Explainable Ethereum Fraud Detection for Secure Sustainable Digital Infrastructure

Abstract

Public blockchains such as Ethereum are frequent targets for fraud. Detecting malicious accounts from publicly available transaction records is therefore essential for maintaining secure and trustworthy blockchain-based digital infrastructure. We present an explainable and imbalance-aware machine learning pipeline for flagging fraudulent Ethereum accounts on a public dataset of 9,816 accounts. Along the way we uncover a label leakage in the widely shared version of this data. Two rowindex columns reveal the class through the order of the rows, and keeping them inflates every metric to a fake perfect 1.000. After removing them we compare no correction, class weights, and SMOTE across five model families. Gradient boosting is the clear winner, and LightGBM with SMOTE reaches a test F1 of 0.928 and a cross-validated F1 of 0.938. We explain the detector with SHAP, which shows that inflow value, counterparty count, and account lifetime are the strongest signals of fraud. The result is an honest and reproducible baseline that pairs strong accuracy with a readable account of its reasoning.