Bidirectional fusion heterogeneous graph networks for semi-supervised Bitcoin transaction anomaly detection in dynamic transaction graphs.

Abstract

Detecting anomalies in the Bitcoin transaction network is critical for ensuring blockchain security and stability. The network's heterogeneous structure and dynamic nature, coupled with scarce labeled anomalies, pose significant challenges for traditional graph-based methods. To address these, we propose Bidirectional Fusion Heterogeneous Graph Network (BF-HGN), a semi- dynamic supervised model for Bitcoin transaction anomaly detection task. BF-HGN designs multi-type feature embedding and alignment strategies to effectively unify features across heterogeneous transaction-address nodes. A bidirectional temporal fusion mechanism is proposed to capture long-range temporal dependencies that unidirectional models often miss. To alleviate class imbalance and limited annotations, a Class-balanced Classifier (CBC) combined with Adjacency Adaptation (AA) and Adaptive Feature Space Regulation (AFSR) losses is proposed to generate pseudo-anomalous nodes closely resembling real anomalies, improving discrimination boundaries. Experiments on the Elliptic++ dataset demonstrate that BF-HGN outperforms existing methods, achieving F1 scores of 0.6301 and 0.5784 for transaction and address nodes, respectively, establishing a new benchmark for Bitcoin transaction anomaly detection.