Abstract: To address the limitations of the existing smart contract vulnerability detection technology, including low detection efficiency, inadequate automation, and the inability to realize large-scale smart contract sample detection, this study proposed a method for smart contract vulnerability detection technology based on machine learning. The method first preprocessed the smart contract dataset, converted the source code of the smart contract into a sequence of opcodes, and formulated the opcode abstraction simplification rules for simplification. On this basis, 2025-dimensional bigram features were extracted from the simplified opcode sequence dataset using the N-gram model, and three feature representations were constructed by using the embedding method for feature selection and principal component analysis for feature dimensionality reduction, respectively. Then, the Borderline SMOTE method, an improved algorithm of SMOTE, was used to equalize the positive and negative sample imbalance dataset. Finally, four algorithms, namely, decision tree, support vector machine, random forest, and XGBoost, were applied to construct the vulnerability detection model, respectively. The experimental results show that the vulnerability detection model of random forest has an average accuracy of 93.60%, and the overall performance Macro-F1 reaches 93.91%, which can efficiently detect multiple vulnerabilities.

Key words: blockchain, machine learning, smart contract, vulnerability detection