Hybrid Detection of Vulnerabilities in Smart Contracts
A hybrid tool for auditing smart contracts using signature-based detection and machine learning, achieving highest score at UC Berkeley's Decentralized Finance Research Seminar.
• Developed a hybrid tool for auditing smart contracts using signature-based detection and machine learning, leveraging a dataset of 36,670 contracts to identify vulnerabilities like reentrancy and integer overflows. • Implemented a logistic regression model and fine-tuned BERT classifier to classify contracts as safe or unsafe, achieving an accuracy of 75% through extensive feature engineering and cross-validation. • Implemented a scalable signature-matching algorithm capable of analyzing smart contracts in under 2 seconds, achieving an accuracy of 86% when combined with the ML classifier into a hybrid system.