Design of Switching Attacks with Physics-Informed Neural Networks

This paper proposes a nonlinear optimal switching attack for Cyber-Physical Systems based on Physics-Informed Neural Networks (PINNs). The methodology employs a two-stage training strategy to identify optimal attack sequences: an attack discovery phase that temporarily relaxes physical constraints to explore divergent system behaviors, and a physics refining phase that enforces dynamical consistency to ensure feasibility. Beyond the standard physics loss, the objective function incorporates an attractivity term to bias the search toward divergent trajectories and a damage term to maximize angular velocity. This formulation enables the network to prioritize attack impact while maintaining physical plausibility. Unlike existing methods limited by linearized models or predefined sliding surfaces, the proposed approach automatically learns high-impact strategies directly from the inherent nonlinear system dynamics.