Detecting Bitcoin CoinJoin Transactions Using Machine Learning

Abstract

Anonymity is one of the characteristics that makes Bitcoin mainstream. However, various approaches have been proposed to deanonymize Bitcoin to reveal the identities of the people behind the addresses. In order to enhance anonymity, Bitcoin mixing services were developed to obscure links between Bitcoin addresses. CoinJoin transactions method is one of the most used mixing approaches. The CoinJoin technique is a concept where multiple users merge their transactions into one larger transaction [16]. In this article, we propose a new system of detecting CoinJoin transactions using machine learning. Firstly, we developed a new algorithm to extract connected transactions and trace them back to the Coinbase transaction. Afterwards, we conducted a rigorous analysis of the data and performed an ablation study to identify the most relevant features of our machine learning models. Then, we implemented and fine-tuned them using an automated tool called OPTUNA, after which we trained the models and evaluated their performances, including accuracy, precision, and F1 score.