Adaptive graph neural network protection algorithm based on differential privacy

Abstract

Graph Neural Networks (GNNs) have gained widespread adoption across various fields due to their superior capability in processing graph-structured data. Nevertheless, these models are susceptible to unintentionally disclosing sensitive user information. Current differential privacy algorithms for graph neural networks exhibit constrained adaptability and prolonged runtimes. To address these issues, this paper introduces an adaptive GNN protection algorithm grounded in differential privacy. The algorithm offers robust privacy safeguards at both node and edge levels, employing a bespoke normalization approach based on mean and variance to effectively manage data non-uniformity and outliers, thereby enhancing the model’s adaptability to diverse data distributions. Furthermore, the implementation of an early stopping strategy markedly decreases runtime while exerting negligible influence on accuracy, thus enhancing computational efficiency. Experimental results indicate that this approach not only improves the model’s predictive accuracy but also significantly reduces its computational time.