Ethereum Smart Contract Security: "Deep Learning Approaches for Vulnerability Detection"

Ethereum smart contracts have transformed decentralized applications, however they are prone to a wide range of security vulnerabilities. This paper analyzes smart contract vulnerabilities using the Smart Contract Vulnerabilities Dataset, an improved dataset compared to the MessiQ Smart Contract Dataset. This research comprises 2,217 samples and classify them into four primary vulnerabilities: Reentrancy, Integer Overflow/Underflow, Timestamp Dependency, and Dangerous Delegatecall. To improve security analysis, the dataset is preprocessed by eliminating comments, special characters, and stopwords, then applying label encoding. Two machine learning models, an attention based BiLSTM CNN and DeBERTA V3, are tested for their efficiency in vulnerability detection. The BiLSTM CNN model performed with an accuracy of 92%, better than the DeBERTA V3 model, which performed with an accuracy of 86%. Evaluation metrics like Precision, Recall, and F1 score emphasize the accuracy of our method. Such results contribute to the evolution of automated smart contracts vulnerabilities detection, which will assist developers in enhancing blockchain security.