An efficient federated learning framework for graph learning in hyperbolic space

With the increasing number of graph data, Graph Federated Learning (GFL) has emerged and been used in medicine, chemistry, social networks and other fields. Consequently, the efficiency of graph classification has become a crucial issue in the GFL framework. However, due to the high distortion and redundancy in graph information, the existing works are troubled by the low accuracy of classification. In this paper, we propose a novel efficient GFL framework for graph classification, namely FedHGCN. FedHGCN has two novel features: (1) collaboratively train Graph Neural Network (GNN) in a high-dimensional space to capture the rich hierarchical feature of graphs. (2) build a strategy of node selection to remove the redundancy from the graph representation and highlight key nodes. Our extensive experiments show that FedHGCN outperforms the state-of-the-art approaches up to 15.6% by accuracy on four publicly available graph datasets. Furthermore, we prove that FedHGCN can efficiently deal with various poisoning attacks by experiments.