FRAUD DETECTION USING GRAPH NEURAL NETWORKS

Abstract

The primary objective of fraud detection is to distinguish between fraudulent behavior and legitimate user activity. In the context of graph/network setups, fraudsters and ordinary users are represented as nodes, while the connections between these nodes are displayed as edges. Fraudsters typically employ tactics to appear inconspicuous by engaging in seemingly ordinary activities, such as deliberately initiating a high volume of interactions with regular users. Their appearance inherently contrasts with what is considered normal due to their camouflage, resulting in abnormalities in the graph. Collaborative fraud, which has become increasingly severe in social networks and telecommunications, poses a challenge for traditional fraud detection methods. The proposed model provides empirical evidence of a substantial and favorable correlation between the rise in collaborative fraud and the decline in the efficacy of conventional detection techniques. This suggests that individuals who are difficult to apprehend using conventional approaches often participate in joint fraudulent activities. To investigate novel research directions for collaborative fraud detection, we propose a novel fraud detection model that utilizes higher-order cooperative relationship mining. Various experimental results manifest the effectiveness of our proposed model.