BitSentry: A Graph-Based Approach for Tracking Illicit Transactions in Bitcoin

Abstract

Detecting illicit transactions in financial networks is critical for combating money laundering, particularly in Bitcoin, where obfuscation techniques enable anonymity. While Graph Neural Networks (GNNs) have shown promise in this domain, existing approaches like FraudLens [1] use graph restructuring (e.g., Personalized Page Rank (PPR)-based augmentation and feature-driven edge filtering) to address imbalances. In this work, we propose an enhanced framework that incorporates domain-aware centrality measures-betweenness and eigenvector centrality-to explicitly identify critical transaction pathways (e.g., mixers as high-betweenness nodes, fraud rings as high-eigenvector hubs). By prioritizing these structures, our method enhances transactional relationship representation while filtering noise. Experimental results demonstrate that our centrality-driven approach achieves higher precision for mixer detection compared to the state-of-the-art, particularly in later temporal windows where illicit activity evolves. Our method also maintains robustness to structural perturbations, though at a higher computational cost, making it suitable for realworld deployment in anti-money laundering systems. This work improves fraud detection by combining key graph metrics with GNNs, offering an interpretable, topology-driven alternative to feature-heavy approaches. © 2025 IEEE.