Joint Optimization of Model Pruning and Bandwidth Allocation for Resource-Constrained Wireless Federated Learning
Abstract
Wireless Federated Learning (FL) enables User Equipments (UEs) to collaboratively train a shared model under the coordination of the Base Station (BS) without exposing their private data. However, due to limited computation and communication resources, some UEs may be unable to train the full model, causing the global model to miss important data features and resulting in degraded performance. Even when training is possible, UEs often cannot complete each round at the same pace, leading to the synchronization barrier and increased overall time. In this paper, we propose FedPB, a framework that accelerates wireless FL by dynamically adjusting the pruning ratios and bandwidth allocation. Specifically, we formulate a joint pruning ratio and bandwidth allocation optimization problem that minimizes the total completion time under resource constraints while ensuring convergence. We theoretically analyze the convergence behavior, and based on the analysis, decompose the joint optimization problem into subproblems, which are solved iteratively. Extensive experiments on both CNN-based vision tasks and Transformer-based language tasks demonstrate that FedPB achieves at least a $1.6\times $ speedup under resource constraints while preserving the performance of the global model.