A Deep Learning Approach for Classifying Vulnerability Descriptions Using Self Attention Based Neural Network

Cyber threat intelligence (CTI) refers to essential knowledge used by organizations to prevent or mitigate against cyber attacks. Vulnerability databases such as CVE and NVD are crucial to cyber threat intelligence, but also provide information leveraged in hundreds of security products worldwide. However, previous studies have shown that these vulnerability databases sometimes contain errors and inconsistencies which have to be manually checked by security professionals. Such inconsistencies could threaten the integrity of security products and hamper attack mitigation efforts. Hence, to assist the security community with more accurate and time-saving validation of vulnerability data, we propose an automated vulnerability classification system based on deep learning. Our proposed system utilizes a self-attention deep neural network (SA-DNN) model and text mining approach to identify the vulnerability category from the description text contained within a report. The performance of the SA-DNN-based vulnerability classification system is evaluated using 134,091 vulnerability reports from the CVE details website. The experiments performed demonstrates the effectiveness of our approach, and shows that the SA-DNN model outperforms SVM and other deep learning methods i.e. CNN-LSTM and graph convolutional neural networks.