FedBFGCN: A Graph Federated Learning Framework Based on Balanced Channel Attention and Cross-Layer Feature Fusion Convolution

Graph Federated Learning (GFL) is an emerging distributed training paradigm that combines federated learning with graph data. Due to its ability to effectively handle complex and heterogeneous graph data while protecting user privacy, GFL has shown great potential in processing various types of graph structures and has been proven effective in a wide range of applications. However, existing methods normally assign equal attention to all nodes within a single graph, focusing too much on the information of neighboring nodes, even if some nodes are more important in the graph structure or task (such as high consumption users or popular products), which inevitably leads to inefficient node embedding. To address this issue, this paper proposes an innovative graph federated learning framework called FedBFGCN (Graph Federated Learning Based on Balanced Channel Attention and Cross-Layer Feature Fusion Convolution) to optimize the embedding and analysis efficiency of graph data. This proposed framework converts single graph data into node features and adjacency matrices for processing, and combines a customized Cross-Layer Feature Fusion Convolution(CLF) and an improved Attention Mechanism that is Balanced Channel Attention Mechanism (BCAM). The FedBFGCN improves the attention to important nodes by dynamically weighting and adjusting the weights of features through BCAM; Using CLF effectively integrates its own features with neighbor information, enhancing feature expression capability. Through the organic fusion of these two modules, the FedBFGCN achieves efficient, robust, and more comprehensive node embedding representation, demonstrating excellent performance in node classification and prediction tasks. In addition, this framework also uses homomorphic encryption methods to enhance privacy protection and improve data security. The FedBFGCN was evaluated on standard reference network datasets (Cora, Citeseer, Polblogs), and experimental results showed that it has lower losses and higher performance in multiple aspects. This framework is capable of addressing various challenges in graph federated learning, significantly improving learning effectiveness and application capabilities. This study not only provides new ideas for graph federated learning and GCN, but also demonstrates its enormous potential in practical applications.