Triplet-style Dynamic Graph Network with Transformer Encoder for Scam Detection in Cryptocurrency Transactions

Abstract

The surge in cryptocurrencies has been accompanied by a significant rise in scams, underscoring the critical need for precise scam detection. Cryptocurrency markets and transaction networks are dynamic, leading to evolving scam tactics and transaction topologies that challenge detection efforts. Compounding this, scammers cleverly mimic benign transactions, blurring the lines between legitimate and illicit activities. To address these intertwined challenges, we introduce the Triplet-style Dynamic Graph Convolutional Network (TD-GCN). TD-GCN is specifically engineered to model the dynamic nature of cryptocurrency transaction networks and to disentangle subtle distinctions between scam and benign transaction patterns. It leverages a Transformer Encoder for dynamic weight updates, effectively capturing network evolution, and employs Triplet Learning to disentangle representations of similar-appearing scam and benign transactions. Evaluations on a real-world Bitcoin transaction network demonstrate TD-GCN's superior scam detection performance. Notably, TD-GCN achieves a leading F1-macro score of 0.8855 and a 10.36%p precision increase over competing models, crucial for minimizing false positives. Ablation studies highlight the Triplet Dynamic Disentangle component's key role, reducing false positives by 81.1%. TD-GCN's performance gains stem from its dynamic updates for evolving networks and Triplet Learning for disentangling subtly different patterns. These features enable TD-GCN to significantly bolster cryptocurrency security by effectively detecting scams and minimizing false positives in dynamic transaction networks. © 2013 IEEE.