Source code vulnerability detection based on deep learning: a review

Abstract Source code vulnerability detection is a crucial aspect of software security development, and the current use of Large Language Models (LLMs) accelerates not only software development but also the generation and propagation of code vulnerabilities. Traditional code vulnerability detection techniques have limited detection efficiency and accuracy. Deep learning techniques have recently gained distinct advantages in multidimensional feature extraction and large-scale data processing, and their application in code vulnerability detection is evolving from simple classification to multimodal approaches. This paper primarily systematizes and summarises deep learning-based source code vulnerability detection, as well as analyzes and anticipates current challenges and future research directions in this area. The distinction between this review and the preceding reviews: This study investigates the literature of the last four years; Not only does it contain datasets, but also includes model-related research and an analysis of multiple different application scenarios. It’s more current and comprehensive than most previous reviews of this type.