LAUSR

Abstract In the face of advanced persistent threat actors, existing information technology (IT) defenses as well as some of the more recent operational technology (OT) defenses have been shown to become increasingly vulnerable, especially for critical infrastructure systems with well-established technical know-how. For example, data deception attacks have demonstrated their ability to mislead human operators and statistical detectors alike for a wide range of systems, e.g., electric grid, chemical and nuclear plants, etc. To combat this challenge, our previous work has introduced a new modeling paradigm, called covert cognizance (C2), serving as an active OT defense that allows a critical system to build self-awareness about its past performance, with the awareness parameters covertly embedded into its own state function, precluding the need for additional courier variables. Further, the embedding process employs one-time-pad randomization to blind artificial intelligence (AI)–based learning and ensures zero impact on system state. This paper employs one of the competing AI-based learning algorithms, i.e., the long short-term memory neural network in a supervised learning setting, to validate the C2 embedding process. This is achieved by presenting the network with many labeled samples, distinguishing the original state function from the one containing the embedded self-awareness parameters. A nuclear reactor model is employed for demonstration.