With the rapid development of the Internet of Things (IoT) and edge intelligence, the volume of data generated by edge devices has grown explosively. Federated learning (FL), characterized by the paradigm of “data remaining local while models are shared,” has emerged as a key approach for adapting to the distributed architecture of edge computing, breaking down data silos, and enabling privacy preservation. However, its practical deployment in edge computing environments still faces significant challenges, including limited device resources and pronounced data heterogeneity. Existing pruning strategies for federated learning are predominantly based on static and single-design schemes, making it difficult to achieve a balanced trade-off among training overhead, communication cost, and model accuracy. To address these issues, this paper proposes FedAHP (Federated Learning with Adaptive Hybrid Pruning), an efficiency optimization scheme for federated learning based on adaptive hybrid model pruning. On the client side, an adaptive pruning mechanism driven by training states is designed to dynamically adjust pruning behavior during local training. On the server side, a counter-based heterogeneous aggregation method is adopted to efficiently align updates from clients with different pruning rates, thereby avoiding additional communication overhead. Furthermore, after training becomes stable, a performance-aware periodic structured pruning strategy is introduced to compress the global model scale and reduce subsequent training costs. Experimental results demonstrate that FedAHP maintains high model accuracy on the MNIST, CIFAR-10 and CIFAR-100 datasets while significantly reducing per-round communication overhead and time cost, making it well suited to the resource-constrained requirements of edge computing scenarios.
Hu et al. (Thu,) studied this question.