Online Adversarial Learning of Reactor State

This paper is in support of our recent efforts to designing intelligent defenses against false data injection attacks, where false data are injected in the raw data used to control the reactor. Adopting a game-model between the attacker and the defender, we focus here on how the attacker may estimate reactor state in order to inject an attack that can bypass normal reactor anomaly and outlier detection checks. This approach is essential to designing defensive strategies that can anticipate the attackers moves. More importantly, it is to alert the community that defensive methods based on approximate physics models could be bypassed by the attacker who can approximate the models in an online mode during a lie-in-wait period. For illustration, we employ a simplified point kinetics model and show how an attacker, once gaining access to the reactor raw data, i.e., instrumentation readings, can inject small perturbations to learn the reactor dynamic behavior. In our context, this equates to estimating the reactivity feedback coefficients, e.g., Doppler, Xenon poisoning, etc. We employ a non-parametric learning approach that employs alternating conditional estimation in conjunction with discrete Fourier transform and curve fitting techniques to estimate reactivity coefficients. An Iranian model of the Bushehr reactor is employed for demonstration. Results indicate that very accurate estimation of reactor state could be achieved using the proposed learning method.