A Deep Learning-based Approach for an Automated Vulnerability Detection in Source Code

With the advancements and significant increase in today’s software projects, software vulnerabilities have become a crucial and serious concern to the software industry. These software defects pose familiar cybersecurity threats and identifying potential vulnerable code is difficult because of various elements related to volume, diversity, and complexity. Traditionally, vulnerabilities in source code have been detected either statically by relying on expert knowledge, which is time consuming, or through dynamic analysis tools. However, lacking robust investigation of source code resulted in decreased efficacy of the existing solutions. In this respect, as an alternative to traditional source code detection methods, deep learning seems to have become an attractive solution. Thus, in this research work, we propose a deep learning-based neural networks approach to construct a function-level vulnerability analysis for an automated identification of vulnerabilities in source code. The proposed solution can handle source code vulnerability detection on function level without expert knowledge by utilizing deep learning convolutional neural networks (CNNs) and artificial neural networks (ANNs). To generate meaningful numerical feature representation of input source code functions, we have generated abstract syntax tree (AST). Moreover, we enhanced the deep learning neural network models through the selection of optimal parameter values and employing a validation methodology. To evaluate the performance of the proposed models we used a public dataset with verified vulnerability labels using various performance metrics such as Precision, Accuracy, F1-score, Recall and ROC-AUC. After conducting multiple experiments, the best result (Accuracy = 84.22%) was achieved using our ANN model.