Robust Graph Learning Against Camouflaged Malicious Actors

Venus Haghighi, Nasrin Shabani, Behnaz Soltani, Lina Yao, Quan Z. Sheng, Jian Yang, Amin Beheshti

Abstract: Graph Neural Networks (GNNs) have achieved great success in various graph-related applications such as fraud detection. However, GNN-based fraud detection models suffer from the camouflage behavior of malicious actors. Camouflaged fraudsters establish many normal connections to benign entities in the network to alleviate their suspiciousness, and eventually bypass the detection systems. To tackle this problem, we propose a new Multiple Adaptive Channels Aggregation Graph Neural Network for Detecting Camouflaged Fraudsters (named MAGNET for short). First, MAGNET includes a graph-agnostic edge labeling module to generate edge labels and domination signals (i.e., homophily-domination or heterophily-domination) for a given neighborhood. Second, MAGNET leverages multiple adaptive aggregation channels to improve graph learning against camouflaged fraudsters. Third, MAGNET adopts a multi-relational combination module to obtain final representations based on different relations for a multi-relational fraud graph. We conduct extensive experiments on two real-world fraud datasets, and our results show that MAGNET outperforms the state-of-the-art baselines. The source codes and datasets of our work are available at https://github.com/VenusHaghighi/MAGNET.