A One-Class Explainable AI Framework for Identification of Non-Stationary Concurrent False Data Injections in Nuclear Reactor Signals

AI can detect sophisticated cyberattacks on nuclear reactors that traditional security systems miss: The paper demonstrates that conventional linear detection methods fail to catch false data injections designed to bypass them, but the proposed AI framework successfully identifies these "replay attacks" where attackers replay recorded sensor data to mask malicious changes.

Explainability is critical for nuclear safety applications: The framework uses SHAP (a technique that explains AI decisions) and rule-based correlations to make its threat detection transparent and interpretable, which is essential for nuclear operators who need to understand why the system flagged a potential attack before taking action.

AI trained only on normal operations can protect against unknown threats: Rather than requiring extensive datasets of past attacks (which are rare in nuclear facilities), the system trains exclusively on normal reactor behavior, then flags deviations—making it practical for real-world deployment where attack data is limited.

Real-world nuclear data validation matters: The framework was tested on actual data from Purdue's nuclear reactor facility, not just simulations, which strengthens confidence that this approach could work in protecting real advanced reactor systems as they transition to digital controls.

Handling noisy, complex sensor data requires specialized techniques: The paper introduces an "adaptive windowing" method for residual analysis to improve accuracy despite the inherent noise and nonlinear behavior of multi-sensor nuclear systems, addressing a practical engineering challenge in real deployments.