Tamper-resistant controller using neural network and time-varying quantization

In this paper, we consider a tamper-resistant control system aiming at protecting the knowledge of the controller from attackers. In this control system, the controller operates normally only for a limited number of time-varying specific states; otherwise, it outputs an incorrect value. We propose to realize the tamper-resistant controller by employing a neural network and time-varying quantization. Furthermore, we make it possible for only one trained neural network to be used for all quantization based on the local approximation linearity of the trained neural network. Without this approach, the neural network needs to be trained for every possible quantization, which leads to huge computation. We provide simulations to demonstrate the security and feasibility of the proposed method.