FedLASE: Performance-balanced system-heterogeneous FL via layer-adaptive submodel extraction

Federated Learning (FL) has gained significant attention for its privacy-preserving capabilities in distributed learning environments. However, the inherent system heterogeneity across edge devices brings significant challenges in deploying a unified global model. Although many submodel extraction methods are designed to address these challenges by selecting a subset of parameters from the global model to accommodate client constraints, our experiments show that existing submodel extraction methods exhibit significant performance discrepancies between submodels with different resource levels, limiting the overall performance of the federated learning system. To overcome these limitations, we propose FedLASE - a novel Layer-Adaptive Submodel Extraction framework that selects important parameters while preserving the structural integrity of the client models, thereby achieving balanced performance across heterogeneous FL clients and improving the convergence. Specifically, our approach quantifies layer importance based on parameter importance and hierarchically extracts critical parameters within each layer while strictly satisfying resource constraints. Theoretically, we rigorously analyze the convergence of FedLASE and investigate the influence of system heterogeneity on its performance. Extensive experiments demonstrate the superiority of FedLASE over the state-of-the-art methods and its robustness across various system-heterogeneous scenarios.