Explainable AI–Based Threat Modeling for Trustworthy Cyber–Physical Systems under Intelligent Adversaries

XAI-Guard (Explainable Threat Modeling for Cyber-Physical Systems) is a comprehensive framework designed to detect, explain, and categorize cyber-physical attacks in real-time. This project focuses on enhancing the trustworthiness of anomaly detection systems by providing multi-modal explanations and automated threat taxonomy, specifically tailored for critical infrastructure like the Secure Water Treatment (SWaT) plant.

---

🚀 Key Objectives

1. Detection under Intelligent Adversaries: Implementing robust classifiers that can identify stealthy manipulations designed to bypass traditional threshold-based alarms.
2. Trustworthy Explainability: Leveraging independent XAI methods (SHAP & LIME) to validate model reasoning.
3. Threat Taxonomy: Automatically clustering attack patterns to understand adversary strategies (e.g., sensor spoofing vs. actuator sabotage).

---

🛠 Model Engineering & Detection Strategy

The core of the system is an XGBoost (Extreme Gradient Boosting) binary classifier, optimized for the high-dimensional, imbalanced nature of CPS data.

1. Handling Class Imbalance with scale_pos_weight

In the SWaT dataset, attacks represent only ~3.8% of the total data. To prevent the model from becoming biased toward "Normal" operation:

• We compute SPW (Scale Pos Weight) = count(negative samples) / count(positive samples).
• In our implementation, this ratio is approximately 25.4:1.
• This forces XGBoost to penalize misclassifications of the minority (Attack) class 25 times more heavily than the majority class.

2. Decision Threshold Optimization (0.3)

Unlike standard classifiers that use a 0.5 threshold, XAI-Guard uses a 0.3 threshold:

• Rationale: In critical infrastructure security, a False Negative (missing an attack) is far more dangerous than a False Positive.
• By lowering the threshold, we significantly boost Recall (Detection Rate), ensuring that even low-probability threat signals are flagged for human review.

3. Temporal Integrity

Data is split using a Temporal Split (80/20) rather than random shuffling to respect the time-series nature of CPS processes. This ensures the model learns the physical correlations (e.g., if Pump A turns on, Flow Meter B should rise after X seconds) without seeing the "future" during training.

---

🧠 Explainable AI (XAI) Deep Dive

To achieve "Trustworthy AI," the system utilizes two independent engines to explain every detection.

1. SHAP (SHapley Additive exPlanations)

• The Theory: Based on Coalitional Game Theory. SHAP treats each sensor reading as a "player" in a game and calculates its "payout" (contribution to the final prediction).
• Why it's helpful: It provides Global Consistency. If SHAP identifies LIT101 as the top feature, it means that sensor is mathematically the most responsible for the anomaly based on the model's entire learned logic.
• Local Explanation: For a specific attack at 12:00 PM, SHAP tells the operator exactly which sensors shifted the probability from 0.0 to 0.9.

2. LIME (Local Interpretable Model-agnostic Explanations)

• The Theory: LIME ignores the internal math of XGBoost. Instead, it takes a single attack instance, creates thousands of slightly "perturbed" versions of it, and sees how the model's prediction changes. It then builds a simple Local Linear Surrogate Model to explain that specific point.
• Why it's helpful: It provides an External Validation. Because LIME is model-agnostic, it acts as a "sanity check" for SHAP.

3. XAI Agreement & Consensus

The system computes an Agreement Score:

• Top-1 Agreement: Do SHAP and LIME agree on the #1 most important sensor?
• Jaccard Overlap: How much do the Top-3 features from both methods overlap?
• Trust Factor: High agreement (e.g., >60%) gives the operator confidence that the explanation is a physical reality, not a mathematical artifact.

---

🔬 Attack Pattern Clustering (Taxonomy)

Instead of just labeling a row as "Attack", the system performs KMeans Clustering on SHAP vectors:

• Why Cluster on SHAP? Clustering on raw sensor data is often noisy. By clustering on the Explanations, we group attacks by their Root Cause.
• Auto-Labeling:
  • The system compares the Statistical Deviation (which sensor moved most physically) with the SHAP Importance (which sensor the model focused on).
  • If they agree, the cluster is labeled: [Sensor]-dominant attack.
  • If they disagree, it's labeled: Mixed disturbance (C/D conflict), indicating a complex, multi-stage attack.
• Validation: Uses Silhouette Scores to ensure clusters are distinct and meaningful.

---

📊 Dashboard UI Features (Fixes & Enhancements)

The app.py has been overbuilt for professional use:

• 📡 Key Sensor Grid: Fixed display of the Top-8 SHAP sensors. Users see the exact same sensors every loop, making it easier to track process stability.
• 📈 Delta Tracking: Every sensor metric shows a Delta (Δ) arrow. If FIT101 rises, the red/green arrow shows the immediate shift from the previous second.
• 🚨 Pulse Alerts: A high-end "Pulse-Red" animation with box-shadow effects replaces cheap blinking, providing a premium SOC (Security Operations Center) feel.
• 🗂 Session Log: A persistent table of all detected attacks, including timestamps and the consensus between SHAP and LIME.

---

📁 File Manifest & Usage

Running the System

1. Execute Pipeline: python ctmas_pipeline.py
   • This generates all .json, .csv, and .png artifacts in /outputs.
2. Launch Dashboard: streamlit run app.py

Key Outputs in /outputs

• shap_attack_clusters.png: The PCA-reduced map of threat signatures.
• attack_cluster_summary.csv: The automated taxonomy of attack types.
• xai_agreement_summary.txt: Detailed metrics on explanation trustworthiness.
• shap_local_explanations.csv: The human-readable reasoning for every threat.

---

🏛 Project Context

Full Title: Explainable AI–Based Threat Modeling for Trustworthy Cyber–Physical Systems under Intelligent Adversaries

This framework demonstrates that for AI to be deployed in critical infrastructure, it must not only be accurate but also interpretable, consistent, and transparent. By combining Gradient Boosting with Dual XAI and Clustering, we provide a complete toolset for modern cyber-defense.