CAREER: Towards Non-Conservative Learning-Aided Robustness for Cyber-Physical Safety and Security

该项目的目标是通过开发控制合成工具，为理解和利用物理系统、人工智能/网络人类代理及其环境之间的相互作用提供科学基础，以推断现实世界不确定因素下的安全和安保。这类网络物理系统包括许多支撑现代社会的重要基础设施(例如交通系统、电力分配)，通常是安全关键的。如果受到损害，可能会对受控制的物理实体和操作或使用这些实体的人造成严重损害，并造成重大经济损失。然而，真实系统和系统的不完美模型之间的模型失配，再加上其他不确定因素(例如，测量误差)，使现有的安全和安全保护失效，而没有学习的健壮解决方案可能过于保守。这些挑战表明，需要设计新的计算工具，在不牺牲性能的情况下，确保真实世界不确定环境下网络物理系统的稳健安全。该项目包括与教育和推广相结合的研究活动，以使学生和行业合作伙伴认识到安全和安保对与计算相关的技术的重要性。为了使网络物理系统能够实现学习辅助的稳健安全和安保，该项目将开发数学基础和控制综合算法，基于集合成员和学习方法，用于不确定性量化、安全/抗攻击估计和设计安全控制。这项研究工作将产生新的科学基础，代表：1)从基于机器学习和/或物理的模型的不确定性的传统平均或随机表征转变为使用混合包含的集成员表示，2)从安全点估计器设计到具有中间人攻击模型/策略的运行时学习的安全集成员估计器的转变，以及3)从具有不妥协的状态反馈的固定按设计安全的控制算法到通过从运行时数据学习的攻击弹性输出反馈设计的进步。总而言之，这些贡献为网络-物理安全和安保的学习辅助控制综合奠定了基础，为广泛的网络-物理系统实现了非保守的安全和可靠的解决方案，包括主要应用于用于驾驶研究计划的自动驾驶汽车。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。

The goal of this project is to provide a scientific basis to understand and leverage the interaction among physical systems, artificial intelligence/cyber-human agents and their environment through the development of control synthesis tools to reason about safety and security under real-world uncertainties. Such cyber-physical systems, which include many vital infrastructures that sustain modern society (e.g., transportation systems, electric power distribution) are usually safety-critical. If compromised, serious harm to the controlled physical entities and the people operating or utilizing them as well as significant economic losses can result. However, model mismatches between the real system and an imperfect model of the system, in addition to other sources of uncertainties (e.g., measurement errors) disable existing safety and security protection, while robust solutions without learning may be overly conservative. These challenges demonstrate the need to design novel computational tools that can guarantee robust safety and security of cyber-physical systems under real-world uncertainties without sacrificing performance. The project includes research activities that are integrated with education and outreach to engage students and industry partners to appreciate the importance of safety and security for computing-related technologies.To enable learning-aided robust safety and security for cyber-physical systems, this project will develop mathematical foundations and control synthesis algorithms based on set-membership and learning approaches for uncertainty quantification, secure/attack-resilient estimation and safe-by-design control. The research endeavor will produce novel scientific foundations representing: 1) a shift from the conventional average or stochastic characterization of uncertainty of machine learning- and/or physics-based models to a set-membership representation using hybrid inclusion, 2) a transition from secure point estimator designs to secure set-membership estimators with run-time learning of man-in-the-middle attack models/strategies, and 3) a progression from fixed safe-by-design control algorithms with uncompromised state feedback to attack-resilient output feedback designs with learning from run-time data. Together, these contributions lay the foundations in learning-aided control synthesis for cyber-physical safety and security, enabling non-conservative safe and secure solutions for a broad range of cyber-physical systems, including the main application to self-driving cars used to drive the research program.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.