FedDAAW: Dynamic client selection and accuracy-adaptive aggregation for cross-task federated learning under heterogeneity
Abstract
Background. Federated learning (FL) enables collaborative model training across distributed clients while preserving data privacy; however, its performance is often compromised in heterogeneous environments due to statistical divergence and system variability among clients. Most existing methods address heterogeneity along a single axis—such as convergence robustness or model personalization—and are primarily validated on classification tasks, with little evidence of generalizability to broader learning paradigms like regression.
Methods. To address these limitations, we propose FedDAAW (Federated Learning with Dynamic Client Selection and Accuracy-Adaptive Aggregation), a lightweight yet effective framework for heterogeneous FL. FedDAAW introduces a multi-factor dynamic client selection mechanism that identifies high-value participants based on historical validation accuracy, computational resources, and data similarity. It further employs an accuracy-adaptive aggregation strategy that assigns weights to local updates using temporally smoothed historical accuracy, amplifying the contribution of consistently high-performing clients without requiring gradient sharing, contrastive learning, or auxiliary public data.
Results. We evaluate FedDAAW across four benchmark datasets spanning diverse machine learning tasks: image classification (MNIST, CIFAR-10), text sentiment analysis (IMDB), and regression (California housing, Cal_housing). Under highly non-IID settings, FedDAAW improves test accuracy by 2.61% on CIFAR-10 and 3.80% on IMDB compared to FedAvg, while reducing mean squared error by 6.5% on Cal_housing. It also achieves faster convergence, reaching stable performance up to five communication rounds earlier than baselines. Conclusions. FedDAAW demonstrates consistent gains in accuracy, convergence speed, and cross-task robustness under heterogeneity. By leveraging only locally computable metrics, it offers a practical, privacy-preserving, and easily deployable solution that advances the state of the art in heterogeneous federated learning without complex architectural modifications.