Ethereum fraud detection using Smote-Tomek and stacked ensemble learning technique

Abstract

This research addresses the critical challenge of detecting fraudulent Ethereum transactions, which remains notoriously difficult due to the overwhelming prevalence of legitimate activities compared to illicit ones. This pronounced class disparity frequently results in skewed outcomes when employing conventional detection frameworks. Our investigation implements a composite resampling strategy, SMOTE-Tomek, which concurrently augments the under-represented fraud category whilst eliminating ambiguous instances from the dominant category. Subsequently, the equalised dataset undergoes processing via a stacked ensemble framework comprising three robust gradient boosting methodologies: XGBoost, LightGBM, and CatBoost, synthesised by a logistic regression meta-classifier.  Through stratified 5-fold cross-validation conducted against established benchmarks, the proposed ensemble attained a mean accuracy of 0.9879, whilst precision, recall, and F1-score each achieved an exceptional score of 0.98. Benchmark comparisons verified that our stacked methodology consistently surpasses individual base algorithms. The study demonstrated that combining targeted resampling with ensemble stacking offers highly effective solutions for Ethereum fraud detection, though real-time implementation and cross-blockchain generalisability require further investigation. The proposed model provides cryptocurrency exchanges, investors, and regulatory bodies with a robust mechanism for identifying fraudulent activities, thereby enhancing ecosystem security and maintaining user confidence. This work represents the first comprehensive integration of SMOTE-Tomek resampling with stacked gradient boosting ensembles for Ethereum fraud detection, validated through rigorous k-fold cross-validation rather than simple train-test splits.

Received 29 December 2025

Accepted 02 February 2026

Published 26 April 2026