Adversarial Link Prediction in Social Networks

Link prediction is one of the fundamental tools in social network analysis, used to identify relationships that are not otherwise observed. Commonly, link prediction is performed by means of a similarity metric, with the idea that a pair of similar nodes are likely to be connected. However, traditional link prediction based on similarity metrics assumes that available network data is accurate. We study the problem of adversarial link prediction, where an adversary aims to hide a target link by removing a limited subset of edges from the observed subgraph. We show that optimal attacks on local similarity metrics---that is, metrics which use only the information about the node pair and their network neighbors---can be found in linear time. In contrast, attacking Katz and ACT metrics which use global information about network topology is NP-Hard. We present an approximation algorithm for optimal attacks on Katz similarity, and a principled heuristic for ACT attacks. Extensive experiments demonstrate the efficacy of our methods.