An Attack-Mechanism-Informed Informer for Sensor Deception Attack Detection in Process Industries

Stealthy sensor deception attacks in process industries are difficult to detect because they often exhibit weak anomalies, strong persistence, and can be obscured by normal process fluctuations. To address these challenges, this study proposes an attack-mechanism-informed Informer (AMI-Informer) detection framework that uses an Informer encoder as the detection backbone and incorporates an attack-mechanism-informed loss constructed from four mechanism-informed features into joint optimization with the data-driven loss. Experiments were conducted on the Tennessee Eastman Process benchmark, with reactor pressure under the min–max attack as the primary scenario, and were further extended to different attacked variables, surge attacks, and an additional Secure Water Treatment (SWaT) benchmark under the LIT101 min–max attack. In the XMEAS7 min–max attack scenario, AMI-Informer achieved an accuracy of 0.9782, a recall of 0.9627, and an F1-score of 0.9579, outperforming Informer, recurrent neural network (RNN), gated recurrent unit (GRU), and long short-term memory (LSTM); the standard Informer achieved a recall of 0.8702 and an F1-score of 0.9188. Across the five attacked-variable scenarios considered in this study, the average number of missed detections was reduced by 69.1% compared with Informer, and the framework also maintained higher recall and F1-score under surge attacks. On the SWaT benchmark, AMI-Informer achieved an F1-score of 0.8313, outperforming Informer, Anomaly Transformer, and cumulative sum (CUSUM). These results indicate that embedding attack evolution patterns into the loss function improves the detection of stealthy sensor deception attacks and enhances model stability in complex attack scenarios.