Abstract: The available smart contract vulnerability detection schemes mostly rely on expert-defined rules, which lack flexibility and struggle with new unknown vulnerabilities. To address this challenge, we present a novel framework called domain adaptive security analysis framework (DASAF). Firstly, we obtain the execution logic of smart contract opcodes and convert them into meaningful sequential features. Secondly, to overcome the inherent data bias in deep learning models, which leads to model aging and difficulty in achieving strong generalization performance due to insufficient labeled samples in new unknown vulnerabilities, the DASAF framework introduces adversarial generative network structure and domain adaptation techniques. Finally, we evaluate the effectiveness of the DASAF framework in the field of smart contract vulnerability analysis and detection using a public benchmark dataset, and compare it with similar schemes. The experimental results demonstrate the superiority of the DASAF framework over comparable approaches.

Key words: smart contract, domain adaptation techniques, generative adversarial network, vulnerability detection, deep learning