(702d) Cyberattack Detection and Handling for Neural Network-Approximated Economic Model Predictive Control

Cyberattacks on control systems can create unprofitable and unsafe operating conditions. A common approach to enhance safety and attack resiliency of control systems is to develop attack detection mechanisms. In prior work in our group, a number of cyberattack detection strategies were elaborated for chemical processes integrated with an advanced control formulation known as Lyapunov-based economic model predictive control (LEMPC) [1], in the sense that the controller properties can be used to analyze closed-loop stability in the presence or absence of undetected attacks [2,3]. This raises two questions of whether these strategies are more conservative than alternative detection methodologies such as those based on reachability analysis for attack detection (e.g., [4], [5]) and also to what extent integrating detection strategies with LEMPC alone is useful.

In this work, we move toward addressing these questions [7]. First, we demonstrate that the LEMPC-based cyberattack detection strategies can be modified to enable extension to a neural network (NN) version of the controller (i.e., using a NN as a function approximator [6]). This NN then approximates the LEMPC (which will be abbreviated as NN-LEMPC). We provide sufficient conditions for an LEMPC which ensure that if an NN-LEMPC is then designed based on data from that LEMPC, the NN-LEMPC will inherit safety properties from the original LEMPC formulation. We focus on this for one of the passive detection strategies for LEMPC from [2], in which a threshold on the difference between state predictions and estimates is used to detect attacks. Then, we discuss a second detection and handling strategy based on reachability analysis and input repair (inspired by the concept of repairing a NN in [8]) for ensuring safety of controllers (e.g., EMPC) that are not based on a rigorous control law formulation like LEMPC (in the sense of boundedness of the closed-loop state in a safe operating region) for a sampling period after an undetected cyberattack. We use the same detection metric related to state predictions and estimates to facilitate a comparison between the LEMPC-based methods and the reachable set and input repair technique. This enables use to examine the potential conservatism differences between the LEMPC-based safety strategy and the detection and handling concept based on repairing problematic control actions that correspond to false state measurements due to a sensor attack, and faciliates discussion regarding how this concept can inspire ideas for fighting back against attacks.

References:

[1] Heidarinejad, Mohsen, Jinfeng Liu, and Panagiotis D. Christofides. "Economic model predictive control of nonlinear process systems using Lyapunov techniques." AIChE Journal 58, no. 3 (2012): 855-870.

[2] Oyama, Henrique, and Helen Durand. "Integrated cyberattack detection and resilient control strategies using Lyapunov‐based economic model predictive control." AIChE Journal 66, no. 12 (2020): e17084.

[3] Oyama, Henrique, Dominic Messina, Keshav Kasturi Rangan, and Helen Durand. "Lyapunov-based economic model predictive control for detecting and handling actuator and simultaneous sensor/actuator cyberattacks on process control systems." Frontiers in Chemical Engineering 4 (2022): 810129.

[4] Liu, Hao, Ben Niu, and Jiahu Qin. "Reachability analysis for linear discrete-time systems under stealthy cyber attacks." IEEE Transactions on Automatic Control 66, no. 9 (2021): 4444-4451.

[5] Narasimhan, S., El-Farra, N.H. and Ellis, M.J., 2023. A reachable set-based scheme for the detection of false data injection cyberattacks on dynamic processes. Digital Chemical Engineering, 7, p.100100.

[6] Lucia, Sergio, and Benjamin Karg. "A deep learning-based approach to robust nonlinear model predictive control." IFAC-PapersOnLine 51, no. 20 (2018): 511-516.

[7] Abou Halloun, Jihan and Helen Durand. "Cyberattack Detection and Handling for Neural Network-Approximated Economic Model Predictive Control." ADCHEM 2024 (2024), accepted.

[8] Yang, Xiaodong, Tom Yamaguchi, Hoang-Dung Tran, Bardh Hoxha, Taylor T. Johnson, and Danil Prokhorov. "Neural network repair with reachability analysis." In International Conference on Formal Modeling and Analysis of Timed Systems, pp. 221-236. Cham: Springer International Publishing, 2022.