CAREER: Characterizing Attack Resilience of Multi-agent Dynamical Systems with Applications to Connected Autonomous Vehicles

目前，无人驾驶汽车、无人驾驶巴士、无人驾驶飞机等无人驾驶汽车等新型交通技术正在蓬勃发展。这些技术有望带来诸多好处，比如减少道路死亡和拥堵，以及提高最后一英里的配送效率。然而，实现自动驾驶的竞赛可能会使自动驾驶汽车的传感器、通信和控制信号越来越容易受到攻击。该CAREER项目将创建一个综合研究和教育计划，重点关注包括自动驾驶汽车在内的移动动力系统（或代理）的迫切需求，确保它们即使在感知、通信和控制攻击下也能实现预期目标。有三个核心研究目标。第一个涉及设计一个基于博弈论的新颖建模和分析框架，该框架考虑了当单个代理在对抗环境中操作时固有的不对称性。第二是扩展框架，包括多个代理之间的合作和学习的使用，以共同减少攻击的影响。第三个是解决在有限的板载计算下解决最终游戏时出现的可扩展性挑战。所提出的方法将通过使用地面机器人试验台和全尺寸自动驾驶汽车的场景进行现实仿真进行评估。教育计划包括创建一个基于与攻击弹性相关的创新游戏的活动池，将工程和艺术结合起来，有效地促进学生参与和人员培训。向公众和执法机构进行的报告将有助于塑造安全事件的根本原因分析和阻止攻击的未来做法。该项目将产生一个整体框架来研究多智能体动态系统中出现的安全问题的数学基础。与专注于安全问题的特定方面的现有框架相比，该框架将是一个重大进步。该项目将推进博弈论和计算的多个子领域。首先，它将结合多代理安全游戏和学习技术，为利用环境不确定性的攻击提供一种新颖的解决方案。其次，这项工作将建立在随机算法的进展上，以解决多智能体博弈中出现的大型优化问题。该方法将为鲁棒多准则优化和多智能体学习提供一个新的概率范式。它将描述系统的攻击弹性和所涉及的计算之间的权衡。第三，包括自动驾驶汽车的真实移动测试平台在内的分层评估将促进该框架向现实世界应用的过渡。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。

Our society is currently witnessing a surge in new transportation technologies that include autonomous vehicles such as self-driving cars, automated buses, and drones. These technologies promise several benefits such as reduced roadway fatalities and congestion, and efficient last-mile delivery. However, the race to achieve autonomy can make autonomous vehicles increasingly susceptible to attacks on their sensors, communications, and control signals. This CAREER project will create an integrated research and education program focusing on the pressing need to make mobile dynamical systems (or agents) including autonomous vehicles resilient – guaranteeing that they meet their intended objectives, even under sensing, communication, and control attacks. There are three core research objectives. The first involves designing a novel modeling and analysis framework based on game theory that factors in the inherent asymmetry when a single agent operates in adversarial environments. The second is to extend the framework to include cooperation and the use of learning between multiple agents to collectively minimize the impact of attacks. The third is to address the scalability challenge that arises in solving the resulting games with limited onboard computation. The proposed methodology will be evaluated through realistic emulations using a ground robotic testbed and scenarios with full-sized autonomous vehicles. The educational plan includes creating a pool of activities based on innovative games related to attack resilience that will combine engineering and the arts for effective student engagement and personnel training. Presentations to the public and law enforcement agencies will help shape future practices for root-cause analyses of security incidents and deter attacks.This project will result in a holistic framework to study the mathematical underpinnings of security problems arising in multi-agent dynamical systems. This framework will be a major advancement compared to existing frameworks that focus on specific aspects of the security problem. The project will advance multiple sub-fields in game theory and computation. First, it will combine multi-agent security games and learning techniques to provide a novel solution to attacks that leverage environmental uncertainty. Second, this work will build on the progress in randomized algorithms to solve large optimization problems arising in multi-agent games. The approach will enable a new probabilistic paradigm for robust multi-criterion optimization and multi-agent learning. It will characterize the tradeoff between the system’s attack resilience and the computation involved. Third, layered evaluations that include a real mobility testbed with autonomous vehicles will facilitate the transition of this framework to real-world applications.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.