A Graph Data Privacy-Preserving Method Based on Generative Adversarial Networks

Abstract

We proposed a graph anonymization method which is based on a feature learning model of Generative Adversarial Network (GAN). We used the differential privacy to ensure the privacy and take both anonymity and utility into consideration. The method consists of the following two parts: Firstly, we designed a graph feature learning method based on GAN. The method used the bias random walk strategy to sample the node sequence from graph data, and trained the GAN model. After training, the GAN generated a set of simulation sequences that are highly like the real sampled sequence. Secondly, we proposed an anonymous graph construction method based on the simulation node sequence. We calculated the number of edges in the node sequences and constructed a probability adjacency matrix. The differential privacy noise is added to get the anonymous probability adjacency matrix. Then we extract the edges from the anonymous matrix and then constructed the anonymous graph. We evaluate our methodology, showing that the model had good feature learning ability through embedding visualization and link prediction experiments, compared with other anonymous graphs. Through experiments such as metric evaluation, community detection, and de-anonymization attack, we proved that the anonymous method we proposed is better than the current mainstream anonymous method.