Security of Control Systems
Security of Cyber-Physical Systems from a Control Theoretical Perspective
The concept of Cyber-Physical System (CPS) refers to the embedding of sensing, communication, control and computation into the physical spaces. Today, CPSs can be found in areas as diverse as aerospace, automotive, chemical process control, civil infrastructure, energy, health-care, manufacturing and transportation, most of which are safety critical. Any successful attack on such systems can cause major disruptions, leading to great economic losses and may even endanger human lives. The first-ever CPS malware (called Stuxnet) was found in July 2010 and has raised significant concerns about CPS security. The tight coupling between the cyber and the physical world, as well as the high reliability requirement, propose new challenges for CPS security, requiring rethinking the existing methodology. In this tutorial, we argue that in order to meet these security challenges, a defense-in-depth framework, which combines cyber security and control-theoretic security, can potentially provide better security guarantees for the CPS. We then consider two control theoretical defense mechanisms for CPS: The first one is an active intrusion detection scheme that uses a random control input known as physical watermarking. We show that it can effectively counter the popular reply attack employed by the Stuxnet virus. Then we discuss how to achieve secure and efficient multi-sensor information fusion which enables graceful performance degradation in the presence of integrity attacks on the sensory data.
- introduction on the security challenges of CPS/control systems and a general defense-in-depth framework on how to meet the challenges (15min)
- overview of the active intrusion detection scheme with physical watermarking (15min)
- how to use data driven method to design watermarking signal as well as the intrusion detector when the system model is unknown (15min)
- overview of the secure hypothesis testing problem (10min)
- overview of the secure state estimation problem (10min)
- a discussion on the secure and efficient state estimation for both static and dynamic cases (20min)
- conclusion and discussion on the future directions (5min)
Instructor: Yilin Mo
Associate Professor, Department of Automation, Tsinghua University, China
Yilin Mo received his Ph.D. In Electrical and Computer Engineering from Carnegie Mellon University in 2012 and his Bachelor of Engineering degree from Department of Automation, Tsinghua University in 2007. Prior to his current position, he was a postdoctoral scholar at Carnegie Mellon University in 2013 and California Institute of Technology from 2013 to 2015. He held an assistant professor position in the School of Electrical and Electronic Engineering at Nanyang Technological University from 2015 to 2018. His research on CPS security and networked control systems has been well recognized, receiving more than 6000 google scholar citations.
Homomorphic Encryption and Its Application to Feedback Control
Security threats to cyber-physical systems are increasing daily; thus, the importance of control technology for enabling the safe operation of automation and control systems is growing. Within the field of control engineering, encrypted control, which incorporates cryptographic methods into control systems, is attracting a great deal of attention as a cybersecurity measure for cyber-physical systems.
This session provides a tutorial on homomorphic encryption and its application to feedback control. Our tutorial session is open to every attendee and is targeted at audiences who have no knowledge of the topic. This session is scheduled to be 180 minutes long, and it consists of three speakers working at the forefront of the control system security field. Dr. K. Kogiso, an associate professor at The University of Electro-Communications, will introduce fundamentals of homomorphic encryption and a concept of encrypted control, using comprehensive mathematics with relevance to existing encrypted control studies. K. Teranishi, a Ph.D. student at The University of Electro-Communications, will present comprehensive illustrations and methods to encrypt a linear controller using a partially (multiplicative) homomorphic encryption scheme, introducing an encrypted control library (ECLib) that is a Python library for numerical simulation of encrypted control. Dr. J. Kim, a research fellow at KTH Royal Institute of Technology, will present comprehensive illustrations and methods to encrypt a linear controller using a fully homomorphic encryption scheme together with related topics.
- Introduction of Homomorphic Encryption and Encrypted Control (60 min) by Kiminao Kogiso
- Topics on Partially Homomorphic Encryption Application (30 min) by Kaoru Teranishi
- Encrypted Control Software ECLib (30 min) by Kaoru Teranishi
- Topics on Fully Homomorphic Encryption Application (60 min) by Junsoo Kim
Instructor: Kiminao Kogiso
Associate Professor, Department of Mechanical Engineering and Intelligent Systems, The University of Electro-Communications, Tokyo, Japan
Dr. Kiminao Kogiso received the B.S., M.S., and Ph.D. degrees in Mechanical Engineering from Osaka University, Japan, in 1999, 2001, and 2004, respectively. He was a postdoctoral researcher of the 21st Century COE Program and became an Assistant Professor in the Department of Information Systems, Nara Institute of Science and Technology, Nara, Japan, in 2004 and 2005, respectively. In March 2014, he joined the Department of Mechanical Engineering and Intelligent Systems, The University of Electro-Communications. From November 2010 to December 2011, he was a visiting scholar at the Georgia Institute of Technology, GA, USA. His research interests include constrained control, cyber-security of control systems, control of decision makers, and their applications. He currently serves as an Associate Editor for SICE Journal of Control, Measurement, and System Integration.
Instructor: Kaoru Teranishi
Ph.D. candidate, Department of Mechanical Engineering and Intelligent Systems, The University of Electro-Communications, Tokyo, Japan
Kaoru Teranishi is a Research Fellow of Japan Society for the Promotion of Science. He received a B.S. in Electromechanical Engineering and an associate degree in Electronics and Information Engineering from National Institute of Technology, Ishikawa College. He also obtained an M.S. in Mechanical and Intelligent Systems Engineering from The University of Electro-Communications. From October 2019 to September 2020, he was a visiting scholar of the Georgia Institute of Technology. His research interests include control theory and cryptography for cyber-security of dynamical systems. He is also interested in motion control and industrial application of control engineering.
