Scholarlink

<p>In the past few years, the advancement of deep learning based AI technologies has accelerated the research on automatic software vulnerability detection. However, deep learning models trained with vulnerability data have limited learning ability with high rates of false negative (FN) and false positive (FP), which are mainly caused by small and imbalanced sample problems of vulnerability data. Therefore, we propose a multiperspective data augmentation approach (called MPDA) and apply it to enhance data quality to improve deep learning for software vulnerability detection. MPDA automatically augments software vulnerability data from different perspectives by its three augmenting components designed as <i>augmenting by oversampling</i>, <i>augmenting by GAN</i>, and <i>augmenting by fuzzy sampling</i>. We also design three algorithms, the Oversampling Strategy Selection (OSS) algorithm, the GAN-based data generating algorithm, and the Fuzzy Sampling Strategy Selection (FSS) algorithm, to help MPDA automatically achieve the optimal augmentation effect. The evaluation results on the Juliet Java Suite dataset and the application of MPDA to five widely used models in deep learning-based vulnerability detection to detect 29 types of vulnerability demonstrated that our approach consistently improved the performance of each deep learning model for vulnerability detection by at least 12% in terms of the F1 score, except for Transformer, which is 3. 9%.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

MPDA: a data augmentation approach to improve deep learning for software vulnerability detection

  • 期刊论文

  • 发表日期：

    2025年7月14日

查看全文
  • Feiqiao Mao,
  • Yingxiang Yuan,
  • Xingyang Du,
  • Li Gao,
  • Zhihua Du

摘要

In the past few years, the advancement of deep learning based AI technologies has accelerated the research on automatic software vulnerability detection. However, deep learning models trained with vulnerability data have limited learning ability with high rates of false negative (FN) and false positive (FP), which are mainly caused by small and imbalanced sample problems of vulnerability data. Therefore, we propose a multiperspective data augmentation approach (called MPDA) and apply it to enhance data quality to improve deep learning for software vulnerability detection. MPDA automatically augments software vulnerability data from different perspectives by its three augmenting components designed as augmenting by oversampling, augmenting by GAN, and augmenting by fuzzy sampling. We also design three algorithms, the Oversampling Strategy Selection (OSS) algorithm, the GAN-based data generating algorithm, and the Fuzzy Sampling Strategy Selection (FSS) algorithm, to help MPDA automatically achieve the optimal augmentation effect. The evaluation results on the Juliet Java Suite dataset and the application of MPDA to five widely used models in deep learning-based vulnerability detection to detect 29 types of vulnerability demonstrated that our approach consistently improved the performance of each deep learning model for vulnerability detection by at least 12% in terms of the F1 score, except for Transformer, which is 3. 9%.