🤖 AI

Explainable AI-Driven Cyber Risk Analytics and Model Reliability Assessment for Intelligent Governance of U.S. Critical Infrastructure: An XGBoost and SHAP-Based Intrusion Detection Framework

This paper proposes an explainable AI-driven framework utilizing XGBoost and SHAP techniques on the CICIDS2017 dataset to enhance intrusion detection, cyber risk analytics, and model reliability for the intelligent governance of U.S. critical infrastructure.

Original authors: B. M. Taslimul Haque, Md. Arifur Rahman, Md. Serajul Kabir Chowdhury Rubel, Md. Iqbal Hossan

Published 2026-06-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: B. M. Taslimul Haque, Md. Arifur Rahman, Md. Serajul Kabir Chowdhury Rubel, Md. Iqbal Hossan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Guarding the Digital City

Imagine the United States' critical infrastructure (power grids, hospitals, banks, and water systems) as a massive, bustling city. For years, the city guards used a simple rulebook: "If you look like a known criminal, stop you." This worked okay when criminals were predictable. But today, the "criminals" (cyber attackers) have changed. They don't just knock on the door; they disguise themselves as delivery drivers, blend into the crowd, or launch thousands of fake people at the gate at once (DDoS attacks) to hide their real theft.

The old rulebook is too slow and rigid to catch them. So, the researchers in this paper tried to build a smart, AI-powered security guard. But here's the catch: they didn't just want a guard that is accurate; they wanted one that can explain its reasoning. If the AI screams "Stop that person!", the human in charge needs to know why before they lock the gate.

The Problem: The "Black Box" Dilemma

In the past, many AI security systems were like black boxes. You put data in, and a "Risk Score" came out.

The AI says: "99% chance this is an attack."
The Human asks: "Why?"
The AI says: "Because I said so."

For a hospital or a power grid, "because I said so" isn't good enough. If the AI is wrong, it could shut down a life-saving machine or a water treatment plant. The researchers wanted to build a system that is not only smart but also transparent and trustworthy.

The Experiment: Training the AI

The researchers used a massive dataset called CICIDS2017. Think of this dataset as a giant library of video recordings from a city's security cameras, containing millions of hours of footage.

The Good Guys: Normal traffic (people buying coffee, sending emails).
The Bad Guys: Various attacks, including DDoS (a massive crowd blocking the streets), brute force (trying every key on a keyring), and botnets (armies of infected computers).

They trained four different types of "AI guards" (machine learning models) to watch this footage:

XGBoost: A very sharp, fast learner that builds a decision tree.
Random Forest: A committee of many decision trees that vote on the answer.
Decision Tree: A single, easy-to-follow flowchart.
Logistic Regression: A simple, linear rule-follower.

The Results: The Champion Guard

The researchers tested these guards on new footage they hadn't seen before.

The Winner: One specific model (a Support Vector Machine, or SVM) performed almost perfectly. It correctly identified nearly every single attack and almost every single normal person.
The Score: It made only 11 mistakes out of thousands of records. It correctly spotted 2,571 attacks and 1,918 normal events.
The "False Alarm" Problem: In security, a "False Positive" is like a guard screaming "Thief!" when it's just a mailman. This causes panic and wastes time. This model had very few false alarms, which is crucial for keeping the city running smoothly.

The Secret Sauce: The "Why" (Explainable AI)

This is the most important part of the paper. The researchers didn't just stop at "it works." They used a tool called SHAP (SHapley Additive Explanations).

Think of SHAP as a magnifying glass that the AI uses to show the human guard exactly what it saw.

Without SHAP: "I think that car is a bomb."
With SHAP: "I think that car is a bomb because:
1. It's driving 100mph (Flow Bytes/s).
2. It's carrying huge, heavy crates (Packet Size).
3. It's heading straight for the bank vault door (Destination Port)."

The study found that the AI was looking at specific clues to spot the DDoS attacks:

Flow Duration: Normal traffic is quick (like a quick handshake). Attack traffic often stays connected for a long time to drain resources.
Packet Size: Attackers often send huge, heavy packets to clog the pipes.
Destination Port: Attacks often target specific "doors" (like Port 80 for websites) with unusual intensity.

Why This Matters for Governance

The paper argues that for critical infrastructure (like the power grid), trust is just as important as accuracy.

If a model is 99% accurate but you don't know why it flagged something, you can't trust it enough to shut down a hospital.
By using SHAP, the model becomes auditable. A human manager can look at the explanation, say, "Yes, the packet size is huge, that makes sense," and then take action with confidence.

The Limitations (The Fine Print)

The authors are honest about what their study didn't do:

Old Data: The data is from 2017. Attackers have evolved since then, so the AI might need retraining for today's tricks.
Static Test: They tested the AI on a recording, not in a live, real-time city where traffic changes every second.
One Type of Attack: This study focused heavily on DDoS attacks (the "crowd" attacks). It didn't fully test the AI against more subtle, sneaky attacks like data theft or ransomware.

The Bottom Line

This paper proves that we can build AI security systems that are not just "smart" but also honest and explainable. By combining powerful detection (like the SVM model) with a "magnifying glass" (SHAP), we can create a cybersecurity system that human leaders can trust to protect the vital systems that keep our society running. It moves us from "blindly trusting the computer" to "understanding and verifying the computer's decision."

