A multi-task deep learning based vulnerability severity prediction method

With the rapid development of information technology, cyber security issues are becoming increasingly critical. Vulnerability risk assessment plays a key role in ensuring network security. The existing vulnerability warning system is limited by the constant changes and emergence of vulnerabilities, which cannot respond to potential threats in a timely and effective manner and has serious lag. This paper addresses the low accuracy of vulnerability severity prediction using publicly available vulnerability data. It introduces a multitask deep learning method that utilizes DistilBERT and BiLSTM for feature extraction through pre-trained models and multi-task learning. The extracted sentence vectors serve as trained models for all eight Common Vulnerability Scoring System (CVSS) metrics. Then each of the eight predicted results is used to calculate the vulnerability severity score by CVSS. The experimental results show that the DistilBERT+BiLSTM multitask deep learning technique outperforms traditional models by approximately 30%. In comparison to using the BERT model, the DistilBERT+BiLSTM model not only exhibits improved results but also delivers higher performance while requiring fewer computational resources, making it more computationally efficient. The proposed model exhibits significant improvement over existing methods for the F1 values of the eight CVSS metrics, highlighting the effectiveness of applying multi-task deep learning techniques to vulnerability severity prediction.