Experimenting Node Embeddings Techniques for Anti-Money Laundering in Cryptocurrencies

Abstract

Exploiting cryptocurrencies’ characteristics of money laundering has caused a huge loss to the global economy. Cybercriminals take advantage of the pseudonymity and less strict regula- tions associated with blockchain technology for hiding fraudulent funds and make them hard to be detected. Anti-Money Laundering (AML) regulations have a pivotal role in protecting financial institutions, businesses and individuals from fraud. Moreover, utilising Machine Learning (ML) algorithms for AML have emerged as powerful and trustworthy techniques to be used in cryptocurrency forensics. This project aims to first investigate the existing node embeddings algorithms specifically Node2vec and Graph Convolutional Network (GCN) as well as supervised classification mod- els namely Random Forest (RF), Logistic Regression (LR) and Extreme Gradient Boosting (XGBoost), for AML purposes in blockchain transactions. Based on a detailed review of the existing work, we propose and develop a novel GCN-label feature that exploits the adjacency nodes’ labels, in order to efficiently detect illicit transactions. Then, we validate the proposed models, evaluate models’ performance in terms of selected metrics and compare them to most closely related work. Our results show that extracting node embeddings from both Node2vec and GCN meth- ods to concatenate them with the original features boosts the performance of the supervised models. In particular, GCN embeddings have a noticeable impact on all RF and XGBoost models, in comparison with Node2vec. Furthermore, constructing the novel GCN-label fea- ture enhances the generated GCN embeddings significantly which leads to improving the supervised models’ performance. Together these results provide important insights into the leverage of node embeddings algorithms in cryptocurrency AML, and how the construction of the GCN-label feature enhances them.