Instructor: Junsoo Kim
Research fellow, KTH Royal Institute of Technology, Sweden
Dr. Junsoo Kim received his B.S. degree in electrical engineering and mathematical sciences in 2014, and M.S. and Ph.D. degrees in electrical engineering in 2020, from Seoul National University. From 2020 to 2021, he held the post-doc position at Automation and Systems Research Institute, Korea. He is currently a postdoctoral researcher at KTH Royal Institute of Technology, Sweden. His research interests include security problems in networked control systems and encrypted control systems.
Model Predictive Control: Fundamentals and Frontiers
Model Predictive Control (MPC) is a well-established design technique for controlling multivariable systems subject to constraints on manipulated variables and controlled outputs in an optimized way. Following a long history of success in the process industries, MPC has rapidly expanded in several other domains, such as the automotive industry, where it is now used in mass production. This tutorial session aims to provide an in-depth introduction to MPC, from the fundamental concepts to current state-of-the-art methods and new trends in learning algorithms for MPC, focusing on the applicability of the presented ideas to solve control problems in practice. The session will cover the formulation of MPC of linear, linear time-varying, hybrid, stochastic, and nonlinear dynamical systems, and various computational methods that one can use to effectively compute the MPC action in real-time in an embedded system. It will also cover several machine learning approaches for designing and calibrating MPC laws from data to reduce development time, simplify real-time computations, and online adaptation. No previous knowledge of MPC is required to attend the tutorial session.
- Introduction and basic principles of MPC
- Linear MPC
- Observer design and integral action
- Solution methods for embedded linear MPC: quadratic programming solvers, explicit MPC
- Linear time-varying and nonlinear MPC
- Hybrid model predictive control (modeling, MPC formulation, mixed-integer programming)
- Stochastic MPC based on scenarios
- Data-driven linear MPC
- Machine-learning methods for nonlinear and hybrid MPC
- Active-learning methods for automatic and preference-based MPC calibration
Instructor: Alberto Bemporad
Professor, IMT School for Advanced Studies Lucca, Italy
Alberto Bemporad received his Master’s degree in Electrical Engineering in 1993 and his Ph.D. in Control Engineering in 1997 from the University of Florence, Italy. In 1996/97 he was with the Center for Robotics and Automation, Department of Systems Science & Mathematics, Washington University, St. Louis. In 1997-1999 he held a postdoctoral position at the Automatic Control Laboratory, ETH Zurich, Switzerland, where he collaborated as a senior researcher until 2002. In 1999-2009 he was with the Department of Information Engineering of the University of Siena, Italy, becoming an Associate Professor in 2005. In 2010-2011 he was with the Department of Mechanical and Structural Engineering of the University of Trento, Italy. Since 2011 he is Full Professor at the IMT School for Advanced Studies Lucca, Italy, where he served as the Director of the institute in 2012-2015. He spent visiting periods at Stanford University, University of Michigan, and Zhejiang University. In 2011 he co-founded ODYS S.r.l., a company specialized in developing model predictive control systems for industrial production. He has published more than 400 papers in the areas of model predictive control, hybrid systems, optimization, automotive control, and is the co-inventor of 16 patents. He is author or coauthor of various software packages for model predictive control design and implementation, including the Model Predictive Control Toolbox (The Mathworks, Inc.) and the Hybrid Toolbox for MATLAB. He was an Associate Editor of the IEEE Transactions on Automatic Control during 2001-2004 and Chair of the Technical Committee on Hybrid Systems of the IEEE Control Systems Society in 2002-2010. He received the IFAC High-Impact Paper Award for the 2011-14 triennial and the IEEE CSS Transition to Practice Award in 2019. He is an IEEE Fellow since 2010. WEB PAGE: http://imt.lu/ab
Mathematical and practical foundations of backpropagation
Gradient computation via backpropagation is one of the fundamental backbones of the modern deep learning infrastructure. However, although modern deep learning libraries provide a robust implementation of backpropagation, understanding this feature is essential for deep learning scientists. After all, as Andrej Karpathy famously said, backprop is a leaky abstraction, meaning that if you try to ignore how it works under the hood, you will be much less effective at building and debugging neural networks. In this three-part tutorial, we will learn the mathematical foundations of backpropagation and solidify the understanding through basic Python implementations.
- Scalar chain rule for compositions of univariate functions.
- Automatic differentiation as chain rule multiplying scalar derivatives.
- Forward-mode and reverse-mode automatic differentiation and their implementation in Python. (Backprop is reverse-mode automatic differentiation.)
- Review of finite difference and symbolic differentiation and how they differ from automatic differentiation.
- Jacobian derivatives and vector chain rule for compositions of multivariate functions.
- Automatic differentiation as chain rule multiplying Jacobian matrices.
- Optimal matrix chain multiplication via dynamic programming.
- Backprop for multilayer perceptions (MLP).
- Backprop for deep convolutional networks (CNN).
- Python implementation of backprop for MLP.
- Review of directed acyclic graphs (DAGs) and topological sort.
- Review of computation graphs.
- Forward evaluation and backprop via topological sort.
- Python implementation of backprop with DAG.
Instructor: Ernest Ryu
Assistant Professor, Department of Mathematical Sciences, Seoul National University, KR
Ernest Ryu received a B.S. with distinction in physics and electrical engineering at California Institute of Technology in 2010. He received an M.S. in statistics and a Ph.D. with the Gene Golub Best Thesis Award in computational mathematics at Stanford University in 2016. He was at the department of mathematics at the University of California, Los Angeles as an assistant adjunct professor from 2016 to 2019, and joined Seoul National University in 2020. Ernest Ryu’s current research focus is numerical optimization and machine learning.