Project Description

We study about improving security by leveraging massive data with deep learning. Deep learning has already used in various fields of security. One of the typical example is malware detection. We are focusing on new security fields we can apply deep learning.

Security community have studied a lot of techniques to find security vulnerabilities in source code. However, previous studies have fundamental limitations. For example, some needs additional effort of developers, and others are hard to compute large source code. To overcome their limitations, we leverage deep learning. Some studies have already introduced machine learning techniques to find vulnerabilities, but they focused on high level features such as metrics like code complexity, or syntax tree of source code. In contrast with them, we focus on more specific features of source code such as data dependency or control flow dependency. These dependencies are represented by graphs. We train the graphs using existing deep learning models for graphs to classify whether a functions has a vulnerability.