CPS: Small: Collaborative Research: SecureNN: Design of Secured Autonomous Cyber-Physical Systems Against Adversarial Machine Learning Attacks

自动驾驶汽车、无人机和智能交通等网络物理系统严重依赖机器学习技术来实现不断提高的自主性。在自动驾驶汽车的例子中，深度学习或深度神经网络可用于感知、传感器融合、预测、规划和控制任务。无论这种机器学习技术变得多么强大，它们也暴露了一个新的攻击面，这可能导致在安全和安全关键场景中容易受到对抗性攻击并产生潜在的有害后果。本项目研究自主网络物理系统面临的对抗性机器学习挑战，目的是制定防御策略。该项目将与东北大学STEM(科学、技术、工程和数学)教育中心和加州大学欧文分校接入和包容中心办公室合作，让本科生、女性和少数族裔学生参与独立研究项目。该项目由两个相互依存的研究推动力组成，一个用于调查对抗性攻击，另一个用于设计对策，旨在确保自主网络物理系统的关键深度学习软件组件，如感知、障碍预测和车辆规划和控制。对自主网络物理系统感兴趣的主要深度学习技术包括用于检测的卷积神经网络、用于预测的递归神经网络和用于控制的深度强化学习。该项目的技术创新包括基于ADMM(乘子交替方向法)的攻击生成、并发对抗性训练和模型压缩，以及融合对抗性训练和集成学习的多源防御方案。该项目将在调查人员的实验室中实施和评估针对自主飞行器和无人驾驶飞行器的自主网络物理系统的真实世界原型的拟议攻击和防御方法。调查人员将向GitHub发布所有开发的模型、算法和软件，以促进社区使用。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Cyber-physical systems such as self-driving cars, drones, and intelligent transportation rely heavily on machine learning techniques for ever-increasing levels of autonomy. In the example of autonomous vehicles, deep learning or deep neural networks can be employed for perception, sensor fusion, prediction, planning, and control tasks. However powerful such machine learning techniques have become, they also expose a new attack surface, which may lead to vulnerability to adversarial attacks and potentially harmful consequences in security- and safety-critical scenarios. This project investigates adversarial machine learning challenges faced by autonomous cyber-physical systems with the aim of formulating defense strategies. The project will collaborate with the Center for STEM (Science, Technology, Engineering and Math) Education at Northeastern University and the Office of Access and Inclusion Center at University of California at Irvine to engage undergraduates, women, and minority students in independent research projects.This project is composed of two interdependent research thrusts, one for investigating adversarial attacks and one for devising countermeasures, aiming to secure the key deep learning-equipped software components of autonomous cyber-physical systems, such as perception, obstacle prediction, and vehicle planning and control. The main deep learning techniques of interest to autonomous cyber-physical systems include convolutional neural networks for detection, recurrent neural networks for prediction, and deep reinforcement learning for control. The technical innovations of the project include ADMM (Alternating Direction Method of Multipliers) based attack generation, concurrent adversarial training and model compression, and multi-sourced defense schemes incorporating adversarial training and ensemble learning. This project will implement and evaluate the proposed attack and defense approaches on real-world prototypes of autonomous cyber-physical systems for autonomous vehicles and unmanned aerial vehicles in the investigators' labs. The investigators will release all the developed models, algorithms, and software to GitHub to facilitate community usage.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.