Deep Learning for Software Security

Computer software is embedded everywhere in our modern lives. Consequently, software vulnerabilities (SVs), which attackers can carry out malicious actions, have become a serious issue. Although there have been significant research efforts developed for SV detection, there are several challenges that current methods fail to address: i) transferring the learning on SVs from label-rich software projects to unlabelled ones, ii) detecting SVs at a fine-grained level, iii) detecting coherent segments of functions for detecting SVs. Grounded in the theoretical sophistication of recent advances in deep learning, this thesis aims to rigorously investigate and provide solutions to these three challenging problems.