Abstract

Federated Learning (FL) is an important approach for decentralized machine learning in edge environments, enabling model training across distributed devices while preserving data privacy. However, its deployment in mobile edge networks faces challenges due to communication overhead, limited computational resources and constrained device capabilities. This paper proposed an optimized FL framework leveraging knowledge distillation and magnitude-based pruning to address these challenges. Knowledge distillation transfers the predictive capabilities of a larger model (teacher) to a smaller, resource-efficient model (student), while pruning eliminates redundant model parameters, reducing computational and communication demands. Through experiments on CIFAR-10 with MobileNetV2 and ResNet-50, we demonstrate that the proposed framework achieves significant reductions in model complexity and communication overhead while maintaining high accuracy. The integration of knowledge distillation improved student model accuracy by 7%, and further pruning achieved over 50 % compression with minimal degradation in performance. Compared to baseline approaches like ZeroFL, our method exhibited superior accuracy retention under higher compression ratios. This work highlights the potential of combining knowledge distillation and pruning for scalable, efficient FL systems in resource-constrained mobile edge environments, paving the way for further integration of complementary compression techniques.