Scholarlink

<p>Purpose: Security-related software fault prediction is essential yet challenging due to limited labeled vulnerability data and the inconsistency of fault patterns across software projects. This study aims to improve the generalizability of fault prediction models using a cross-project learning approach. Method:We propose a cross-project transfer learning framework for security fault prediction, enhanced by feature distribution alignment techniques including Kolmogorov–Smirnov (KS) test, Anderson–Darling test, and Pearson correlation. To effectively address the issue of class imbalance prevalent in vulnerability datasets, we introduce SMADASYN—a hybrid oversampling method that integrates the strengths of SMOTE (Synthetic Minority Oversampling Technique) and ADASYN (Adaptive Synthetic Sampling). The proposed approach is empirically validated across six Apache software projects using a range of classifiers: Decision Tree, Naive Bayes, Logistic Regression, and Random Forest, along with ensemble methods such as Bagging, Boosting, and Stacking. Results: Experimental results demonstrate that Random Forest consistently achieves the highest performance, particularly in terms of F1-score and recall. Statistical validation using the Friedman test and Nemenyi post-hoc analysis confirms the significance and robustness of the results across datasets. Conclusion: The proposed method effectively enhances security fault prediction in cross-project scenarios. By combining feature alignment and advanced sampling techniques, the framework delivers accurate and generalizable predictions, offering practical value for software security assurance.</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

Smadasyn-boosted cross-project transfer learning for effective security fault prediction”

  • 期刊论文

  • 发表日期：

    2025年6月25日

查看全文
  • Rashmi,
  • Arvinder Kaur

摘要

Purpose: Security-related software fault prediction is essential yet challenging due to limited labeled vulnerability data and the inconsistency of fault patterns across software projects. This study aims to improve the generalizability of fault prediction models using a cross-project learning approach. Method:We propose a cross-project transfer learning framework for security fault prediction, enhanced by feature distribution alignment techniques including Kolmogorov–Smirnov (KS) test, Anderson–Darling test, and Pearson correlation. To effectively address the issue of class imbalance prevalent in vulnerability datasets, we introduce SMADASYN—a hybrid oversampling method that integrates the strengths of SMOTE (Synthetic Minority Oversampling Technique) and ADASYN (Adaptive Synthetic Sampling). The proposed approach is empirically validated across six Apache software projects using a range of classifiers: Decision Tree, Naive Bayes, Logistic Regression, and Random Forest, along with ensemble methods such as Bagging, Boosting, and Stacking. Results: Experimental results demonstrate that Random Forest consistently achieves the highest performance, particularly in terms of F1-score and recall. Statistical validation using the Friedman test and Nemenyi post-hoc analysis confirms the significance and robustness of the results across datasets. Conclusion: The proposed method effectively enhances security fault prediction in cross-project scenarios. By combining feature alignment and advanced sampling techniques, the framework delivers accurate and generalizable predictions, offering practical value for software security assurance.