| Abstract [eng] |
Cryptocurrencies are becoming an important part of the economy. As much as a quarter of Bitcoin wallets are involved in illegal activities, almost every second transfer on the network is made by malicious actors. The market is growing in need of regulation of cryptocurrency cash flows, and investigators lack tools capable of performing a detailed analysis of crypto wallets. This paper investigates the possibilities of machine learning for analyzing bitcoin network transactions. The main objectives of this paper are to conduct a review of relevant literature, prepare a dataset, train models to select the best one, and use it to create a tool capable of detecting illegal bitcoin transactions. The created dataset is suitable for training models that would be able to perform transaction analysis in real time using real, unseen bitcoin network data. Three classification algorithms were used (random forest, ADA Boost, XG Boost). During the classification of real transactions, SHAP analysis of the observation is performed, which can explain which transaction features determined the model's decision. The best results were obtained using the XG Boost algorithm, the most important feature being the transaction fee. Using this model, an intuitive tool with a user interface has been created for investigators of financial crimes on the bitcoin network. |
