An Unsupervised Deep Explainable AI Framework for Localization of Concurrent Replay Attacks in Nuclear Reactor Signals

Nuclear reactors are becoming more digitally dependent, creating new cybersecurity vulnerabilities: Next-generation advanced reactors rely heavily on fully digital instrumentation and control systems that generate continuous streams of data, making them potentially vulnerable to cyber attacks that could compromise safety and reliability.

Current detection methods are insufficient for real-world nuclear environments: Existing approaches to detecting replay attacks (where attackers reuse previously recorded data to deceive systems) rely on synthetic data or supervised learning that may not capture the complexity of actual reactor operations, and they typically cannot explain why an attack was detected or identify its source.

The framework uses explainable AI to go beyond simple detection: Rather than just flagging that something is wrong, this approach combines machine learning with interpretability tools (autoencoders and SHAP algorithms) to identify not only that an attack occurred, but also when it happened, what type of attack it was, and where it originated in the system.

Explainability is critical for regulated industries: In nuclear facilities, regulators and operators need to understand not just that an AI system detected a threat, but the specific reasoning behind that detection—this framework addresses that requirement by providing transparent, characterizable results rather than "black box" predictions.

Real-world data validation matters for safety-critical systems: By testing on actual reactor signals rather than only synthetic data, this work demonstrates that AI security solutions can be practically deployed in nuclear environments where failures could have serious consequences.