MGCNLLaMA: Multi-graph convolutional networks with lightweight LLaMA-style transformer encoder integration for dynamic fraud detection in cryptocurrency networks

Financial fraud detection in blockchain networks presents unique challenges due to the dynamic, heterogeneous, and large-scale nature of blockchain transactions. Traditional graph neural networks often struggle to capture both local structural patterns and global temporal dependencies inherent in financial networks, while also failing to account for incomplete or missing information. This paper introduces MGCNLLaMA (A Multi-Graph Convolutional Network with Lightweight Large Language Model Architecture (LLaMA)-style Transformer Encoder). This hybrid architecture synergistically combines Graph Convolutional Networks (GCNs) with lightweight LLaMA-style transformer encoder models for dynamic fraud detection in temporal cryptocurrency transaction networks. Our approach incorporates a random walk-based feature extraction mechanism to capture multi-hop neighborhood relationships and employs temporal slicing to model the evolving nature of transaction patterns. We also proposed a module for predicting missing information to address incomplete node information representations through learnable gating mechanisms. Experimental evaluation on three blockchain transactional datasets from both Bitcoin and Ethereum demonstrates that MGCNLLaMA achieves superior performance with an AUC score improvement of at least 6.9%, a 9.9% increase in Precision, 19.8% increment in recall, and 15% improvement in F1-score, significantly outperforming existing transformer graph-based fraud detection methods. With thorough adversarial testing, we confirmed that our model is robust towards attacks. The architecture of the MGCNLLaMA model successfully identifies fraudulent transaction patterns by leveraging both local graph topology and global sequential dependencies, making it particularly effective for real-world cryptocurrency fraud detection scenarios.