Encrypted machine learning-based model predictive control architectures for nonlinear systems

This work proposes the implementation of encryption in model predictive control of nonlinear systems in which the system dynamics are modeled through machine-learning, denoted ML-based MPC, as a means to improve cybersecurity without significant performance losses. The Pallier cryptosystem is utilized for encryption and the closed-loop stability of the encrypted ML-based MPC is established accounting for the impacts of signal quantization loss due to encryption and sample-and-hold control. A nonlinear chemical process example is used to study the impact of different encryption levels on ML-based MPC closed-loop performance. Finally, we present the implementation of the encrypted ML-based MPC method in a two-layer economic model predictive control framework and in a distributed model predictive control scheme to optimize economic performance and control large-scale processes, respectively.