CAREER: Robustifying Machine Learning for Cyber-Physical Systems

这个用于网络物理系统（CPS）的鲁棒机器学习（ML）项目专注于检测和减少ML模型的漏洞，这些模型已经变得普遍，并正在部署用于现实生活中的CPS应用（包括自动驾驶汽车和机器人飞行器）中的决策。机器学习方法的日益增长的前景，例如深度卷积神经网络（CNN）和深度强化学习（DRL），用于CPS（例如，自动驾驶汽车）已经引起了对自主代理的安全性和鲁棒性的关注。最近关于生成对抗性攻击的工作表明，一个坏行为者戏剧性地欺骗深度学习（DL）模型在计算上是可行的。除了对抗性攻击之外，这种DL模型还可能屈服于所谓的“边缘情况”，即现实生活中的操作情况呈现的数据在训练数据集中没有得到很好的表示。这些案例是最近发生的许多自动驾驶汽车事故的主要原因。虽然初步研究已经开始解决特定攻击模型的场景，但在CPS背景下，关于ML模型对"边缘案例"和对抗性攻击的检测和适应，仍然存在显著的知识差距。这个项目建立了一个元学习，基于监督框架和相关算法来检测和缓解ML系统漏洞，这将大大降低使用ML进行安全的风险时间关键型系统。科学驱动应用是自动驾驶汽车和机器人。算法验证和评估使用了爱荷华州的实验性自动驾驶汽车和机器人测试台，并与交通运输研究所和英伟达合作。研究与教育相结合，以支持在系统理论和数据科学的关键跨学科领域培养学生的目标，这迫切需要快速和高质量的劳动力发展，以促进美国的持续经济和社会增长。教育计划还包括研究生和本科生一级的课程开发、本科生研究经验、学术竞赛和涉及高中学生和教师的外联活动。该项目的成果将支持NSF的"利用数据革命"的使命，用于目前涉及ML或未来将涉及ML的许多关键CPS，例如制造流程，电网，智能城市和交通系统，使它们更安全，该奖项反映了NSF的法定使命，并被认为值得通过使用基金会的知识产权进行评估来支持。优点和更广泛的影响审查标准。

This robustifying machine learning (ML) for cyber-physical systems (CPSs) project focuses on detecting and reducing the vulnerabilities of ML models that have become pervasive and are being deployed for decision-making in real-life CPS applications including self-driving cars, and robotic air vehicles. The growing prospect of machine learning approaches such as deep Convolutional Neural Networks (CNN) and deep Reinforcement Learning (DRL) being used in CPSs (e.g., self-driving cars) has raised concerns around safety and robustness of autonomous agents. Recent work on generating adversarial attacks have shown that it is computationally feasible for a bad actor to fool a deep learning (DL) model dramatically. Apart from adversarial attacks, such DL models can also succumb to the so-called 'edge-cases' where the real-life operational situation presents data that are not well-represented in the training data set. Such cases have been the primary reason for quite a few self-driving car accidents recently. Although initial research has begun to address scenarios with specific attack models, there remains a significant knowledge gap regarding detection and adaptation of ML models to 'edge-cases' and adversarial attacks in the context of CPS.With this motivation, this project builds a meta-learning-based supervisory framework and associated algorithms to detect and mitigate ML system vulnerabilities which will substantially reduce the risk in using ML for safety and time-critical systems. The science driver applications are self-driving cars and robotics. The algorithm validation and evaluation use experimental self-driving cars and robotics test beds at Iowa State in collaboration with the Institution of Transportation and NVIDIA.Research is integrated with education to support the goal of training students in the critical interdisciplinary area of system theory and data science, which is in dire need of rapid and quality workforce development for sustained economic and social growth of the United States. Education plans also include curriculum development at graduate and undergraduate level, undergraduate research experience, academic competitions and outreach activities involving both high school students and teachers. Outcomes of this project will support NSF's mission of "Harnessing the Data Revolution" for many critical CPSs that currently involve ML or will involve it in future, such as manufacturing processes, power grid, smart cities and transportation systems, to make them safer, more efficient and cost effective.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.