Communication-Efficient Federated Learning Based on Dynamic Gradient Compression and Client Selection

Federated Learning is a distributed machine learning paradigm designed to protect data privacy and security while allowing multiple participants to collaboratively train a shared global model without uploading their data to a central server.Modern machine learning models often consist of millions of parameters, resulting in significant computational complexity and a heavy communication burden during the collection and distribution of training data.This becomes a key bottleneck that limits the effectiveness of federated learning.In order to reduce communication costs, we first propose the CS client selection strategy, which integrates model performance and fairness to construct scheduling priority metrics.Client selection is then carried out based on these metrics.And then, we propose a dynamic gradient sparsification algorithm that adjusts the level of sparsification applied to gradient updates for each client, ensuring that every client uploads a suitably compressed update.Finally, we conduct comprehensive experiments and the experimental results show that the method proposed in this paper can effectively reduce communication costs while ensuring model accuracy.