LSTM-AE-Based Control Signal Protection and Cyber Attack Detection

This paper presents a point-to-point encryption framework for nonlinear cyber-physical systems, integrating data protection, resilient control, and attack detection. A long short-term memory autoencoder (LSTM-AE) is employed to protect control signals against cyber attacks. The LSTM-AE model is trained offline and deployed online. When system dynamic models are unknown, adversarial training is adopted to preserve reconstruction accuracy and reduce the impact of attacks for control. When model knowledge is available, it is embedded into the training process. For attack detection, a dual-decoder architecture is proposed, where the discrepancy between decoder outputs serves as the detection residual. The proposed approach is more lightweight than conventional encryption schemes, works for secure control and anomaly detection simultaneously, and is well applicable for nonlinear systems. Simulation results on a three-tank system demonstrate the effectiveness of the proposed scheme.