CAREER: A Dynamic Program Monitoring Framework Using Neural Network Hardware

软件漏洞和安全攻击每年给美国经济造成超过1500亿美元的损失。然而，在这方面还没有重大的创新。本研究项目旨在借助基于神经网络的硬件来改变这一事实。如果项目成功，它将对当前的行业实践产生重大影响，并引发新的趋势。它将鼓励公司投资使用神经网络硬件进行调试和安全攻击分析的新技术，并为硬件实现制定引人注目的用例，从而影响对神经网络硬件的持续投资。此外，该项目将有助于少数民族服务机构的研究和教育活动。学生将通过论文、论文工作和本科研究工作紧密地融入项目。PI将在本科和研究生课程中纳入新兴建筑设计及其编程。此外，PI还将通过暑期实习，让当地高中生参与计算机科学相关项目。神经网络是一种模仿人类大脑的机器学习技术。因此，神经网络硬件提供了一些独特的功能，可以在许多不同的方式中使用。本课题提出利用神经网络硬件进行“程序监控”。程序执行监控通常用于检测软件错误、性能问题、安全攻击等。神经网络硬件将学习程序的正常“行为”。然后，它将检测到这种行为的任何偏差。这种偏差可以归因于软件错误、性能问题或安全攻击。建议的方法提供了处理这些问题的一般框架。由于神经网络硬件的在线学习和测试能力，该框架将适应程序输入、代码和平台的任何变化。

Software bugs and security attacks cripple US economy by costing more than $150 billion a year. However, there has been no major innovation in this context. This research project aims to change that fact with the help of neural network based hardware. If the project is successful, it will significantly affect current industry practices and spur a new trend. It will encourage companies to invest in new techniques for debugging and security attack analysis using neural network hardware and make a compelling use case for the hardware implementation, thereby influencing continuous investment in neural network hardware. In addition, the project will contribute to the research and educational activities of a minority serving institution. Students will be tightly integrated into the project through dissertation, thesis work, and undergraduate research work. The PI will incorporate emerging architecture design and its programming in undergraduate and graduate coursework. Moreover, the PI will involve local high school students in computer science related projects through summer internships.Neural network is a machine learning technique that mimics human brain. Therefore, neural network hardware provides some unique capabilities that can be utilized in many different ways. This project proposes to utilize neural network hardware for "program monitoring". Program execution monitoring is often used to detect software bugs, performance issues, security attacks etc. Neural network hardware will learn the normal "behavior" of the program. Then it will detect any deviation of such behavior. Such deviation can be attributed to software bugs, performance issues or security attacks. The proposed approach provides a general framework for handling these issues. Due to online learning and testing capability of neural network hardware, the framework will be adaptive to any change in program inputs, code, and platforms.