Technical Summary: Explainable AI-Driven Cyber Risk Analytics for U.S. Critical Infrastructure

Problem Statement
U.S. critical infrastructure sectors (energy, healthcare, transportation, finance, communications) face an expanding cyberattack surface driven by AI-governed digital technologies. Traditional, rule-based cybersecurity mechanisms are increasingly inadequate against sophisticated threats such as Distributed Denial of Service (DDoS) attacks, Advanced Persistent Threats (APTs), botnets, and ransomware. While machine learning (ML) offers promise for intrusion detection, existing frameworks often prioritize classification accuracy at the expense of model reliability, explainability, and governance transparency. This "black box" nature creates a fundamental trust deficit for infrastructure operators and policymakers who require auditable, interpretable decision support to justify operational actions.

Methodology
The study employs a quantitative, experimental research design to develop a resilient cyber risk analytics and model reliability assessment framework. The methodology proceeds through six sequential phases using the CICIDS2017 dataset, a realistic benchmark containing labeled network flows for DDoS, brute force, botnet, web attacks, and infiltration traffic.

Data Preprocessing: Raw network flows underwent cleaning (removing duplicates, imputing missing/infinite values), encoding (converting string labels and protocol types to integers), normalization (Min-Max scaling for continuous features), and feature selection. The selection process utilized Pearson correlation analysis (removing pairs with $|r| > 0.95$ ) and Random Forest-based importance ranking (removing features below the 10th percentile), reducing the initial 79 features to a set of approximately 40–45 discriminative features.
Model Training: Four supervised classifiers were trained and comparatively evaluated: XGBoost, Random Forest, Decision Tree, and Logistic Regression. The study also highlights the performance of a Support Vector Machine (SVM) classifier on a binary DDoS vs. benign task. Experiments utilized an 80/20 stratified train-test split with five-fold stratified cross-validation.
Explainable AI (XAI): To address the opacity of ML models, SHAP (SHapley Additive Explanations) analysis was integrated. Using the TreeSHAP algorithm, the study generated feature-level explanations (beeswarm plots, bar plots of mean absolute SHAP values, and force plots) to interpret model predictions.
Reliability Assessment: Performance was evaluated not only on accuracy but on a comprehensive suite of metrics: Precision, Recall, F1-score, ROC-AUC, and specifically the False Positive Rate (FPR), which is critical for preventing alert fatigue in operational environments.

Key Results

Classification Performance: The SVM classifier achieved near-perfect discrimination on the binary DDoS versus benign classification task, yielding an ROC-AUC approaching 1.00. The confusion matrix revealed 2,571 true positives and 1,918 true negatives, with only 11 total misclassifications (7 false positives and 4 false negatives).
Feature Discriminative Power: Exploratory analysis confirmed that specific network features carry substantial discriminative information:
- Flow Duration: Long-duration flows were associated with persistent sessions and TCP connection-exhaustion attacks.
- Packet Size: DDoS traffic exhibited significantly larger average packet sizes compared to benign traffic, reflecting volumetric flooding strategies.
- Destination Ports: Anomalous volumes on standard ports (80, 443, 53) combined with other features served as indicators of attack activity.
- Flow Bytes/s: High-intensity traffic anomalies were clearly distinguishable in the distribution tail.
Explainability: SHAP analysis successfully identified the most influential network features driving model predictions, providing the feature-level transparency required for governance actors to audit classification outputs with informed confidence.

Significance and Claims
The paper asserts that its primary contribution is not merely demonstrating that ML can detect attacks—a capability already established in literature—but rather showing how to do so in a manner that is reliable, transparent, and operationally useful for human decision-makers.

Governance Trustworthiness: The framework advances "trustworthy, explainable AI" by combining reliable intrusion detection with transparent, governance-ready decision support. It addresses the gap between high accuracy and the need for auditable, explainable models in critical infrastructure.
Operational Accountability: By integrating SHAP, the study moves beyond detection accuracy toward systems that are interpretable, auditable, and operationally accountable. This allows analysts to understand why a model flagged an event (e.g., specific packet size anomalies or flow duration patterns), enabling informed actions such as system isolation or further investigation.
Resilience over Static Protection: The study aligns with the shift from static protection to resilience-based governance, arguing that models must be explainable and their failure modes understood to support adaptive security management.

Limitations and Scope
The authors maintain a modest stance regarding the generalizability of their findings:

The analysis focused on binary classification (DDoS vs. benign) within a specific subset of the CICIDS2017 dataset; performance on the full taxonomy of 15 attack categories (including subtle threats like brute force or infiltration) was not fully evaluated in this study.
Experiments were conducted in an offline, static environment; the framework's behavior under real-time conditions, concept drift, and adversarial evasion techniques remains to be validated.
Class imbalance was addressed via stratified sampling but not synthetic augmentation (e.g., SMOTE), which may impact recall for minority attack classes in broader applications.

In conclusion, the study establishes a credible foundation for an integrated framework where detection, reliability assessment, and explainability operate together to support the human governance processes essential for protecting U.S. critical infrastructure.