Federated Learning With Dynamic Pruning for Efficient Model Aggregation in Heterogeneous Environments

With the widespread adoption of 5G networks, Federated Learning (FL) has gradually become an important paradigm for distributed artificial intelligence. However, clients in FL environments often exhibit significant heterogeneity in computational resources and data distributions, and under Non-Independent and Identically Distributed (Non-IID) scenarios, such heterogeneity not only slows model convergence but also affects the final accuracy. To reduce the computational burden on clients, this study adopts structured pruning by trimming filters in convolutional neural networks (CNNs) and dynamically adjusting pruning ratios according to client resources. Nevertheless, variations in model structures resulting from different pruning ratios make it difficult for traditional FL aggregation to maintain consistency. To address this issue, we propose FedCHAP (Federated Customized Heterogeneous Aggregation with Pruning), which integrates dynamic pruning with a heterogeneous aggregation mechanism. FedCHAP enables clients to flexibly adjust pruning ratios based on their available resources while ensuring the feasibility of aggregation through padding weight and structural alignment. Experiments conducted on a network traffic classification dataset under multiple Non-IID scenarios demonstrate that, in the most extreme case of quantity-based label imbalance, FedCHAP reduced the overall FL runtime by 8.1% and improved global accuracy by 4% compared with applying a fixed 30% pruning ratio to all clients. Furthermore, we benchmark the performance of FedCHAP against state-of-the-art FL methods and demonstrate its generalizability across various application domains. Overall, the findings validate that FedCHAP can effectively lower resource consumption while preserving model performance and stability.