Stealth FDI Attack Detection Method Based on Adversarial Spatial-Temporal Attention LSTM_DAE for ICPS

With the high degree of integration between industrial physical dynamic processes and cyberspace, the potential risks posed by various cyber-attacks on Industrial Cyber-Physical Systems (ICPS) have become increasingly prominent. Existing attack detection methods have not fully considered the spatial-temporal correlations and high stealthiness of False Data Injection (FDI) attacks, nor the impact of various types of noise on attack detection. As a result, it is difficult to effectively detect highly stealthy FDI attacks. This paper proposes an unsupervised attack detection method based on adversarial spatial-temporal attention LSTM-DAE. First, a spatial-temporal attention autoencoder is constructed to extract the spatial-temporal dependencies in the multivariate long time series and assign weights according to the importance of the data during training, effectively extracting key information. Second, to mitigate the impact of various noises, such as environmental noise, sensor noise, and system operation noise, on the accuracy of attack detection in ICPS, this paper enhances the model's robustness to noise by using a denoising autoencoder (DAE). Finally, to prevent highly stealthy FDI attacks from being missed due to small reconstruction error, an adversarial training method is adopted to amplify the reconstruction error of anomalous data, ensuring the model's sensitivity to attack detection. By establishing an anomaly scoring, this paper achieves effective detection of stealthy FDI attacks and conducts experimental analysis on the SWAT and SKAB datasets to validate the effectiveness of the proposed model.