Fault Tolerance vs. Attack Detection in Industrial Control Systems: A Deep Learning Approach

Industrial Control Systems (ICS) are based on faulttolerant mechanisms to ensure operational stability. However, these mechanisms can mask anomalies, potentially suppressing indicators of cyber-attacks. This study investigates the interplay between fault tolerance and attack detection using the Tennessee Eastman Process (TEP) dataset. By integrating thresholdbased fault masking, feature engineering, and temporal models (LSTM), we evaluate how different approaches impact detection performance. Our results indicate that, while fault masking provides stability, it significantly reduces the recall of attack detection. PCA analysis suggests high feature overlap between faulty and attack states, further complicating classification. We compare fixed-threshold masking with probabilistic masking to assess trade-offs in accuracy and detection robustness. These findings highlight key challenges in designing resilient ICS capable of balancing fault tolerance and cybersecurity.