CRAFIC Framework: Multi-Account Collaborative Fraud Detection, Efficient Feature Extraction and Relationship Modelling Combined with CNN-LSTM and Graph Attention Network

Abstract

This study proposed a complex fraud detection framework called CRAFIC (complex relationship analysis for fraud identification and cost management), which combines deep learning and graph neural networks to address complex fraud behaviours such as multi account collaborative fraud. The study used the DataCo Global supply chain dataset and the IEEE-CIS fraud detection dataset to extract order features using convolutional neural networks and long short term memory networks, and analysed the relationships between orders using graph attention networks to reveal complex fraud patterns. The results show that the CRAFIC framework performs well in both single order and collaborative fraud detection tasks. In single order fraud detection, the accuracy of the CRAFIC framework increased from the initial 45.21% to 93.75%, and the loss value decreased from 1.19 to 0.14, significantly better than other models. In collaborative fraud detection, the accuracy of the CRAFIC framework reached 90.3%, once again surpassing other models. These results validate the advantages of the CRAFIC framework in multimodal data fusion and complex relationship modelling. The CRAFIC framework reveals complex fraud patterns, optimizes internal controls and audit processes, enhances data security measures, prevents system vulnerabilities from being exploited, and enhances market reputation and customer trust.