Fedavg-Db: A Dual-Loop Dynamic Batch Optimization Framework for Efficient and Robust Federated Learning

Queen Mary School Hainan, Beijing University of Posts and Telecommunications, Beijing, China

This email address is being protected from spambots. You need JavaScript enabled to view it.

Abstract

Federated Learning (FL), as a distributed machine learning paradigm, demonstrates significant potential in privacy-preserving scenarios but faces dual challenges of client data non-Independent and Identically Distributed (non-IID) characteristics and heterogeneous computing resources. Existing methods like FedAvg employ fixed batch sizes, leading to constrained convergence rates and suboptimal resource utilization. This paper proposes FedAvg-DB, a dynamic batch adjustment framework featuring a dual-loop optimization architecture: At the client level, a three-phase sliding window mechanism monitors loss improvement rate (r_k) and variance (σ_k) in real-time, establishing a threshold-triggered adaptive adjustment system; At the server level, an innovative aggregation strategy integrating median calibration and momentum smoothing effectively mitigates data distribution bias. Experiments on CIFAR-10/100 and FEMNIST datasets demonstrate that the proposed method achieves 18.3% faster convergence, 34.6% lower client loss variance, and 10.3% energy reduction compared to traditional FedAvg. Ablation studies confirm the synergistic effects of dynamic batch adjustment and learning rate square-root scaling rules, validating the framework's effectiveness in practical applications such as medical image segmentation (5.2% Dice coefficient improvement) and industrial predictive maintenance (12.7% RMSE reduction).