Abstract:
Encrypted computation opens up promising avenues across a plethora of application domains, including machine learning, healthcare, finance, and control. Arithmetic homomorphic encryption, in particular, is a natural fit for cloud-based computational services. However, computations are essentially limited to polynomial circuits, while comparisons, transcendental functions, and iterative algorithms are notoriously hard to realize.
Against this background, this talk presents a reliable algorithm for calculating the general inverse of an encrypted matrix. More precisely, we exploit encryption-friendly iterative algorithms for matrix inversions and present reliable initializations as well as certificates for the achieved accuracy without compromising the privacy of provided I/O-data. Clearly, being able to realize encrypted matrix inversions opens the door for encrypted solutions of least-squares problems. In this context, we illustrate the effectiveness of our approach by implementing an encrypted system identification as-a-service. Furthermore, we briefly discuss future usage of our method for more complex encrypted optimization problems and applications.
Abstract:
Encrypted dynamic controllers that operate for an unlimited time have been a challenging research subject. The fundamental difficulty is the accumulation of errors and scaling factors in the internal state during operation. Bootstrapping, a technique commonly employed in fully homomorphic cryptosystems, can be used to avoid overflows in the controller state but can potentially introduce significant numerical errors. In this talk, we analyze dynamic encrypted control with explicit consideration of bootstrapping. By recognizing the bootstrapping errors occurring in the controller’s state as an uncertainty in the robust control framework, we can provide stability and performance guarantees for the whole encrypted control system. Further, the conservatism of the stability and performance test is reduced by using a lifted version of the control system.
Abstract: While the integration of information technology and physical infrastructure systems have already made smart infrastructure systems a reality, the advent of AI has the promise of making such systems even more responsive and efficient. However, making this promise a reality requires solution of multiple challenges. In this two-part talk, I will present my work on two of these challenges. In the first part, leveraging concepts from bi-similarity frameworks, I will present a principled framework towards designing, in a compositional and robust manner, approximate but computationally more efficient models of the physical system that can be used in control design and contingency planning. In the second part, I will present a framework to design robust learning algorithms for multiple non-cooperative AI agents. Both frameworks that I will present are general in the sense that they can also be used with preexisting methods.
Abstract: In this talk, I will introduce a hypergame framework to analyze the leader-follower Stackelberg security game with typical misinformation like misperception and deception. A stability criterion will be proposed with the help of the concept of Hyper Nash equilibrium to investigate both strategic stability and cognitive stability of equilibria. The robustness of equilibria will also be discussed to reveal whether players can keep profits under the perturbation of misinformation. Furthermore, I will extend the methodology to players’ different strategy preferences by a hyper-Bayesian game framework and show the corresponding core concept, the Hyper Bayesian Nash equilibrium. The equilibrium stability of both strategies and cognitions, and the equilibrium robustness under perturbed information, will be investigated as well under these circumstances.
Abstract: Networked Control Systems (NCSs) are integral to many safety-critical infrastructures, including power grids, transportation networks, and industrial automation systems. Ensuring their resilient operation in the presence of cyber-attacks is crucial to maintaining societal and economic stability. A systematic approach to this challenge is provided by the risk management framework, which consists of two key components: risk assessment - the quantitative evaluation of potential threats and their impacts, and risk mitigation, which involves designing effective countermeasures such as resilient controllers, optimal security resource allocation, and robust attack detection mechanisms.
This talk presents recent advances in applying the risk management framework to the analysis and design of cyber-resilient NCSs. It highlights how risk-based approaches can guide security decisions under uncertainty and constrained resources. In particular, the talk will also address the scalability challenges of existing risk analysis methods when applied to large-scale or distributed NCSs and introduce novel, computationally efficient techniques that enable risk evaluation and mitigation at scale.