APPLICATION OF GRAPH ATTENTION NETWORKS (GAT) WITH SELF-ATTENTION MECHANISM FOR ANOMALY DETECTION IN BLOCKCHAIN TRANSACTION NETWORKS

The rapid adoption of blockchain technology and cryptocurrencies has introduced new paradigms in decentralized finance. However, the pseudonymous nature of blockchain networks has also attracted illicit activities, including money laundering, phishing, and ransomware payments. Traditional machine learning approaches often struggle to detect these anomalies because they analyze transactions as isolated tabular data, ignoring the complex, non-Euclidean topological structures inherent in transaction networks. This study applies Graph Attention Networks (GAT) integrated with a self-attention mechanism to detect anomalous transactions within a blockchain network and compares it against baseline models under an expedited training regime. Evaluated using the standard Elliptic Data Set representing Bitcoin transactions, the empirical experiments demonstrate a significant trade-off between traditional algorithms and Graph Neural Networks (GNN). The Random Forest (RF) model achieved the highest F1-Score of 82.2% driven by a high precision of 95.9%. Conversely, the GAT models, while struggling with precision under limited epoch training, demonstrated a persistent strength in identifying hidden anomalies, reaching a recall of up to 83.2% (16-heads). The findings suggest that while ensemble methods like RF are highly effective for immediate precision-based classification, dynamically weighing neighborhood topologies using GAT reveals hidden illicit patterns (high recall), providing a foundational insight for developing hybrid detection frameworks in decentralized financial systems.