Abstract: Subgraph federated learning trains on subgraphs of a global graph on local clients using graph convolutional networks and updates these parameters on the server, thereby safeguarding user privacy. Existing methods lack attention to important nodes in specific tasks or graph structures, which may reduce the efficiency of node embedding. This paper proposes a novel graph federated learning algorithm, called FedMFG, which introduces multi-scale information fusion convolution to integrate node features with neighbor information, thereby enhancing node feature representation capabilities. The algorithm passes parameters during the pre-training phase to reduce communication costs and applies an attention mechanism at the server to dynamically adjust weights for better aggregation of global parameters. Experimental results on standard benchmark datasets demonstrate that FedMFG achieves higher accuracy, higher stability, and lower communication costs than previous algorithms.

Key words: subgraph federated learning, multi-scale information fusion convolution, graph convolutional network(GCN) model, attention mechanism