Detection of illicit accounts over the Ethereum blockchain

Abstract

Take-down policy If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim.

Application of Gradient Boosting Algorithms for Anti-money Laundering in Cryptocurrencies

“…To circumvent the lack of unlabelled data indicating financial cryptocurrency crimes, various unsupervised approaches, including k-means and kd-trees, were investigated to label accounts [2,33]. Others relied on online sources (flagged by communities) [15,46] or by employing heuristic-based reasoning processes (e.g., reuse of the same addresses) to categorise instances [45,49]. Some researchers focus on detecting a specific kind of criminal activity such as Bitcoin accounts linked to High Yielding Investment Programs (HYIP) [46] or Ponzi Schemes [3], while others link cryptocurrency accounts or transactions to multiple labels (multiclass classification), e.g., mixer services, dark marketplaces, exchanges, wallet providers, scams, gambling, terrorists, and ransomware [19,29,30,45,47,58].…”

“…Some researchers focus on detecting a specific kind of criminal activity such as Bitcoin accounts linked to High Yielding Investment Programs (HYIP) [46] or Ponzi Schemes [3], while others link cryptocurrency accounts or transactions to multiple labels (multiclass classification), e.g., mixer services, dark marketplaces, exchanges, wallet providers, scams, gambling, terrorists, and ransomware [19,29,30,45,47,58]. Others focus on a binary classification, grouping illegal/illicit vs legal/licit activities [15,26,33,49]. Classifying accounts to their service is important to combat money laundering, as Fanusie and Robinson [14] pointed out that this activity stems from various areas in the network, with the highest exploited services being: mixing services, dark marketplaces, and gambling sites.…”

“…Most research focuses on supervised learning [3, 15, 19, 29, 45-47, 49, 58] or utilises unsupervised learning to label the data and then applies a supervised learner to predict the label [2,33]. Moreover, the classification usually falls within two paradigms: account [3,15,19,33,46,47,58] or transaction level [26,49] [3,19,26,30,45,49,58]. However, the most common models are DT ensembles, more particularly RF [2, 3, 19, 26, 45-47, 47, 49, 58] and gradient boosting (e.g., XGBoost and LGBM) [15,19,30,45,46,58].…”

“…Moreover, the classification usually falls within two paradigms: account [3,15,19,33,46,47,58] or transaction level [26,49] [3,19,26,30,45,49,58]. However, the most common models are DT ensembles, more particularly RF [2, 3, 19, 26, 45-47, 47, 49, 58] and gradient boosting (e.g., XGBoost and LGBM) [15,19,30,45,46,58]. In some cases, RF outperformed gradient boosting [46], or even more complex deep learning models, i.e., Graph Convolutional Network (GCN) [49], and in others gradient boosting outperformed RF [30,45,58].…”

“…First, although it is evident from the literature that tree-based ensembles are one of the most widely used models in this domain, in particular, RF [2,3,26,30,46,49,58] and tree-based gradient boosting [15,19,30,45,46,58], it is still unclear which ensemble is most effective in detecting illicit activities on cryptocurrency networks. Previous research primarily focuses on detecting illicit activities on one of the following approaches: at an account level [15,30,45], or at a transaction level [26,49]. When these ensembles were compared head-to-head against each other [30,45,46] at an accountlevel, some reported that RF outperforms tree-based gradient boosting [46], while others suggested otherwise [30,45].…”