Date of Award


Degree Type


Degree Name

Doctor of Philosophy (PhD)


Electrical Engineering and Computer Science


Fanxin Kong


attack detection, attack recovery, CPS, cyber-physical systems, security, sensor attack


Cyber-physical systems (CPSs) utilize computation to control physical objects in real-world environments, and an increasing number of CPS-based applications have been designed for life-critical purposes. Sensor attacks, which manipulate sensor readings to deceive CPSs into performing dangerous actions, can result in severe consequences. This urgent need has motivated significant research into reactive defense. In this dissertation, we present an adaptive detection method capable of identifying sensor attacks before the system reaches unsafe states. Once the attacks are detected, a recovery approach that we propose can guide the physical plant to a desired safe state before a safety deadline.Existing detection approaches tend to minimize detection delay and false alarms simultaneously, despite a clear trade-off between these two metrics. We argue that attack detection should dynamically balance these metrics according to the physical system's current state. In line with this argument, we propose an adaptive sensor attack detection system comprising three components: an adaptive detector, a detection deadline estimator, and a data logger. This system can adapt the detection delay and thus false alarms in real-time to meet a varying detection deadline, thereby improving usability. We implement our detection system and validate it using multiple CPS simulators and a reduced-scale autonomous vehicle testbed. After identifying sensor attacks, it is essential to extend the benefits of attack detection. In this dissertation, we investigate how to eliminate the impact of these attacks and propose novel real-time recovery methods for securing CPSs. Initially, we target sensor attack recovery in linear CPSs. By employing formal methods, we are able to reconstruct state estimates and calculate a conservative safety deadline. With these constraints, we formulate the recovery problem as either a linear programming or a quadratic programming problem. By solving this problem, we obtain a recovery control sequence that can smoothly steer a physical system back to a target state set before a safe deadline and maintain the system state within the set once reached. Subsequently, to make recovery practical for complex CPSs, we adapt our recovery method for nonlinear systems and explore the use of uncorrupted sensors to alleviate uncertainty accumulation. Ultimately, we implement our approach and showcase its effectiveness and efficiency through an extensive set of experiments. For linear CPSs, we evaluate the approach using 5 CPS simulators and 3 types of sensor attacks. For nonlinear CPSs, we assess our method on 3 nonlinear benchmarks.


Open Access