Heterogeneous Graph Federated Learning with Structure-Aware Data-Free Distillation

Subgraph heterogeneity is a critical challenge in federated learning, significantly impacting model performance. Data-free knowledge distillation overcomes the limitation of sharing private label distributions in federated learning, yet existing research lacks in-depth exploration of subgraph heterogeneity. Our work builds upon the fact that node and structural variations in heterogeneous graphs cause significant disparities in the reliability of knowledge from local graph neural networks. We propose a structure-aware bidirectional data-free federated distillation technique, where the generator and global model engage in an adversarial training process to ensure reliable bidirectional knowledge transfer between local and global models. Extensive experiments across multiple public datasets demonstrate the model's effectiveness.