Anti-Money Laundering in Bitcoin using Deep Learning

A classification model called temporal-GCN is introduced, which utilizes transaction features to detect illicit transactions in Elliptic data, one of the largest Bitcoin transaction graphs. The model integrates long-short-term memory (LSTM) with graph convolutional networks (GCN). Although previous research has shown promising results using classical supervised learning and graph convolutional networks for anti-money laundering, few studies have incorporated temporal information from the dataset, with limited success. Furthermore, active learning techniques have rarely been applied to blockchain data. To address this gap, an active learning framework is applied to the Bitcoin transaction dataset, using Bayesian approximations such as Monte-Carlo dropout (MC-dropout) and Monte-Carlobased adversarial attacks (MC-AA) to estimate uncertainties. These methods are compared within active learning frameworks using various acquisition functions found in existing literature. The MC-AA approach has not been explored in the context of active learning before. Results show that the temporal-GCN model achieves considerable success compared to previous research using the same experimental settings and dataset. Additionally, the performance of the proposed acquisition functions with MC-AA and MC-dropout is evaluated and compared against the baseline random sampling model.