Dp-graph2vec: Learning Differential Privacy Representations of Graphs

Many machine learning tasks based on graph embedding achieved significant results in a wide range of fields, such as node clustering and link prediction. However, the original graph data may contain confidential information about individuals or organizations, which could be revealed through the publication of graphs or even via embeddings of graphs. While datasets contain highly private sensitive information, graph embedding methods which do not consider any privacy protection measures become impractical. Researchers recently have explored how differential privacy can be employed to a number of machine learning tasks, but differential privacy embedding methods for graphs have not been considered yet to the best of our knowledge. To address this problem, we propose a differential privacy embedding methods for graphs, named dp-graph2vec, based on gradient perturbation. In dp-graph2vec, we design a training mechanism called Average-Negative Sampling Update(ANSU) to solve the problem of privacy accounting. We proved that the mechanism satisfies the definition of differential privacy. Finally, we demonstrated results on various public datasets. Experiments and theory show that dp-graph2vec can strike a balance between privacy protection and data utility.