<link rel="stylesheet" href="styles.cf390c6bfbe7555f.css">
dc.contributor.authorPaudel, Bipin
dc.date.accessioned2026-11-13T14:25:58Z
dc.date.graduationmonthDecember
dc.date.issued2026
dc.description.abstractCyber-physical systems increasingly rely on networked sensing, computation, and super- visory control, creating attack surfaces through which adversaries can manipulate physical processes while remaining undetected. Stealthy attacks are particularly challenging because they can exploit partial observability and system dynamics to alter internal states with- out immediately producing anomalous behavior in observed measurements. Classical zero- dynamics attacks generally assume substantial knowledge of the underlying system model. This dissertation investigates how learning-based adversaries can relax this assumption, how detection mechanisms can adapt to increasingly capable attackers, and how transient oper- ating conditions and privacy-preserving mechanisms influence this interaction. The dissertation first studies model-free generation of zero-dynamics-like attacks using deep reinforcement learning. Using the quadruple-tank process, a Soft Actor-Critic-based adversary is trained to manipulate control inputs and drive unobserved states away from desired operating conditions without access to the plant’s state-space matrices. An iterative attacker-detector framework is then developed in which progressively stronger attackers are trained against existing detectors, while new learning-based detectors are trained using tra- jectories generated by the latest attacker. The results show that reinforcement learning can reproduce important characteristics of theoretically constructed zero-dynamics attacks and can provide challenging attack data for detector development. The framework is next extended to inverter-based power systems. Deep reinforcement learning attackers manipulate control-layer signals of a grid-forming microgrid while su- pervisory detection relies primarily on point-of-common-coupling measurements. Iterative adversarial training is used to construct increasingly robust detectors and to examine attack impact and detection timing. The dissertation then considers attacks during transient rather than steady-state oper- ation. Load-induced transients naturally produce deviations in voltage and frequency that can make malicious behavior more difficult to distinguish from legitimate dynamics. An adaptive adversary is trained to induce mislocalization, in which a disturbance originating from one compromised inverter is attributed to another. The study also compares detector aggregation strategies, showing that strict aggregation can progressively suppress successful attacks, whereas consensus-based aggregation can preserve exploitable ambiguity and allow persistent mislocalization under some conditions. Finally, the dissertation examines a complementary defense that changes the measure- ments available for system identification. Real and surrogate measurements are combined for feedback control and attacker identification, while a separate trusted detector retains real measurements. The mixture weight is selected from disturbance recovery, and the surrogate parameters are optimized to reduce damage before alarm, defined as a physical-limit vio- lation before detection. The optimized design reduces the observed successful attack rate relative to real-only feedback and an unperturbed surrogate, while meeting the aggregate tracking limit and tested recovery requirements. This benefit is obtained at the cost of increased normal-operation tracking error. Together, these studies present an adaptive view of cyber-physical security in which attackers and defenders co-evolve, and in which resilience must be evaluated across model uncertainty, transient operation, detector design, and control requirements.
dc.description.advisorGeorge Amariucai
dc.description.degreeDoctor of Philosophy
dc.description.departmentDepartment of Computer Science
dc.description.levelDoctoral
dc.identifier.urihttps://hdl.handle.net/2097/47504
dc.language.isoen_US
dc.subjectCyber-physical systems security
dc.subjectAdversarial reinforcement learning
dc.subjectStealthy cyberattacks
dc.subjectAttack detection and localization
dc.subjectPrivacy-preserving control
dc.titleLearning-based stealthy attacks and adaptive defenses in cyber-physical systems
dc.typeDissertation

Files

Original bundle

Now showing 1 - 1 of 1
Loading...
Thumbnail Image
Name:
BipinPaudel2026.pdf
Size:
5.59 MB
Format:
Adobe Portable Document Format

License bundle

Now showing 1 - 1 of 1
Loading...
Thumbnail Image
Name:
license.txt
Size:
1.65 KB
Format:
Item-specific license agreed upon to submission
Description: