FedGraph-KD: An Effective Federated Graph Learning Scheme Based on Knowledge Distillation
Shiyu Wang, Jiahao Xie, Mingming Lu, Naixue Xiong
Abstract
Graph Neural Networks (GNNs) have achieved success in a variety of domains due to their potent graph-data processing skills. However, gathering graph structure data from various universities and applying GNNs for centralized training is highly challenging due to privacy concerns and regulatory limitations. As a solution, Federated Graph Neural Networks (Fed-GNNs) do not require sharing data but support collaborative training of public models by sharing parameters or features between multiple parties. Thus, Fed-GNNs has gained more attention recently. However, existing Fed-GNNs schemes do not consider the problem of participants of the public model often having different private GNN models, i.e., the model heterogeneity problem, which can lead to failure in the model heterogeneity scenario. To address this issue, this paper explores an effective and novel Federated Graph Learning scheme Based on Knowledge Distillation models (FedGraph-KD). On one hand, each client trains its local models through knowledge distillation. On the other hand, this paper uses a federated learning framework to update the shared model parameters. Extensive experiments and analyses on several different graph classification datasets demonstrate the effectiveness of our approach.