Differentially private adaptive noise for graph neural network in online social networks

Online Social Networks (OSNs) have become a significant domain for studying social behaviors and information dissemination due to their unique user interactions and relational structures. These graph-structured data contain important information, and in-depth research on such data can effectively mine users’ social influence, thereby promoting the study of social applications such as information dissemination, behavior prediction, and social recommendation. Graph Neural Network (GNN) have demonstrated superior performance compared to traditional neural networks in multiple domains due to their advantages in handling graph-structured data. However, the inherent node-link structure in graph data makes privacy leakage an urgent issue to be addressed. In response to privacy protection issues in online social networks, w·e propose an adaptive differential privacy noise GNN scheme. This scheme can dynamically adjust the noise value introduced at each iteration based on historical and current model parameters, ensuring that the model meets differential privacy requirements while minimizing the impact on model accuracy. Through empirical experiments on multiple real-world datasets, this method maintains high accuracy under different privacy budgets. Even when the privacy budget $ϵ=2$, compared to the GAP algorithm and the baseline model based on Multilayer Perceptron (MLP), the model still achieved a 2.07 $\%$ increase in average accuracy, providing a superior trade-off between privacy and accuracy under a range of privacy protection requirements.