Explainable AI (XAI) in Industrial Intrusion Detection: Stability, Agreement and Correctness of KernelSHAP and LIME Under a Data-Quality Audit

An explanation of an industrial intrusion detection can be trusted only as far as the chain behind it allows, namely the dataset, the detector, and the explanation. Four research questions test this chain on combinations of seven variants of industrial control system and industrial Internet of Things (ICS/IIoT) datasets with three model families (XGBoost, CNN, logistic regression), using two explainers, the KernelSHAP and the LIME. We find that the detector is worth explaining in 19 of 21 combinations. Each explainer mostly repeats its answer, where a new seed keeps about 9 of the top 10 features for KernelSHAP and 8 for LIME. The two explainers agree more than random feature lists would on 18 of 19 combinations. Yet, although both explainers are local, little of this agreement appears to be instance-specific, since lists for two different instances overlap nearly as much as lists for the same instance. In process datasets, where we injected attacks into known variables, their agreement showed no clear link to KernelSHAP being right, so agreement does not imply correctness. We propose four checks per explanation, one per research question. For reproducibility, the settings for each explanation are also reported.