The Random Neural Network as a Bonding Model for Software Vulnerability Prediction

Software vulnerability prediction is an important and active area of research where new methods are needed to build accurate and efficient tools that can identify security issues. Thus we propose an approach based on mixed features that combines text mining features and the features generated using a Static Code Analyzer. We use a Random Neural Network as a bonding model that combines the text analysis that is carried out on software using a Convolutional Neural Network, and the outputs of Static Code Analysis. The proposed approach was evaluated on commonly used datasets and led to 97% training accuracy, and 93%- 94% testing accuracy, with a 1% reduction in false positives with respect to previously published results on similar data sets.