Large Language Models for Synthetic Dataset Generation: A Case Study on Ethereum Smart Contract DoS Vulnerabilities

Abstract

The use of Ethereum smart contracts has significantly influenced sectors that depend on decentralized control and automated financial transactions. However, ensuring their security and reliable execution remains a complex task. Among the most serious challenges is the Denial of Service (DoS) attack, which can make a contract nonfunctional. The broad range of vulnerabilities that enable these attacks complicates prevention efforts. While dynamic security tools exist, they require substantial computational resources, and machine learning-based approaches face limitations due to a lack of training data. To address this issue, we propose a methodology using Large Language Models (LLMs), specifically Antropic’s Claude and OpenAI’s GPT-4, to generate synthetic examples of Ethereum smart contracts exposed to DoS attacks. Our results show that, with properly designed prompts, these models can produce high-quality synthetic examples, enabling the development of classification and anomaly detection models.