ProRLearn: boosting prompt tuning-based vulnerability detection by reinforcement learning

ProRLearn: boosting prompt tuning-based vulnerability detection by reinforcement learning

연구 분야: Infrastructure

학회: Automated Software Engineering

Software vulnerability detection is a critical step in ensuring system security and data protection. Recent research has demonstrated the effectiveness of deep learning in automated vulnerability detection. However, it is difficult for deep learning models to understand the semantics and domain-specific knowledge of source code. In this study, we introduce a new vulnerability detection framework, ProRLearn, which leverages two main techniques: prompt tuning and reinforcement learning. Since existing fine-tuning of pre-trained language models (PLMs) struggles to leverage domain knowledge fully, we introduce a new automatic prompt-tuning technique. Precisely, prompt tuning mimics the pre-training process of PLMs by rephrasing task input and adding prompts, using the PLM’s output as the prediction output. The introduction of the reinforcement learning reward mechanism aims to guide the behavior of vulnerability detection through a reward and punishment model, enabling it to learn effective strategies for obtaining maximum long-term rewards in specific environments. The introduction of reinforcement learning aims to encourage the model to learn how to maximize rewards or minimize penalties, thus enhancing performance. Experiments on three datasets (FFMPeg+Qemu, Reveal, and Big-Vul) indicate that ProRLearn achieves performance improvement of 3.27–70.96% over state-of-the-art baselines in terms of F1 score. The combination of prompt tuning and reinforcement learning can offer a potential opportunity to improve performance in vulnerability detection. This means that it can effectively improve the performance in responding to constantly changing network environments and new threats. This interdisciplinary approach contributes to a better understanding of the interplay between natural language processing and reinforcement learning, opening up new opportunities and challenges for future research and applications.