Robust graph learning for misbehavior detection

Abstract

Recent years have witnessed the thriving of online services like social media, e-commerce, and e-finance. Those services facilitate our daily lives while breeding malicious actors like fraudsters and spammers to promote misinformation, gain monetary rewards, or reap end users' privacy. Graph-based machine learning models have been playing a critical and irreplaceable role in modeling and detecting online misbehavior. With the observation that misbehaviors are different from massive regular behaviors, the graph models can leverage the relationship between data entities from a holistic view and reveal suspicious behaviors as anomalous nodes/edges/subgraphs on the graph. In this proposal, we investigate the graph-based misbehavior detection models from an adversarial perspective, considering the adversarial nature of malicious actors and real-world factors that impair graph models' robustness. We first introduce two published works enhancing the robustness of several graph-based misbehavior detectors using reinforcement learning. Then, we propose to explore: 1) the robustness of graph neural networks for misinformation detection on social media; and 2) the general robustness of graph neural networks towards unknown perturbations.