GraphFA: Graph Enhanced Fraud Detectors with Camouflage Detection for Financial Anti-Fraud

Financial anti-fraud has attracted increasing attention from the industry, leading to the constant development of related methods. Existing methods mainly treat customer information as independent, ignoring that fraudulent behaviors often have similar patterns and occur in the form of organized groups. To address this challenge, we propose a novel graph enhanced fraud detector with camouflage detection for financial anti-fraud (GraphFA). Firstly, we model customer information as a graph capturing relationships between customers and reformulate the financial fraud detection problem as a node classification problem on graph. Secondly, to address the issue of fraudulent edge camouflage on large-scale graphs, which hinders the performance of graph-based fraud detectors, GraphFA introduces a neural camouflage detection sampler to identify camouflaged edges during subgraph sampling. To the best of our knowledge, GraphFA is the first financial anti-fraud method with graph construction and camouflage detection. Our experiments on a genuine loan fraud detection dataset prove that GraphFA outperforms both state-of-the-art methods and commonly used gradient boosting decision tree methods in the financial domain, especially reaching 90.5% AVC and +4.7% increase in Recall on individual fraud dataset. On group fraud dataset, GraphFA achieved +7.5% increase in F1-score and +5.9% increase in Recall compared to the best baseline.