FedPA: A Performance-Based Approach for Resource-Optimized and Imbalance-Resilient Federated Learning

Document Type

Article

Publication Date

1-1-2026

Abstract

Federated Learning (FL) has emerged as a key paradigm for privacy-preserving collaborative model training, particularly in data-sensitive domains such as healthcare, telecommunications, and finance. However, conventional FL aggregation methods including Federated Averaging (FedAvg) and Federated Stochastic Gradient Descent (FedSGD) treat all client updates equally making them vulnerable to data heterogeneity, client imbalance, and low-quality updates in non-IID settings. These limitations often result in degraded global model performance, slow convergence, and unnecessary communication overhead. To address these challenges, this article proposes Performance-Based Federated Partial Averaging (FedPA). FedPA is a selective aggregation strategy that incorporates client updates based on local performance indicators such as accuracy and macro-F1 score. Unlike FedAvg, FedPA aggregates only high-performing client updates while still synchronizing underperforming clients with the global model, enabling progressive improvement without permanent exclusion. This performance-aware design reduces the influence of noisy or unreliable updates, accelerates convergence, and improves communication efficiency particularly in the early training stages. FedPA is evaluated on the MNIST dataset under a simulated non-IID federated setting. Experimental results demonstrate that FedPA achieves competitive accuracy and F1 performance while requiring fewer communication rounds than FedAvg. Moreover, FedPA exhibits robustness to client-side class imbalance and heterogeneous data distributions, yielding more stable convergence and improved fairness across clients. In addition, we discuss an extension of FedPA, termed Performance- and Resource-Aware Federated aggregation (FedPA-RA). Which incorporates system-level resource indicators to further control communication overhead in resource-constrained environments. In general, the proposed framework highlights the effectiveness of performance-aware aggregation for scalable and adaptive federated learning. The results suggest that FedPA and its resource-aware extension are well-suited for deployment in edge and wireless network environments, including emerging 5G/6G systems, where communication efficiency and heterogeneous client conditions are critical considerations.

Keywords

5G, federated learning, FedPA, FedPA-RA, industrial growth, Industry innovation, resource optimization

Publication Title

IEEE Access

DOI

10.1109/ACCESS.2026.3665401

Volume

14

First Page

42105

Last Page

42118

Share

COinS