ASA: Adversary Situation Awareness via Heterogeneous Graph Convolutional Networks

Abstract

Given a large graph with millions of vertices and different types, how can we spot anomalies and find potential adversaries in time? Most graph-based fraud detection algorithms focus on finding dense blocks, discovering local subgraphs, designing belief propagation, and latent factor models. However, as fraudsters in online social networks gradually become adversarial, distributed, and invisible, it is easy for them to evade traditional detection methods. Even worse, existing detection systems are fragile to the new attacks. Once attackers update their tragedies, they will be inefficient.

Topics & Concepts