Predictive Intrusion Detection for Industrial IoT Using World Models and Counterfactual Safety Analysis

Industrial Internet of Things (IIoT) systems are increasingly exposed to cyberattacks in which control commands remain protocol-compliant and statistically plausible while gradually driving physical processes toward unsafe operating conditions. Conventional intrusion detection approaches based on network signatures, reconstruction errors, or deviations from normal behaviour may therefore identify such attacks only after a noticeable process deviation has already occurred. We present PREVENT-IIoT, a consequence-aware intrusion detection framework designed to detect potentially harmful control actions before physical impact. The framework uses an action-conditioned probabilistic world model, trained on trusted normal operation, to estimate the future evolution of the physical process under the observed control action. A safety-reachability module then estimates the probability of leaving the safe operating envelope and the expected time to an unsafe state. To distinguish harmful manipulation from legitimate operational changes, a counterfactual risk mechanism compares the predicted risk of the observed action with the risk associated with a nominal supervisory action. These predictive signals are combined with dependency-graph consistency and network-telemetry indicators through a sequential fusion mechanism; the fusion weights are the only supervised element, and we also report a fully unsupervised variant that requires no attack labels. Experiments on a physics-based industrial process testbed covering nine attack families and ten random seeds show that PREVENT-IIoT achieves an AUPRC of 0.674, compared with 0.643 for the strongest classical baseline and 0.56 for the strongest of four deep baselines, namely, USAD, TranAD, LSTM-AE and MAD-GAN. Crucially, every reconstruction-, forecasting- and distance-based detector we evaluate, whether classical or deep, provides essentially no pre-impact warning and has a pre-impact rate near zero, whereas PREVENT-IIoT detects 53% of consequence-bearing attacks before physical impact. The safety model achieves a Brier score of 0.047, an expected calibration error of 0.056, and a mean absolute error of approximately 2 s for time-to-unsafe estimation. Leave-one-attack-out experiments show that the framework generalises to previously unseen attack families. We further add a long-memory CUSUM channel on the world-model residual that recovers the slow stealth-bias family, raising its recall from 0 to 0.84, which the windowed detector misses, at no cost to the false-alarm budget.