Resource-and Data-Aware Client Selection for Heterogeneous Edge Federated Learning
Abstract
This paper investigates client selection for federated learning (FL) in heterogeneous edge networks with statistical and system heterogeneity. A Resource- and Data-Aware (RDA) method is proposed to jointly consider data representativeness, training latency, energy consumption, and participation fairness. The four factors are normalized into a unified utility score, while a greedy mechanism dynamically updates marginal data utility during selection. Experiments on non-IID Fashion-MNIST compare RDA with Random, Fastest, Energy-aware, and Fairness-aware strategies. At a participation rate of 0.5, RDA achieves a final test accuracy of 0.8642 while reducing training time and energy consumption by 27.5% and 20.2%, respectively, relative to Random. It also reduces the time and energy required to reach 85% accuracy by 24.8% and 22.4%. Ablation and weight-sensitivity analyses confirm the complementary roles of data, resource, and fairness information. Overall, RDA provides a balanced trade-off among learning performance, resource efficiency, and participation fairness in heterogeneous edge FL.