VDSimilar: Vulnerability detection based on code similarity of vulnerabilities and patches

摘要

Vulnerability detection using machine learning is a hot topic in improving software security. However, existing works formulate detection as a classification problem, which requires a large set of labelled data while capturing semantical and syntactic similarity. In this work, we argue that similarity in the view of vulnerability is the key in detecting vulnerabilities. We prepare a relatively smaller data set composed of both vulnerabilities and associated patches, and attempt to realize security similarity from (ⅰ) the similarity between pair of vulnerabilities and (ⅱ) the difference between a pair of vulnerability and patch. To achieve this, we setup the detection model using the Siamese network cooperated with BiLSTM and Attention to deal with source code, Attention network to improve the detection accuracy. On a data set of 876 vulnerabilities and patches of OpenSSL and Linux, the proposed model (VDSimilar) achieves about 97.17% in AUC value of OpenSSL (where the Attention network contributes 1.21% than BiLSTM in Siamese), which is more outstanding than the most advanced methods based on deep learning.