Federated Graph Learning (FGL) allows multiple participants to collaboratively train powerful Graph Neural Network (GNN) models in a distributed fashion without sharing raw data. It has shown advantages in various domains, including social network analysis and drug discovery. However, one of the core challenges in FGL is the nonIndependent and Identically Distributed (non-IID) nature of data. Particularly in graph-structured data, local clients may display heterogeneous feature distributions and skewed node degree distributions. To address this issue, we propose FedPT, a method within the FGL framework that incorporates the concept of prototype learning to aggregate feature representations across heterogeneous graph data through class prototypes. To capture core features of each class, each client integrates normalized and dimension-reduced features within each class, while the server performs global aggregation based on data weighting to harmonize class-specific core representations across clients. Subsequently, nodes are categorized based on their degrees, and hierarchical class vectors are computed and aggregated globally, allowing for finer-grained representation of graph structural differences. Finally, we design a contrastive learning approach to maximize the similarity between local and global prototypes of the same class and minimize the similarity between prototypes of different classes. We conduct extensive experiments on ten real-world datasets, demonstrating the superior performance of the FedPT method.

Graph Neural Networks (GNNs) have demonstrated remarkable capabilities in processing graph data, showcasing immense potential across various applications. However, concerns arise regarding privacy leakage when GNNs are applied to graphs containing sensitive data. Existing privacy protection methods for GNNs often introduce excessive noise, leading to prolonged computation time. In this study, we propose a GNN privacy-enhancing algorithm based on local differential privacy. Specifically, we transmit data processed by the encoder and rectifier to a graph convolutional layer named MSMA. This graph convolutional layer utilizes multi-hop aggregation of node features as a denoising mechanism to further mitigate the impact of injected noise. Subsequently, the denoised data is forwarded to the GNN layer for processing, and its eligibility for early termination after validation testing is assessed. Additionally, we devise an early termination strategy that significantly reduces runtime with minimal impact on accuracy, thereby conserving computational resources. Extensive experiments conducted on real-world datasets demonstrate that our approach effectively mitigates privacy loss while maintaining satisfactory accuracy and computational efficiency.