EUVIMEDEUVIMED entdecken
Crossref

Marvel: Effective AI-Enabled Cyber-Physical System Falsification Guided by Novelty Detection

Xuan Xie, Lei Ma

IEEE Transactions on Reliability · 2026 · 75 · 2247-2259

DOI: 10.1109/tr.2026.3705058 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Transactions on Reliability, 2026. Verfasst von Xuan Xie und 1 weiteren Autor:innen. Die Arbeit nennt 55 Literaturangaben. Gefördert von: Macau University of Science and Technology Faculty Research, Canada CIFAR AI Chairs Program, Natural Sciences and Engineering Research Council of Canada.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Cyber-physical systems (CPS) integrate digital controllers in physical plants, and have been widely deployed in many safety-critical domains, such as transportation, health care, energy, etc. In recent years, there has been a surge of interest in the application of deep neural networks (DNNs) for the control of CPS, giving birth to the emerging class of AI-enabled CPS. Since the decision logic of DNNs is notoriously hard to interpret, the systems under their control are exposed to great uncertainties and safety risks, which poses an urgent need for effective safety assurance techniques for these systems. In the domain of CPS, falsification is an established safety validation approach that explores the system behavior space in order to search for a counterexample. However, in the case of AI-enabled CPS, the system behavior triggered by DNN controllers is subject to the novelty of the controller inputs, i.e., the system is more error-prone if a controller input is novel w.r.t. the training data. This important feature of AI-enabled CPS is not adopted in the search strategy of the classic falsification, and hence the exploration in the complicated system behavior space lacks effective guidance. Our key insight is that more attention should be paid by falsification to the system inputs that are novel for the DNN controllers. In this article, we propose a framework Marvel for falsification of AI-enabled CPS, which incorporates this insight and adopts the novelty detection of DNN controller inputs as an effective guidance in the loop of falsification. Specifically, we introduce two novelty criteria for the sequential inputs of DNN controllers over time in AI-enabled CPS. We experimentally evaluate the performance of Marvel on two industry-level AI-enabled CPS, namely, an Adaptive Cruise Control system, and an Abstract Fuel Control system. Our results show that Marvel outperforms the state-of-the-art falsification tool, and provides a powerful weapon for the safety assurance of AI-enabled CPS.

Förderung

  • Macau University of Science and Technology Faculty Research
  • Canada CIFAR AI Chairs Program
  • Natural Sciences and Engineering Research Council of Canada
  • JSPS KAKENHI
  • Autoware Foundation

Zitieren

APA

Xie, X., & Ma, L. (2026). Marvel: Effective AI-Enabled Cyber-Physical System Falsification Guided by Novelty Detection. IEEE Transactions on Reliability, 75, 2247-2259. https://doi.org/10.1109/tr.2026.3705058

Vancouver

Xie X, Ma L. Marvel: Effective AI-Enabled Cyber-Physical System Falsification Guided by Novelty Detection. IEEE Transactions on Reliability. 2026;75:2247-2259. doi:10.1109/tr.2026.3705058

BibTeX
@article{Xie2026,
  author = {Xie, Xuan and Ma, Lei},
  title = {Marvel: Effective AI-Enabled Cyber-Physical System Falsification Guided by Novelty Detection},
  journal = {IEEE Transactions on Reliability},
  year = {2026},
  volume = {75},
  pages = {2247--2259},
  doi = {10.1109/tr.2026.3705058}
}
Kostenlos starten →