Oversampling Techniques for Detecting Bitcoin Illegal Transactions

Abstract

Bitcoin users are guaranteed to be anonymous, increasing the number of cryptocurrency trading related to crimes and fraudulent activities. While most studies about detecting illegal transactions try to distinguish trading patterns and classify them from legitimate ones, classification performance is poor since the class distributions of transaction data are highly imbalanced. In general, the Synthetic Minority Over-sampling TEchnique (SMOTE) is used to deal with class-imbalanced data, but SMOTE has a problem that it does not fully represent the diversity of the data. In this paper, we introduce another oversampling technique using Generative Adversarial Networks (GAN) to generate artificial training data for classification model. In order to verify similarity between artificial data and the actual one, oversampled dataset is evaluated with a classification model using XGBoost algorithm. We show classification performance is improved on average with synthetic data generated by both SMOTE and well-designed GAN model.