DR-SGFormer: Dynamic Graph Sparsification with Swin Windows and Hierarchical Fusion for Cryptocurrency Fraud Detection

Graph neural networks (GNNs) show significant potential in anti-money laundering and fraud detection, but Cryptocurrency transaction networks face technical challenges. Existing GNN-based methods are prone to node representation distortion under adversarial attacks, topological pollution from fraudulent nodes in mixing services, and the inability to correlate local high-frequency patterns with global fund-flow closure structures. To address these, we propose DR-SGFormer, which includes the Graph-Adaptive Swin Window Mechanism for robust feature aggregation via high-degree nodes and noise-robust attention, the Dynamic Risk-Aware Sampling Strategy for adaptive edge pruning based on feature similarity, and the Hierarchical Collaborative Feature Fusion for multi-scale integration. Experimental results on the Elliptic Bitcoin transaction dataset, Elliptic[Formula: see text] (Actors dataset), and Ethereum dataset demonstrate that DR-SGFormer achieved the highest average ranking among the comparison methods. Although the limited number of datasets and statistical adjustments for multiple comparisons restrict definitive assertions that our model outperforms strong baselines (e.g., Graphformer, XGBoost), the consistent improvements in F1-scores, large effect sizes, and moderate-to-strong Bayesian evidence underscore its competitive performance in cryptocurrency fraud detection.