Achieving Adaptive Privacy-Preserving Graph Neural Networks Training in Cloud Environment

With the widespread adoption of Graph Neural Network (GNN) technology in industry, concerns regarding graph data privacy have become increasingly prominent. Differential privacy has been demonstrated as an effective method to ensure privacy in graph learning. However, existing differential privacy-based GNN methods often overlook the individual privacy protection needs of users, offering uniform privacy guarantees to all. This approach can result in either over-protection or insufficient protection for certain users. To address this issue, we propose an adaptive privacy-preserving GNN training method that accommodates the varying privacy requirements of nodes while achieving high model training accuracy. Specifically, APPGNN allocates adaptive privacy budgets based on individual user privacy needs. Additionally, to mitigate the impact of noise on data utility, APPGNN incorporates a weighted neighborhood aggregation mechanism to enhance GNN model accuracy. Theoretical analysis indicates that APPGNN provides adaptive privacy protection while ensuring ε-differential privacy on node data. Experimental evaluations on four real-world graph datasets validate the effectiveness of APPGNN.

I. Introduction

Graph data is widely used in data modeling and management in industrial scenarios. This data structure demonstrates its unique advantages in various industrial applications such as anomaly detection [1], supply chain management [2], and smart grids [3]. However, despite the significant potential of graph data in the industrial field, its complexity and inherent sparsity pose substantial challenges for graph analysis. With the development of deep learning, Graph Neural Networks (GNNs) have been widely applied in intelligent industries due to their unique advantages. GNNs can effectively handle graph-structured data, learning and inferring from node features and neighborhood information to provide more accurate analysis results than traditional methods. However, graph data often contains sensitive information from the participants, such as trade secrets in the production process and equipment operation data. If this information is leaked during graph analysis, it can lead to severe privacy risks. Therefore, it is essential to ensure data privacy protection when leveraging GNNs to advance industrial intelligent applications.