Advancing software security: DCodeBERT for automatic vulnerability detection and repair

Abstract

The exponential growth of software complexity has led to a corresponding increase in software vulnerabilities, necessitating robust methods for automatic vulnerability detection and repair. This paper proposes DCodeBERT, a large language model (LLM) fine-tuned for vulnerability detection and repair in software code. Leveraging the pre-trained CodeBERT model, DCodeBERT is designed to understand both natural language and programming language context, enabling it to effectively identify vulnerabilities and suggest repairs. We conduct experiments to evaluate DCodeBERT’s performance, comparing it against several baseline models. The results demonstrate that DCodeBERT outperforms the baselines in both vulnerability detection and repair tasks across multiple programming languages, showcasing its effectiveness in enhancing software security.