Q-Learning Based Detector Design for State Estimation Under Non-gaussian Noises

This paper addresses the problem of detecting false data injection attacks (FDI) for state estimation with non-Gaussian noises. During the estimation process, potential attacks and non-Gaussian noise can lead to the non-Gaussian property of the innovation, rendering traditional attack detection methods ineffective. To tackle this issue, we propose a novel detection strategy using Kullback-Leibler (KL) divergence as a detection metric, which adapts well to non-Gaussian scenarios. Furthermore, we adopt a Q-learning strategy to train the safety threshold of the detector to improve the reliability of detection. Through verification via Python simulation experiments, we demonstrate that the designed detector has a negligible impact on estimation performance, and provide an effective detection performances against FDI attacks.