Disentangled Prototypical Graph Convolutional Network for Phishing Scam Detection in Cryptocurrency Transactions

Abstract

Blockchain technology has generated an influx of transaction data and complex interactions, posing significant challenges for traditional machine learning methods, which struggle to capture high-dimensional patterns in transaction networks. In this paper, we present the disentangled prototypical graph convolutional network (DP-GCN), an innovative approach to account classification in Ethereum transaction records. Our method employs a unique disentanglement mechanism that isolates relevant features, enhancing pattern recognition within the network. Additionally, we apply prototyping to disentangled representations, to classify scam nodes robustly, despite extreme class imbalances. We further employ a joint learning strategy, combining triplet loss and prototypical loss with a gamma coefficient, achieving an effective balance between the two. Experiments on real Ethereum data showcase the success of our approach, as the DP-GCN attained an F1 score improvement of 32.54%p over the previous best-performing GCN model and an area under the ROC curve (AUC) improvement of 4.28%p by incorporating our novel disentangled prototyping concept. Our research highlights the importance of advanced techniques in detecting malicious activities within large-scale real-world cryptocurrency transactions.