Detecting Fraudulent Wallets in Ethereum Blockchain Combining Supervised and Unsupervised Techniques - Using Autoencoders and XGboost.

The illicit activity in Blockchain reached an all-time high in 2021. In this work, we combined two machine learning techniques, Autoencoder (AE) and Extreme Gradient Boosting (XGBoost), to improve the performance of predicting illicit activity at the account level. The choice of autoencoding technique allows us to be able to detect new MOs (modus operandi) from fraudsters. With an Autoencoder trained only with healthy accounts, we are not misleading the model to focus on specific MOs as it happens with tree-based models. This allows us to introduce a dimension that could capture future fraudulent behaviours. Furthermore, the dataset was generated considering the real applicability of the model, i.e. it mimics what can realistically be obtained in a practical situation. With this approach, we were able to improve the state-of-the-art performance. In our test set the precision-recall AUC (area under the curve) of our final model increased by 12,07% when compared with our baseline.