Decentralized federated learning model based on network propagation dynamics

This study presents a decentralized federated learning model based on network propagation dynamics (NET-D-DFL), designed to enhance the performance of traditional decentralized federated learning (DFL) with respect to communication efficiency. The NET-D-DFL model is based on network propagation dynamics, treating each node in the DFL as an independent entity within the dynamics framework. This approach enables dynamic information transmission between nodes. The state transition of each node is influenced by both its own state and the states of neighboring nodes. We conduct a comprehensive performance evaluation of the NET-D-DFL model using the MNIST public dataset , CIFAR-10 dataset and a dataset of Parkinson’s disease (PD) patients. The experimental results indicate that while the accuracy of the NET-D-DFL model may be slightly lower than that of the traditional NET-DFL model in certain scenarios, it reduces communication time and enhances communication efficiency. This advantage is particularly crucial in distributed learning environments characterized by high communication costs, as it effectively alleviates the communication burden and improves system scalability. In summary, the NET-D-DFL model introduces a novel approach to DFL by incorporating network propagation dynamics, making a contribution to improving communication efficiency.