基于机器学习的小样本软件缺陷检测技术的研究

Software quality is the fundamental basis for the reliability of a software system. Software defect is the major cause that leads to poor software quality, and effective detection of software defect can help to improve the quality of a software system. The key for software defect detection is to construct a predictive model from the current software in order to correctly identify the defective modules in the software. However, the inherent small-sample property of software defect detection makes the information for model construction insufficient, which accounts for the unsatisfactory performance of the state-of-the-art methods. In this project, we are planning to study the novel software defect detection methods given the training sample is small, based on which five typical results would be achieve, including proposing a novel software defect detection method which is able to utilize the undetected module to improve the modeling performance; proposing a novel software defect detection method which is able to exploit the heterogeneous data sources to improve modeling performance; proposing a novel software defect detection method which is able to actively ask for the modules that is the most helpful for performance improvement; proposing a novel software defect detection method which is able to enhance the sensitivity of the learner to the defects based on unequal misclassification cost; proposing a novel software defect detection method which is able to transfer knowledge from other software to favor the training of the defect detection model in current software systems. The expected outcomes include the publication of 8-10 papers on international journals/conference/native top journals, application for 2-3 National patents, supervision for a number of graduate students.

软件质量是软件可靠运行的基础，而软件缺陷是破坏软件质量的元凶之一，有效检测软件缺陷是保障软件质量的重要手段。软件缺陷自动检测的关键在于对当前软件的缺陷模式进行有效建模。然而，软件缺陷检测任务固有的小样本特性造成了可供建模的信息不充分，而现有缺陷检测技术的建模方法难以有效适应这一特性，从而造成缺陷检测性能不佳。本项目拟基于新型机器学习风范，对能够适应小样本特性的软件缺陷检测技术进行深入研究，提出一种能利用当前软件中待检测模块进行学习的检测方法；提出一种能利用当前软件中多种异构数据资源进行学习的检测方法；提出一种能通过主动获取最有价值模块进行学习的检测方法；提出一种通过引入非均衡缺陷错检代价增强对缺陷敏感性的检测方法；提出一种能借鉴已发布软件的缺陷模式以辅助学习当前软件缺陷模式的检测方法。本项目可望在国际期刊/会议/国内一级学报上发表论文8-10篇，申请国家发明专利2-3项，培养研究生多名。

本项目围绕软件缺陷检测固有的训练集小样本性，以弥补因小样本造成的信息不充分为主线，就能够针对软件缺陷检测固有的小样本特性进行学习建模的软件缺陷检测技术进行了研究，提出了能够有效利用当前软件中待检测模块信息进行学习的方法、能够利用当前软件中其他非代码数据资源进行学习的方法、能够通过主动获取最有价值的模块进行学习的方法、通过引入非均衡的缺陷检测错误代价以提升学习器对缺陷的敏感性的方法、能够借鉴已发布软件中的缺陷模式辅助学习的方法，提升了小样本软件缺陷检测的性能。本项目发表(录用)论文14篇，包括IJCAI、AAAI等一流国际会议与《ACM/IEEE Trans.》等重要国际期刊，获国家发明专利1项，另有1项在审；主编国际研讨会论文集4部，组织国际研讨会4次。在人才培养方面取得良好进展。