| As the software supply chain becomes increasingly complex,source code security vulnerabilities in open source software will also have farreaching adverse effects.Therefore,it is becoming more and more urgent to invent a tool that can accurately detect vulnerabilities in open source software.Due to the rise of AI technology in the CV and NLP fields at this stage,many researchers have applied deep learning technology to source code vulnerability detection.However,everyone is working towards improving the accuracy of vulnerability detection models,ignoring the most important issue of model robustness.In the field of software security,a model that is not robust enough cannot be truly applied.Once exploited by attackers in the production environment,it will cause immeasurable consequences.In order to prevent such consequences,this paper refers to many research works related to vulnerability detection,and makes the following contributions:First of all,this work points out the problems of noise and single sample mode in vulnerability datasets commonly used by researchers in the deep-learning-based vulnerability detection task,and reveals the important reasons for the low robustness of the model.Secondly,this topic explores the source of noise and the actual impact of noise on model training,and proposes a denoising method based on ensemble learning and confidence pruning,which can reduce the impact of noise on the model through denoising.Then,aiming at the problem of single sample pattern in the vulnerability dataset,this topic proposes a data augmentation method for semantic protection code transformation.Through data augmentation,the sample pattern in the dataset can be enriched and the performance of the model can be improved.Through data augmentation,the problem of low robustness caused by single sample pattern and insufficient data volume is solved.Finally,in order to meet the needs of a reasonable vulnerability detection model robustness measurement method in the field,we propose a multi-dimensional vulnerability detection model robustness measurement method,and conduct a multi-dimensional analysis of some classic vulnerability detection models.With this evaluation method,we also demonstrate that denoising and data augmentation have a significant effect on improving model robustness.And we provide two feasible solutions for optimizing the robustness of the vulnerability detection model. |
