A Decentralized Federated Learning Model Based on Global Channel Selection

With the increasing challenges of data privacy protection and computational resource constraints, Decentralized Federated Learning has gained widespread attention as an emerging distributed learning framework. However, frequent communication and model updates in decentralized environments significantly increase the communication overhead of the system. To address this problem, this paper proposes a decentralized federated learning scheme with a global channel selection algorithm. In this scheme, nodes dynamically decide which channels should be shared by locally training and evaluating the importance of model channels and aggregating global channel scores through a voting mechanism of neighboring nodes. During the aggregation process, we pass only the parameters of the important channels to reduce the communication burden, and introduce a Learn-to-Rank regularization strategy in the channel selection optimization to further improve the accuracy and performance of the sparse model. Experimental results show that the proposed scheme is effective in reducing communication overhead and maintaining high model performance across multiple datasets and model architectures, especially in resource-constrained distributed environments, demonstrating strong adaptability and scalability. The approach provides new ideas and effective technical support for the application of decentralized federated learning in large-scale, heterogeneous environments.