VulDeBERT: A Vulnerability Detection System Using BERT
Soolin Kim, Jusop Choi, Muhammad Ejaz Ahmed, Surya Nepal, Hyoungshick Kim
Abstract
Deep learning technologies recently received much attention to detect vulnerable code patterns accurately. This paper proposes a new deep learning-based vulnerability detection tool dubbed VulDeBERT by fine-tuning a pre-trained language model, Bidirectional Encoder Representations from Transformers (BERT), on the vulnerable code dataset. To support VulDeBERT, we develop a new code analysis tool to extract well-represented abstract code fragments from C and C++ source code. The experimental results show that VulDeBERT outperforms the state-of-the-art tool, VulDeePecker [1] for two security vul- nerability types (CWE-119 and CWE-399). For the CWE-119 dataset, VulDeBERT achieved an Fl score of 94.6 %, which is significantly better than VulDeePecker, achieving an Fl score of 86.6 % in the same settings. Again, for the CWE-399 dataset, VulDeBERT achieved an Fl score of 97.9 %, which is also better than VulDeePecker, achieving an Fl score of 95 % in the same settings.