Plausible Heterogeneous Graph $k$-Anonymization for Social Networks

The inefficient utilization of ubiquitous graph data with combinatorial structures necessitates embedding methods, aiming at learning a continuous vector space for the which is amenable to be adopted in traditional machine algorithms favor representations. Graph methods build an important bridge between social network analysis and analytics as networks naturally generate unprecedented volume continuously. Publishing not only bring benefit public health, disaster response, commercial promotion, many other applications, but also give birth threats that jeopardize each individual's privacy security. Unfortunately, most existing works publishing focus on preserving structure less attention paid issues inherited from networks. To specific, attackers can infer presence sensitive relationship two individuals by training predictive model exposed embedding. In this paper, we propose novel link-privacy preserved framework using adversarial learning, reduce adversary's prediction accuracy links while persevering sufficient non-sensitive information such topology node attributes Extensive experiments are conducted evaluate proposed ground truth datasets.