BC-GFL: blockchain-based graph federated learning
Yufei Xie, Zijiao Zhang, Lin Zhu · 2023
Graph Neural Networks (GNNs) can better handle non-euclidean space data, and when combined with Federated Learning (FL) can address their data volume shortage and privacy security issues during model training. However, the traditional centralized Graph Federated Learning architecture faces threats such as unreliable centralized servers and attacks from malicious clients, and there may be missing edge structures between subgraph data of different clients, which reduces the performance of GNN models. In this study, a blockchain platform is introduced to implement decentralized Graph Federated Learning to improve the robustness of models. An approximation evaluation mechanism for finding similar subgraphs through Transfer Learning (TL) is also proposed, which can discover subgraph clients with structural associations and aggregate models of similar subgraphs without leaking data, thus solving the structural missing problem. Simulation results show that compared with existing algorithms, the approximation evaluation mechanism in this paper can better compensate for the missing subgraph structure and significantly improve the accuracy of model prediction, and the introduction of blockchain also effectively prevents the influence of malicious nodes and ensures the stability of the training process.