AIdetectorX: A Vulnerability Detector Based on TCN and Self-attention Mechanism

Abstract

A vulnerability detector should have both excellent detection capabilities (such as high accuracy, low false positive rate, low false negative rate, etc.) and little time overhead. However, existing vulnerability detection methods often rely on manual intervention by human experts or result in high false positives and high false negatives. Additionally, the development of deep learning techniques has prompted many scholars to conduct research in the field of vulnerability detection. Since Temporal Convolutional Networks (TCN) have causal relationships between their convolutional layers and can process information in parallel, while self-attention mechanism can attach more attention to the information related to vulnerabilities. Therefore, in this paper, we combine TCN and self-attention mechanism for vulnerability detection. This leads to the design and implementation of an improved deep learning-based vulnerability detector, called AIdetectorX. We conduct experiments on publicly available and widely used datasets for evaluating the effectiveness of AIdetectorX. Evaluation results suggest that AIdetectorX is effective for vulnerability detection and that combining TCN and self-attention mechanism can lead to higher detection capabilities and decrease time overhead.