Machine learning methods to detect money laundering in the Bitcoin blockchain in the presence of label scarcity (2005.14635v2)

Abstract: Every year, criminals launder billions of dollars acquired from serious felonies (e.g., terrorism, drug smuggling, or human trafficking) harming countless people and economies. Cryptocurrencies, in particular, have developed as a haven for money laundering activity. Machine Learning can be used to detect these illicit patterns. However, labels are so scarce that traditional supervised algorithms are inapplicable. Here, we address money laundering detection assuming minimal access to labels. First, we show that existing state-of-the-art solutions using unsupervised anomaly detection methods are inadequate to detect the illicit patterns in a real Bitcoin transaction dataset. Then, we show that our proposed active learning solution is capable of matching the performance of a fully supervised baseline by using just 5\% of the labels. This solution mimics a typical real-life situation in which a limited number of labels can be acquired through manual annotation by experts.

• The paper demonstrates that active learning with just 5% of labels achieves performance comparable to full supervision in detecting illicit Bitcoin transactions.
• It reveals that standard unsupervised anomaly detection methods are inadequate because illicit activities often mimic legitimate behavior.
• The study advocates for integrating active learning into AML systems to effectively counter fraud in environments with limited labeled data.

Machine Learning Approaches for Detecting Money Laundering in Bitcoin with Limited Labels

This paper explores machine learning methodologies to tackle the challenge of detecting money laundering activities within the Bitcoin blockchain, particularly under the constraints of label scarcity. Money laundering, a significant global issue, benefits greatly from the anonymity features offered by cryptocurrencies, making it imperative to develop robust detection systems. Traditional supervised learning methods require substantial amounts of labeled data, which is often unavailable in practical scenarios. Consequently, this paper investigates alternative approaches, namely unsupervised learning and active learning (AL), to recognize illicit transactions when only a limited number of labels are accessible.

Key Findings and Methods

The authors first examined the efficacy of unsupervised anomaly detection methods. Specifically, they evaluated seven common algorithms including Local Outlier Factor, K-Nearest Neighbours, and Isolation Forest, among others. Contrary to previous studies suggesting the effectiveness of anomaly detection for anti-money laundering (AML), these methods performed poorly in identifying illicit Bitcoin transactions. The paper attributes this to the non-outlying nature of illicit transactions in real-world data, highlighting that sophisticated criminals tailor their activities to mimic licit behavior, thereby evading typical anomaly detection.

Given the limitations of unsupervised approaches, the paper explores the potential of active learning. AL involves iterative querying of the most informative instances for labeling, thereby minimizing the dependency on large labeled datasets. The researchers tested various AL query strategies, including uncertainty sampling and expected model change, combined with classifiers like Random Forest, XGBoost, and Logistic Regression. Remarkably, their results reveal that using only 5% of the available labels in the Bitcoin dataset suffices to achieve performance equivalent to a fully supervised model. This finding is significant as it illustrates the practical feasibility of implementing machine learning-based AML systems with constrained labeled data.

Implications and Future Prospects

The paper's outcomes have several implications for both theoretical research and practical implementation in financial crime detection. The inadequacy of unsupervised methods in detecting non-outlying illicit transactions underscores the necessity for models that can handle complex adversarial behaviors. Additionally, the success of AL in achieving competitive performance with minimal labels opens avenues for its integration into real-world AML processes within various financial systems, including bank transfers and loans.

Future research should explore the applicability of these findings across different financial datasets and money laundering schemes with varying levels of sophistication. Additionally, developing models that can dynamically adapt to evolving fraudulent tactics remains a critical challenge. In summary, this paper contributes valuable insights into combating financial crimes within decentralized systems like Bitcoin, advocating for innovative solutions that require minimal manual intervention while maximizing efficiency in detecting illicit activities.