Active Learning-Based Smart Contract Vulnerability Detection Using Ensemble Classifiers

Abstract

Smart contract vulnerabilities continue to threaten the security and reliability of blockchain-based applications, especially as blockchain systems are increasingly integrated into digital finance, supply chains, identity management, and other intelligent service environments. In these convergent digital systems, dependable vulnerability detection is important before deployment because deployed contracts are often difficult to modify and may expose users and organizations to operational, financial, and security risks. This paper presents an active learning-based framework for smart contract vulnerability detection using ensemble classifiers. Using the BCCC-VulSCs-2023 dataset, the proposed approach applies preprocessing, feature normalization, feature selection, and iterative uncertainty-based sample refinement to improve classification performance while making efficient use of labeled data. Three ensemble classifiers, namely Random Forest, XGBoost, and CatBoost, are evaluated under a unified pipeline using accuracy, precision, recall, and F1-score. Experimental results show that Random Forest with active learning achieved the strongest overall performance, obtaining a mean precision of 97.34%, mean recall of 95.95%, mean F1-score of 96.64%, and mean accuracy of 96.67%. The findings indicate that active learning can strengthen ensemble-based smart contract vulnerability detection and that robust tree-based methods can provide a practical and efficient option for securing blockchain-enabled intelligent applications.