Privacy-preserving graph representation learning

Abstract

Among the various machine learning algorithms created to handle data with underlying graph structures are graph neural networks. There are several disciplines in which graph representation learning is used. Graph neural networks in particular, as a novel kind of link prediction method, can extract hidden link information from accessible network data in the field of link prediction. However, certain edges or connections between network nodes that are sensitive may be exposed by the learnt graph representation. In this dissertation, we investigate techniques for graph representation learning that safeguard connections’ privacy. We achieve privacy protection with link prediction in two ways. The first aspect is to view the privacy preservation problem as an optimization problem. Through optimization iterations we can achieve effective privacy preservation. The second aspect is to introduce a graph attack strategy, which attacks the target graph against the graph neural network algorithm in order to reduce the accuracy of the link prediction of the graph neural network, so that a certain degree of privacy protection can be realized.