Detection of Software Vulnerabilities in Source Code Using Interpretable Deep Learning

Abstract

Detection of Software Vulnerabilities in Source Code Using Interpretable Deep Learning WORA SALAMI Aicha Dept of Computer Science and Engineering Graduate School Hanyang University The detection of software vulnerabilities plays a crucial role in ensuring the security and reliability of software systems. Traditional vulnerability detection methods often rely on manual code inspection, which is time- consuming and error-prone. To address this challenge, many researchers have explored applying Deep Learning techniques to automate vulnerability detection. However, the need for interpretability in Deep Learning models poses a significant limitation, as understanding the reasoning behind their predictions is essential for practical vulnerability analysis and remediation. In this research, we propose an approach for detecting software vulnerabilities in source code using Interpretable Deep Learning. Our method combines Deep Learning models with interpretability techniques to provide explanations for vulnerability detections. To evaluate the performance of our proposed method, we conduct experiments using a Python environment with PyTorch as the Deep Learning framework. We train and validate our model using a diverse dataset of source code samples, including both vulnerable and non-vulnerable functions. The evaluation metrics used include accuracy, precision, recall, and F1-score, providing a comprehensive assessment of the model's performance.