Sensor Attack Detection and Mitigation Using Hybrid Physics Informed Neural Networks - A Real-Time Implementation for DC-DC Converter

This article introduces a novel approach to enhance the cybersecurity of cyber-physical systems (CPS) by integrating a Hybrid Physics Informed Neural Network (HPINN) into the control framework of a DC-DC converter. The proposed system leverages the strengths of both physics-based modeling and data-driven techniques to detect and mitigate false data injection attacks (FDIA) affecting the converter's performance. This article proposes a novel HPINN based methodology for FDIA attack detection and mitigation. The proposed methodology includes the design of PINN based attack detection module and PINN prediction module, the combination of these two modules ensures robust control during system uncertainties and reliable operation during FDIA scenarios. PINN detection module is a combination of state estimator and NN classifier. The article presents the architecture of the hybrid system and details its training process, emphasizing the integration of physical principles to enhance the model's interpretability and generalization. Simulation is performed in MATLAB Simulink 2023b; Simulation results demonstrate the effectiveness of the proposed approach in detecting and mitigating FDIA, including those attempting to manipulate the converter's output or compromise its stability. Later, real-time implementation of the proposed method is performed by developing experimental hardware and deploying the HPINN model into the Atmel SAM3X8E ARM Cortex-M3 microcontroller. The proposed scheme is analyzed for various scenarios like load changes, constant FDIA, and time-varying FDIA and its results are explored.