Graph embedding has been demonstrated to be a powerful tool for learning latent representations for nodes in a graph. However, despite its superior performance in various graph-based machine learning tasks, learning over graphs can raise significant privacy concerns when graph data involves sensitive information. To address this, in this paper, we investigate the problem of developing graph embedding algorithms that satisfy local differential privacy (LDP). We propose LDP-GE, a novel privacy-preserving graph embedding framework, to protect the privacy of node data. Specifically, we propose an LDP mechanism to obfuscate node data and adopt personalized PageRank as the proximity measure to learn node representations. Then, we theoretically analyze the privacy guarantees and utility of the LDP-GE framework. Extensive experiments conducted over several real-world graph datasets demonstrate that LDP-GE achieves favorable privacy-utility trade-offs and significantly outperforms existing approaches in both node classification and link prediction tasks.

翻译：图嵌入已被证明是一种学习图中节点潜在表示的有效工具。然而，尽管它在各种基于图的机器学习任务中表现出卓越性能，但当图数据涉及敏感信息时，在图上的学习过程可能引发严重的隐私问题。为解决这一问题，本文研究了满足局部差分隐私（LDP）的图嵌入算法开发问题。我们提出了一种新型隐私保护图嵌入框架LDP-GE，用于保护节点数据的隐私。具体来说，我们设计了一种LDP机制来混淆节点数据，并采用个性化PageRank作为邻近度量来学习节点表示。随后，我们从理论上分析了LDP-GE框架的隐私保证与效用。在多个真实图数据集上进行的广泛实验表明，LDP-GE在隐私-效用权衡方面表现优异，并在节点分类和链接预测任务中显著优于现有方法。