Collaborative Research: SHF: Medium: Learning Semantics of Code To Automate Software Assurance Tasks

深度学习在完成软件工程任务方面已经显示出巨大的潜力。然而，它的能力对于具有挑战性但非常重要的软件保证任务是有限的，例如错误检测、调试、测试输入生成和测试套件优先级排序。这些任务很难形成一个学习问题。困难的主要部分是这些复杂的任务需要对程序语义进行建模。据我们所知，即使是最先进的深度学习模型对程序语义的理解也不足。因此，这些模型无法达到足够的精度和召回率，无法得到更广泛的部署。这些工具不能很好地泛化到未见过的项目中，并且对于源代码中的小扰动也不够健壮。训练模型也需要大量的计算资源和数据。在这个项目中，研究团队的目标是提高用于软件保证的深度学习模型的性能、鲁棒性、通用性和效率，并使深度学习能够用于尚未成功使用深度学习的复杂任务。解决方案的目标是将程序分析、软件工程和深度学习专业知识结合起来，将程序语义编码到程序表示中，以开发新的公式，通过深度学习有效地减少软件保证问题。该项目有三个研究重点：为了学习抽象语义，该项目将研究如何将静态分析算法和静态分析结果与深度学习模型相结合。为了学习具体的语义，该项目将研究如何使用程序执行跟踪来指导深度学习。最后，该项目将研究如何识别当前模型使用的虚假特征，然后应用因果学习来阻止具有虚假特征的模型。研究成果、数据集和工具将被传播到研究界，并将组织研讨会以加强代码深度学习的研究界。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Deep learning has demonstrated great potential for accomplishing software engineering tasks. However, its capabilities are limited for challenging yet very important software assurance tasks such as bug detection, debugging, test input generation, and test suite prioritization. These tasks are hard to formulate into a learning problem. A major part of the difficulty is that these complex tasks require modeling of program semantics.  To the best of our knowledge, even state-of-the-art deep learning models have an insufficient understanding of program semantics. As a result, the models fail to achieve sufficient precision and recall to be more widely deployed. The tools do not generalize well to unseen projects and are not robust to small perturbations in source code. It also takes large amounts of computational resources and data to train the models. In this project, the team of researchers aims to improve the performance, robustness, generalizability and efficiency of deep learning models for software assurance and to enable deep learning for complex tasks that have not yet successfully used deep learning. Solutions will target encoding program semantics into the program representation by combining program analysis, software engineering, and deep learning expertise to develop novel formulations to effectively reduce software assurance problems via deep learning. The project has three research thrusts: To learn with abstract semantics, the project will study how to combine static analysis algorithms and the results from static analysis with deep learning models. To learn with concrete semantics, the project will study how to use program execution traces to guide deep learning. Finally, the project will investigate how to identify spurious features used by the current models and then apply causal learning to discourage models that have spurious features.  Research results, datasets, and tools will be disseminated to the research community, and workshops will be organized to strengthen the research community of deep learning for code.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.