| Abstrakt: |
Software vulnerabilities reported every year increase exponentially, leading to the exploitation of software systems. Hence, when a vulnerability is reported, a requirement arises to patch it as early as possible. Generally, this process requires some time and effort. For proper channelizing of the efforts, a requirement comes to predict the severity of the vulnerability so that the more critical ones can be given a higher priority. Therefore, a need arises to build a model that can analyze the data available on vulnerabilities and predict their severity. The experiment of this study is conducted on vulnerability reports of five software of Mozilla. As the data is textual, text mining techniques are applied to preprocess the data and form feature vectors. This input as text creates very high dimensional feature vectors leading to the requirement of dimensionality reduction. Hence, feature selection is done using chi-square and information gain. To develop the classifier, seven machine learning algorithms are chosen. Hence, fourteen software vulnerability severity prediction models (SVSPM) are developed. The result analysis allowed us to find the best-performing SVSPM. It is concluded that the model performed better for the medium and the critical severity level of the vulnerability. Out of the two feature selection techniques, information gain gave better results. An optimum number of features is also determined at which SVSPM gave good results. The best SVSPM using a machine learning algorithm corresponding to each dataset is found as well. A comparison is also made to identify significant differences among various SVSPMs developed using Friedman and Wilcoxon Signed Rank test. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink