Improving software vulnerability severity prediction model performance with HDLN & FWFS: a two-stage feature selection approach

Abstract

Software Vulnerability Severity Prediction (SVSP) is an evolving domain of software engineering. Ensuring the robustness and continuous operation of software systems relies heavily on the accuracy of SVSP models. However, the effectiveness of SVSP models relies heavily on the features used for training. The presence of redundant, correlated, or irrelevant features can significantly degrade model performance. To address this issue, researchers employ Feature Selection (FS) methods to streamline feature sets, reducing computational costs in model development. However, selecting appropriate FS methods and utilizing their combined strength to achieve a more efficient model remains a challenge in many existing techniques documented in the literature. In this paper, to overcome the problems associated with a large number of features, two new FS modules are proposed which have not been explored so far. The first module employs a combination of Convolutional Neural Network and Bidirectional Long Short-Term Memory, referred to as the Hybrid Deep Learning Network (HDLN) module while the second module utilizes Information Gain and Grey Wolf Optimization techniques, forming the Filter Wrapper Feature Selection (FWFS) module. The efficacy of the proposed feature selection approach is evaluated through the Area under the Curve (AUC). Upon analyzing the results, it is determined that the proposed module enhances the performance of the SVSP model. The prediction models crafted using the HDLN module exhibited the most superior performance, boasting an average AUC value of 0.88. Following closely, the FWFS module, employing Extreme Gradient Boosting as a classifier, demonstrated the second-best performance, achieving an average AUC value of 0.86.