PPFedGNN: An Efficient Privacy-Preserving Federated Graph Neural Network Method for Social Network Analysis

Exploring the intrinsic value of social data has long been a focal point for researchers. Presently, diverse social network data is dispersed across various platforms. While federated learning enables collaborative training with multiple clients, enhancing model performance while safeguarding client-specific information, it often overlooks global user relationships and node-level semantic information, and still faces privacy breaches. Therefore, to address the above shortcomings, this study proposes an efficient privacy preserving federated graph neural network method (PPFedGNN) for social network analysis, thereby achieving dual guarantees of model performance and privacy security. To obtain global user relationships while protecting privacy, we designed a secure coding-based social subgraph aggregation method (SecureSA). This method improves model performance and algorithm efficiency by securely encoding and aggregating the node adjacency relationships across different clients. Additionally, to capture richer global node-level semantic information, we developed a secure social node augmentation method (SecureNA) based on local differential privacy mechanism (LDP). This method enhances model performance while maintaining security by adding noise perturbation to important weights and integrating overlapped node embeddings from different clients. Through experimental verification, it has been found that on Facebook, Blogcatalog, Flickr, and TeleComm datasets, the classification accuracy of PPFedGNN was 0.926, 0.838, 0.662, and 0.901, respectively, outperforming other algorithms. Through ablation experiments, the effectiveness of the global user relationships and node augmentation has been further demonstrated. In addition, we also conducted a theoretical analysis of the security techniques used throughout the training process to demonstrate their safety and efficiency.