Anomaly Detection in the Bitcoin Network Using a Semi-Supervised LSTM Autoencoder

Abstract

The increasing use of cryptocurrencies, especially Bitcoin (BTC), has created new challenges for financial investigation. Although blockchain transactions are publicly accessible, the pseudo-anonymous nature of cryptocurrency networks can facilitate illicit financial activity. This work explores anomaly detection in the Bitcoin network using a semi-supervised Long Short-Term Memory Autoencoder (LSTM-AE). The focus is on the analysis of wallet activity over time in order to capture temporal behavioral patterns that may be related to illicit activities. Experiments are conducted on the Elliptic++ dataset. The model is trained exclusively on licit behaviour and the results indicate that the proposed formulation is able to retrieve a large proportion of illicit wallets despite the highly imbalanced setting.