MGGPT: A Multi-Graph GPT-enhanced framework for dynamic fraud detection in cryptocurrency networks

Abstract

The rapid increase in cryptocurrency transactions has increased demand for advanced fraud detection systems. Conventional methods are often rigid and do not effectively capture cryptocurrency networks’ intricate temporal and structural patterns, while existing dynamic approaches struggle with incomplete or missing information. To tackle this issue, we present MGGPT, a new hybrid framework that integrates Graph Attention Neural Networks (GAT) with GPT-based transformers to improve fraud detection within cryptocurrency transaction networks. Our approach utilizes temporal graph structures through reachability networks (reach-nets) to derive essential node features, while also directly integrating edge labels into the embedding vectors, and introduces an innovative mechanism for predicting missing information to address the challenges posed by incomplete data in blockchain networks. The model features a dual-perspective learning strategy, employing local graph structures via GAT Networks and global contextual patterns through GPT-based sequence modeling to capture both structural and temporal dynamics in transaction networks. Our MGGPT framework implements a sophisticated edge classification mechanism using Support Vector Machines (SVM) for the final prediction. Experimental findings on actual cryptocurrency transaction datasets indicate superior efficacy in identifying fraudulent patterns, achieving notable improvements of 8.5% AUC, a 10.2% increase in Precision, 29.5% increment in recall, and 20.5% improvement in F1-score. Compared to baseline models such as STA-GT and CTGN, the proposed MGGPT improves the representation of dynamic relationships and faster convergence. Overall, the analysis reveals that our framework is not only more accurate but also more robust and scalable for real-world temporal graph applications. Ultimately, we assessed the robustness of our framework against adversarial attacks to show its practical applications in blockchains.