Causal Digital Twins for Cyber-Physical Security: A Framework for Robust Anomaly Detection in Industrial Control Systems

Abstract: Industrial Control Systems (ICS) face growing cyber-physical attacks that exploit both network vulnerabilities and physical processes. Current anomaly detection methods rely on correlation-based analysis, which cannot separate true causal relationships from spurious associations. This limitation results in high false alarm rates and poor root cause analysis. We propose a novel Causal Digital Twin (CDT) framework for cyber-physical security in medium-scale ICS. Our method combines causal inference theory with digital twin modeling. The framework enables three types of causal reasoning: association for pattern detection, intervention for understanding system responses, and counterfactual analysis for attack prevention planning. We evaluate our framework on three industrial datasets: SWaT, WADI, and HAI, with validation through physical constraint compliance (90.8\%) and synthetic ground truth testing (structural Hamming distance 0.13). Results show significant improvements over seven baseline methods. Our CDT achieves F1-scores are $0.944 \pm 0.014$ for SWaT, $0.902 \pm 0.021$ for WADI, and $0.923 \pm 0.018$ for HAI with statistical significance ($p < 0.0024$, Bonferroni corrected). The framework reduces false positives by \SI{74}{\percent} and achieves \SI{78.4}{\percent} root cause analysis accuracy compared to \SI{48.7}{\percent} for existing methods. Counterfactual analysis enables defense strategies that reduce attack success by \SI{73.2}{\percent}. The system keeps real-time performance with \SI{3.2}{ms} latency, which is suitable for industrial deployment, while providing interpretable explanations for operators.