Hybrid Feature Engineering and Explainable Ensemble Learning for Ethereum Smart Contract Vulnerability Detection

Abstract

Ethereum smart contracts have become a fundamental component of decentralized blockchain applications. However, vulnerabilities in Solidity smart contracts can lead to severe financial losses and security risks. Traditional vulnerability detection approaches often rely on opcode analysis or deep learning architectures with limited interpretability and high computational complexity. This study proposes a hybrid feature engineering and explainable ensemble learning framework for Ethereum smart contract vulnerability detection. The proposed framework integrates lexical TF-IDF features, structural software metrics, security-oriented execution patterns, and metadata attributes extracted from Solidity source code. A dataset containing 691 Ethereum smart contracts was utilized, consisting of 358 non-vulnerable and 333 vulnerable contracts generated through rule-assisted vulnerability annotation. Ensemble learning models including Random Forest, XGBoost, and LightGBM were evaluated to classify vulnerable and non-vulnerable smart contracts. Experimental results demonstrate that the proposed framework achieved an accuracy of 98.56%, precision of 98.51%, recall of 98.51%, and F1-score of 98.51%. In addition, SHAP explainability analysis revealed that execution-related Solidity operations such as success, gas, call, and staticcall significantly influence vulnerability prediction outcomes. Compared with previous CNN-based, LSTM-based, and opcode-based approaches, the proposed framework achieved superior classification performance while additionally providing interpretable security analysis. The findings indicate that hybrid feature engineering combined with explainable ensemble learning provides an effective and lightweight solution for Ethereum smart contract vulnerability detection and blockchain security analysis.