Detection of CoinJoin Transactions on the Bitcoin Blockchain: A Machine Learning Approach with Imbalanced Data

Abstract

Blockchain technology enables semi-anonymous transactions, where user identities are not directly revealed but instead linked to cryptographic wallet addresses. While this design enhances privacy and security, the intrinsic transparency of public blockchains raises concerns about the true anonymity of users. To address this, privacy-enhancing techniques such as CoinJoin were developed to obscure transaction flows. CoinJoin is a Bitcoin-based mixing technique that combines multiple inputs and outputs into a single larger transaction, making it difficult to trace the original senders and recipients. Although CoinJoin was intended to support privacy-preserving transactions, it is often exploited for money laundering. As a result, there is a growing need to identify CoinJoin transactions in order to prevent such misuse. This research aims to detect unidentified CoinJoin transactions on the Bitcoin blockchain using a machine learning approach. Unlike prior work that relied on heuristics or traditional machine learning algorithms, we propose an improved methodology relying on random forest ($\mathbf{R F}$) models designed to handle imbalanced datasets. Specifically, we introduce a novel variant, the biased random forest (BRAF), aimed at improving detection performance under significant class imbalance.