A One-class variational autoencoder for smart contract vulnerability detection

Abstract

Smart contracts and blockchain technology have revolutionized our transactions and interactions with digital systems, yet their vulnerabilities can lead to devastating consequences such as financial losses, data breaches, and compromised system integrity. Existing detection methods, including static analysis, dynamic analysis, and machine learning-based approaches, have their limitations, such as requiring large amounts of labeled data or being computationally expensive. To address these limitations, we propose a novel approach that leverages a One-Class Variational Autoencoder (VAE) with CodeBERT for data pre-processing to detect vulnerabilities in smart contracts. Our approach achieved a higher F1 score (88.93%) compared to the baselines evaluated, even when labeled data is limited. This paper contributes to the development of effective and efficient vulnerability detection methods, ultimately enhancing the security and reliability of smart contracts and blockchain-based systems. By demonstrating superior performance in imbalanced data scenarios, our method offers a practical solution for real-world applications in blockchain security.