Community hiding using a graph autoencoder

摘要

Community detection can reveal real social relations and enable great economic benefits for enterprises and organizations; however, it can also cause privacy problems such as the disclosure of individual or group information amongst community members, which goes against the hidden wishes of individuals and groups. Therefore, community hiding has received increasingly more attention in recent years. However, the network generation mechanism has not been considered in previous studies on community hiding. Generation models can reflect the generation process of the network and show the strength of the connection between nodes. To this end, we propose a new graph autoencoder for the community hiding algorithm, namely, GCH, which not only hides the community structure but also embodies the generation mechanism of the network. It uses the rules of the generation process from underfitting to overfitting in the community network to select the connections that have the greatest impact on the community structure for rewiring. After analyzing the essence of community detection algorithms and graph neural networks, an improved graph autoencoder is used to reconstruct the probabilistic adjacency matrix; and under the constraint of an ”invisible perturbation” of the network structure, the existing mainstream community detection algorithm is attacked, which greatly reduces the accuracy of community detection results. For the verification of model effectiveness, two widely used indicators NMI and AE are used to compare the performance of our attack on the community detection algorithm with other baselines under different dimension settings. Compared with several baseline algorithms, extensive experimental results are obtained.