ERI: Operator-Automation Shared Protection for Security and Safety Assured Industrial Control Systems: Learning, Detection, and Recovery Control

工业控制系统(ICSS)通常用于关键基础设施，包括电力、水处理和分配以及交通运输。然而，ICSS日益数字化，涉及传感、通信和控制，带来了先进的功能，但也暴露了易受恶意网络攻击的漏洞。保护ICSS免受此类攻击至关重要，因为它们可能造成灾难性的物理损害。该项目旨在开发一种全面的解决方案，将人对环可解释的机器学习(ML)、检测和恢复控制集成到操作员自动化共享保护框架中，以提供针对恶意网络攻击的安全可靠的ICSS。此外，该项目将纳入工程研究和教育，以培训学生，特别是那些来自代表性不足的少数群体(URM)的学生，并培养一支多元化的、具有全球竞争力的网络安全劳动力队伍。该项目旨在降低ICSS安全研究和教育的门槛，旨在通过为不同群体的初级和专业网络安全学生和工程师提供可访问的试验台来学习和实践，从而产生重大影响。用于检测网络攻击的进程异常检测的基本概念涉及基于物理不变量比较观察到的和预期的行为。数据驱动的方法具有无需领域专业知识即可自动发现这些物理不变量的优势。然而，现有的基于黑盒机器学习(ML)的方法往往忽略了系统操作员在安全关键ICSS中的作用。检测结果中缺乏有洞察力的解释阻碍了系统操作员进行故障排除和隔离受到攻击的异常传感器和执行器，这是计划在线恢复所必需的。为了解决这个问题，PI的团队建议开发一个操作员自动化共享保护框架，将人在回路上可解释的ML、检测和恢复控制统一起来。该框架旨在使用尖端ML和控制技术进行实时决策，同时评估操作员的反馈，以防止过度信任安全关键系统中的自主性。该研究项目有三个主要目标：1)PI团队将开发有洞察力的混合自动机学习，以捕获物理不变量，使系统操作员能够理解模型和检测结果，验证和修正模型，并定位异常；2)将设计一种实时可证明的安全控制，在不违反安全约束的情况下，将系统恢复到正常运行；3)PI的团队将在3D模拟和真实测试床上评估、演示和传播拟议框架的最佳实践，这些测试床具有广泛的可移植性，以降低ICSS安全研究和教育的障碍。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。

Industrial control systems (ICSs) are commonly utilized in critical infrastructures, including power, water treatment and distribution, and transportation. However, the increasing digitization of ICSs, involving sensing, communication, and control, brings advanced features but also exposes vulnerabilities to malicious cyber-attacks. Protecting ICSs from such attacks is crucial due to the potential catastrophic physical damages they can cause. The project aims to develop a comprehensive solution, integrating human-on-the-loop explainable machine learning (ML), detection, and recovery control in an operator-automation shared protection framework, to provide security and safety-assured ICSs against malicious cyber-attacks. Moreover, the project will incorporate engineering research and education to train students, particularly those from Under-Represented Minorities (URM), and cultivate a diverse, globally competitive cybersecurity workforce. With the goal of lowering barriers to ICSs security research and education, this project aims to have a significant impact by providing accessible testbeds for a diverse population of beginning and expert cybersecurity students and engineers to learn and practice.The underlying concept of process anomaly detection, which is used for detecting cyber-attacks, involves comparing observed and expected behaviors based on physical invariants. The data-driven approach has the advantage of automatically discovering these physical invariants without requiring domain expertise. However, existing approaches based on black-box Machine Learning (ML) often overlook the role of system operators in safety-critical ICSs. The lack of insightful explanations in detection results hinders system operators from conducting troubleshooting and isolating anomalous sensors and actuators under attack, which is necessary for scheduling online recovery. To address this issue, the PI's team proposes to develop an operator-automation shared protection framework that unifies human-on-the-loop explainable ML, detection, and recovery control. This framework aims to enable real-time decision-making using cutting-edge ML and control techniques while valuing the feedback of human operators to prevent over-trust in autonomy in a safety-critical system. The research project has three major objectives: 1) The PI's team will develop insightful hybrid automata learning that captures physical invariants in a way that system operators can understand the model and the detection results, verify and correct the model, and localize anomalies; 2) A real-time provably safe control under uncertainty will be designed to restore the system to normal operation without violating safety constraints; and 3) the PI’s team will evaluate, demonstrate, and disseminate best practices of the proposed framework on 3D simulated and real testbeds with portability across a wide range of ICSs for lowering the barriers to ICSs security research and education.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.