CAREER: Securing the AI Stack in Autonomous CPS under Physical-Layer Attacks: A Systems Perspective

近年来，自动驾驶汽车和送货无人机/机器人等自主网络物理系统在现实世界中的开发和部署出现了巨大的激增。为了在复杂环境中实现高度自主，人工智能(AI)堆栈发挥着核心作用，这是一种“大脑”，这使它们具有高度的安全关键型。以前的工作研究了用于自主网络物理系统中的针对AI算法的对抗性攻击，但大多集中在完全或部分孤立于物理上下文的AI算法级别的安全特性。然而，由于这些算法只是整个系统的组成部分，从系统的角度研究和解决它们的安全问题更有实际意义和有效性，特别是在更一般和更基本的物理层攻击模型下。该项目旨在创建一套系统的方法论、解决方案框架和平台，以实现物理层攻击下自治网络物理系统的AI组件的系统级安全分析和防御设计。随着自主网络物理系统在现实世界中的不断部署和商业化，这方面的成功将直接惠及日常生活的安全。该项目包括两个研究推力，涵盖拟议的系统级安全研究的攻击和防御两个方面。首先，为了实现系统级的安全分析，该项目将开发新的系统到人工智能和人工智能到系统映射方法，通过克服各种设计挑战，例如在物理层攻击生成中系统地维护物理可实现性和语义等价性，以及有效地适应现实世界系统设计和实现的多样性。其次，为了开发系统级防御设计，该项目将从单个系统和作战生态系统的角度系统地识别和利用新的设计机会，包括新的物理不变量类别、利用系统级特性的新型攻击弹性传感器融合设计，以及利用作战生态系统和基础设施支持中的其他参与者的新型设计。该项目还将开发一个基于模拟的评估平台，提供统一和可扩展的攻击和防御开发支持，用于促进自主网络物理系统安全的研究和教育。该奖项反映了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。

Recent years have witnessed a massive surge in real-world development and deployment of autonomous cyber-physical systems such as autonomous driving cars and delivery drones/robots. To achieve high-level autonomy in complex environments, the Artificial Intelligence (AI) stack plays a central role, a type of “brain,” which makes them highly security-critical. Prior works have studied adversarial attacks against AI algorithms used in autonomous cyber-physical systems, but mostly focus on the AI algorithm-level security properties in complete or partial isolation of the physical context. As these algorithms are only components of the entire system, however, it is both more practically meaningful and effective to study and address their security problems from a systems perspective, especially when under the more general and fundamental physical-layer attack model. This project aims to create a suite of systematic methodologies, solution frameworks, and platforms that can achieve system-level security analysis and defense designs for the AI component of autonomous cyber-physical systems under physical-layer attacks. With the growing deployment and commercialization of autonomous cyber-physical systems in the real world, success in this should directly benefit the safety of everyday lives.This project consists of two research thrusts to cover both the attack and defense sides of the proposed system-level security research. First, to enable system-level security analysis, this project will develop novel system-to-AI and AI-to-system mapping methodologies, by overcoming various design challenges such as systematically maintaining physical realizability and semantic equivalency in physical-layer attack generation, and effectively accommodating the diversity of real-world system designs and implementations. Second, to develop system-level defense designs, this project will systematically identify and leverage novel design opportunities from both individual system and the operation ecosystem perspectives, including new classes of physical invariants, novel attack-resilient sensor fusion designs leveraging system-level properties, and novel designs that leverage other participants in the operation ecosystem and infrastructure support. This project will also develop a simulation-based evaluation platform with uniform and extensible attack and defense development support, which will be used to facilitate both research and education of autonomous cyber-physical system security.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.