An Advanced Blockchain-Based Cryptocurrency Pattern Detection Using Machine Learning: An Enhanced CNN and RF Approach with Mixers, Tumblers, and Privacy Coins

Abstract

Cryptocurrency-enabled money laundering has emerged as a critical threat to global financial integrity, with advanced obfuscation techniques such as mixing services, tumblers, CoinJoin protocols, and privacy-oriented coins making illicit transaction detection increasingly complex. This paper presents an Enhanced Random Forest (ERF) framework for detecting cryptocurrency laundering patterns within the Bitcoin transaction network. Operating on the Elliptic dataset comprising 203,769 labeled transactions with 167 transactional and aggregate features, the proposed methodology integrates graph-based feature engineering (in-degree, out-degree, total-degree centrality computed via NetworkX), adaptive Synthetic Minority Over-sampling Technique (SMOTE), Recursive Feature Elimination (RFE) for dimensionality reduction, and systematically optimized hyperparameters through 5-fold cross-validated GridSearchCV. The resulting model achieves an F1-score of 0.973, precision of 0.975, recall of 0.971, and ROC-AUC of 0.992 on a strictly temporal holdout split, surpassing standard Random Forest (+6.1%), XGBoost (+3.9%), and Logistic Regression (+13.9%) baselines. Crucially, graph degree features rank among the top-six most discriminative features by Gini importance, confirming that transaction-network topology provides powerful laundering signals beyond raw transactional attributes. The complete pipeline executes in under 9 minutes on free CPU infrastructure, demonstrating production viability. These results advance the state-of-the-art in explainable, computationally efficient Anti-Money Laundering (AML) detection applicable to mixer, tumbler, and privacy-coin obfuscation schemes.