strategies, communication efficiency, privacy-preserving mechanisms, and system architecture. This structured analysis offers a much-needed roadmap for a field where data privacy, low latency, and bandwidth constraints are paramount.

Using standard benchmarking datasets—MNIST, CIFAR-10, FEMNIST, and Shakespeare—the researchers assessed five prominent FL algorithms. The results deliver clear, actionable insights for practitioners. The SCAFFOLD algorithm emerged as the leader in model accuracy, achieving a score of 0.90, and demonstrated superior robustness when handling challenging non-IID (non-Independent and Identically Distributed) data, a common reality in distributed edge networks. In contrast, the widely used Federated Averaging (FedAvg) algorithm excelled in practical deployment metrics, showing the best performance in communication overhead and energy consumption, which are critical for battery-powered edge devices.

Despite these advancements, the review honestly addresses persistent challenges, including data heterogeneity, energy limitations, and issues with experiment repeatability. By identifying these gaps and providing a structured research agenda, the paper serves as both a performance guide and a call to action. It equips engineers and researchers with the data needed to select the right algorithm for their specific edge AI use case, balancing accuracy needs against system constraints like power and bandwidth.