Identify Vulnerability Fix Commits Automatically Using Hierarchical Attention Network

Abstract

The application of machine learning and deep learning in the field of vulnerability detection is a hot topic in security research, but currently it faces the problem of lack of dataset. Considering vulnerable code can be obtained from vulnerability fix commits, we propose an automatic vulnerability commit identification tool based on hierarchical attention network (HAN) to expand existing vulnerability dataset. HAN can model the input data at the word and sentence levels respectively and pay attention to the changes in the characteristics of different words in different categories, which improves the classification performance. Experimental results show that the accuracy and F1 of our model both achieve 92%. Through the vulnerability fix commit, researchers can quickly locate the vulnerable code. And extracting vulnerable code from open-source software can effectively expand the current dataset due to the enormous number of open-source software.