DPNE: Differentially Private Network Embedding

Learning the low-dimensional representations of the vertices in a network can help users understand the network structure and perform other data mining tasks efficiently. Various network embedding approaches such as DeepWalk and LINE have been developed recently. However, how to protect the individual privacy in network embedding has not been exploited. It is challenging to achieve high utility as the sensitivity of stochastic gradients in random walks and that of edge sampling are very high, thus incurring high utility loss when applying Laplace mechanism and exponential mechanism to achieve differential privacy. In this paper, we develop a differentially private network embedding method (DPNE). In this method, we leverage the recent theoretical findings that network embedding methods such as DeepWalk and LINE are equivalent to factorization of some matrices derived from the adjacency matrix of the original network and apply objective perturbation on the objective function of matrix factorization. We evaluate the learned representations by our DPNE from three different real world datasets on two data mining tasks: vertex classification and link prediction. Experiment results show the effectiveness of DPNE. To our best knowledge, this is the first work on how to preserve differential privacy in network embedding.