Advanced Automated Vulnerability Scoring: Improving Performance with a Fine-Tuned BERT-CNN Model

The ever-growing prevalence of software systems necessitates robust and efficient vulnerability assessment methods. An automated vulnerability assessment model can significantly streamline this process, providing accurate identification of vulnerability severity levels based on the given textual descriptions. This paper addresses this critical need by proposing a novel vulnerability score predictor model. Our model uses Bidirectional Encoder Representations from Transformers (BERT) and deep learning to classify vulnerability metrics accurately. We have fine-tuned the BERT tokenizer on NVD’s descriptions and used TextCNN for better performance. This approach represents a significant advancement in the field of automated vulnerability assessment, offering substantial benefits for security professionals by streamlining vulnerability identification and prioritization. The carried out experimental evaluations in comparison to various deep learning models, including LSTM, CNN, and LSTM-Attention, as well as two state-of-the-art automated vulnerability assessment baselines indicate an accuracy score of over $\mathbf{9 7 \%}$ across all vulnerability metrics.