Secure detection and control in Cyber-physical systems

A Cyber-physical system (CPS) is a complex system embedding advanced computation, communication and control techniques into physical spaces, and is usually built up with a set of networked agents such as sensors, actuators, control processing units, and communication devices. Although the development of CPSfacilitates efficient and real-time collaboration between elements, the open nature of communication networks makes it rather vulnerable to malicious attacks. Given that the applications of CPSs vary from aerospace, transportation, power grids, which are always safety-critical, researchers have acknowledged the importance of designing the system with secure algorithms. This thesis first characterizes the properties required for a secure system and possible security threats. Driven by the concerns of deception attacks on communication channels, we study secure detection and control in adversarial environment. New designs on the detection and control algorithms will be developed in this thesis, providing acceptable system performance in the presence of attacks. Chapter 3 investigates the binary hypothesis testing in adversarial environment, where a detector determines the true state of an unknown parameter using m sensors. Among these sensors, n out of them can be compromised by the adversary and send arbitrary data. The exponential rate, at which the worst-case probability of detection error goes to 0, is adopted to depict the system performance. This problem is then formulated as a game between detector and attacker, where the former player attempts to maximize this rate and the latter intends to minimize it. We study both cases where m > 2n and m<=2n, and obtain an equilibrium strategy pair of detection rule and attack scheme for both cases. Inspired by that the unreliable data transmission can degrade the performance of traditional control algorithms, Chapter 4 discusses the resilient consensus in multi-agent systems, where some of the agents might be misbehaving. Speci fically, a continuous-time second-order system is considered, where the agent's dynamics are governed by both position and velocity states. To avoid continuous communication and control, we propose an impulsive secure algorithm. Based on this strategy, signal transmissions and control actions only occur at (aperiodic) sampling instants. After creating a "safe region" with the position states from neighbors, each benign agent derives its control signal with a value inside this region. Sufficient conditions related to the network topology and the maximum number of tolerable faulty nodes are finally derived. As a result, the position states of benign agents are asymptotically synchronized, and the velocity states converge to 0. Chapter 5 also studies the problem of resilient consensus in multi-agent systems. At this time, we intend to propose secure algorithms which not only facilitate the agreement among benign agents, but also guarantee that the agreement is within the convex hull formed by benign agents' initial states. Towards this end, a resilient consensus algorithm is given, where at each time, the normal agent sorts its received values on one dimension, computes two "middle points" based on the sorted values, and moves its state toward these middle points. An explicit approach is further given for the computation of middle points through linear programming. Compared with the existing works, our approach is applicable to general multi-dimensional systems and introduces lower computational complexity. As the consensus among agents arguably forms the basis of distributed computing, the aforementioned results represent a fi rst step towards the development of secure coordination protocols. Chapter 6 focuses on another important application of multi-agent systems, namely the resilient containment control in the presence of multiple leaders. Both the leaders and followers can be malicious. In contrast to the leaderless consensus, the objective of this problem is not to achieve an agreement, but to drive the normal followers to the convex hull formed by normal leaders. To this aim, we design secure protocols for both the first-order and second-order systems. Through convex analysis and Lyapunov functions, convergence and resiliency of the proposed algorithms is theoretically proved. In summary, this thesis considers the secure detection and control in the presence of deception attacks. All of the proposed approaches are well supported by numerical examples, besides theoretical analysis.Doctor of Philosoph