Application of boosting algorithms for anti money laundering in cryptocurrencies : towards healthier cryptocurrency networks

Abstract

Detecting money laundering is an essential function to protect the global economy, and it is vital to have systems and laws in place to counteract this nefarious activity. The recent emergence of cryptocurrencies has added another layer of complexity in the fight towards financial crime, while also creating an intriguing paradoxical paradigm: blockchain, the core component that underpins cryptocurrencies, functions without a central authority and offers pseudo-anonymity to its users, allowing criminals to disguise themselves amongst them, on the other hand, the openness of data, fuels the investigator’s toolkit to conduct forensic examinations. Meanwhile, the application of machine learning to combat and detect these crimes by leveraging data to build more robust compliance and Anti Money Laundering (AML) systems is advancing and exhibits great potential to safeguard the economy. This study focuses on the initial stages of money laundering, primarily, the detection of illicit activities (such as scams, financing terrorism, Ponzi-schemes) on cryptocurrency infrastructures, on both an ’account’ and ’transaction’ level. The common denominator between these crypto-related crimes is money laundering. Once an unlawful user gains access to these illicitly-gained funds, the primary focus shifts into "washing" them without detection. Utilising 4,681 Ethereum accounts and 46,564 Bitcoin transactions, we attempt to detect illicit activities using three state-of-the-art variants of gradient boosting, eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LGBM) and CatBoost. However, given the widespread issue of class imbalance in this domain and its dynamic environment that is created by the techniques employed by criminals which are continuously evolving to avoid detection, we seek to address these problems in order to mitigate the negative ramifications attributed to them. Employing Neighbourhood Cleaning Rule (NCL), Synthetic Minority OverSampling (SMOTE) and NCL-SMOTE as data-sampling techniques, and using our proposed innovative adaptation of XGBoost, ’Adaptive Stacked eXtreme Gradient Boosting (ASXGB)’, developed to handle non-stationary data, we successfully reduced the impact of concept drift, an issue which is often overlooked in this domain as well as, class imbalance. LGBM obtained the highest F1-Score of 0.820 on the ’transaction’ level data, whilst XGBoost obtained the highest F1-Score of 0.983 on the ’account’ level data, with further improvements made on the ’transaction’ level data when we used data-sampling techniques. We also showed that our ASXGB was one of the fastest model to adapt to concept drift when compared against other state-of-the-art adaptive learners on the ’transaction’ level data. Based on the obtained results, the proposed approaches are highly effective in the detection of illicit activities over cryptocurrency networks, at both an ’account’ and ’transaction’ level, obtaining fewer False Positives and False Negatives rates in comparison to previous work in this domain and industry standard (up to 90% False Positives).