Communication-efficient and Privacy-enhanced Federated Graph Learning via GAN-based Data Augmentation

Graph neural network (GNN) is widely used to solve challenging problems in wireless communications, but traditional centralized training of GNN models may pose significant privacy risks. Federated graph learning offers a reliable solution for safeguarding sensitive information within subgraphs in each client. Nonetheless, there may be connections between subgraphs, and existing federated graph learning frameworks generally require additional client-server communication to reconstruct those missing graph information (e.g., edges and nodes), resulting in high communication overhead and privacy leakage risks. To address this issue, we propose a communication-efficient and privacy-enhanced Federated Graph Learning framework based on generative adversarial networks (GANs), termed FGL-GAN. Specifically, FGL-GAN introduces a GAN-based graph data augmentation module to synthesize realistic neighbor nodes between subgraphs, thereby alleviating local missing graph information and enhancing model performance. Notably, the training and execution of the entire graph data augmentation module are performed locally on each client, which significantly reduces the server-client communication overhead and ensures privacy protection. Extensive experiments demonstrate the effectiveness of FGL-GAN across a diverse set of benchmarks.