Anomaly detection in cryptocurrency transactions with active learning

Abstract

Cryptocurrencies have gained tremendous popularity in recent years, with the rise of Bitcoin and other altcoins. However, this surge in popularity has also attracted fraudulent activities, such as scams, phishing, and money laundering. Particularly, machine learning (ML) algorithms have the potential to detect these fraudulent patterns. However, since in the fraud detection (FD) domain labels are scarce and most times very hard to get, traditional supervised ML models cannot be applied. Additionally, traditional unsupervised anomaly detection (AD) algorithms, generally, lead to high false positive rates. Therefore, this study is intended to explore the feasibility of using AD and active learning (AL) algorithms to uncover new fraudulent patterns in cryptocurrency transactions, assuming minimal access to labels.