Supervised Machine Learning Framework for Deanonymization of Transactional Entities in Decentralized Infrastructure Environment

Abstract

A public ledger is used by Bitcoin, a digital currency, to keep track of transactions. The owner of the Bitcoin keeps their identity secret and is identified only by their unique address. This indicates that because Bitcoin offers anonymity, it may be utilized for illicit purposes on a regular basis. This study presents a supervised machine learning approach for predicting anonymous user activities on the Bitcoin Blockchain. As a training dataset to facilitate the user activities classification, we created a labelled dataset with over 4 million samples from exchanges, gambling, pools, and services whose identities and types were disclosed. The primary goal is to classify transactions on the blockchain in order to deanonymize them and distinguish between legitimate and illegitimate ones. On the class imbalanced dataset, we obtained impressive cross-validation (CV) accuracy using the Gradient Boosting, Random Forest, and eXtreme Gradient Boosting with default parameters and hyperparameters. Using Random Forest helped achieve the best cross-validation accuracy on default parameters and hyperparameters obtained using grid search on the class-balanced dataset using the Synthetic Minority Oversampling Technique, while Bagging and eXtreme Gradient Boosting were used on hyperparameters obtained using randomized search. Empirical results show that the recommended model is up to 98% accurate.