Conference paper Embargoed Access
Ilias Kalouptsoglou; Miltiadis Siavvas; Dimitrios Tsoukalas; Dionysios Kehagias
Vulnerability prediction constitutes a mechanism that enables the identification and mitigation of software vulnerabilities early enough in the development cycle, improving the security of software products, which is an important quality attribute according to ISO/IEC 25010. Although existing vulnerability prediction models have demonstrated sufficient accuracy in predicting the occurrence of vulnerabilities in the software projects with which they have been trained, they have failed to demonstrate sufficient accuracy in cross-project prediction. To this end, in the present paper we investigate whether the adoption of deep learning along with software metrics may lead to more accurate cross-project vulnerability prediction. For this purpose, several machine learning (including deep learning) models are constructed, evaluated, and compared based on a dataset of popular real-world PHP software applications. Feature selection is also applied with the purpose to examine whether it has an impact on cross-project prediction. The results of our analysis indicate that the adoption of software metrics and deep learning may result in vulnerability prediction models with sufficient performance in cross-project vulnerability prediction. Another interesting conclusion is that the performance of the models in cross-project prediction is enhanced when the projects exhibit similar characteristics with respect to their software metrics.
Files are currently under embargo but will be publicly accessible after September 29, 2022.