Stabilising Explainability Fragility in Cybersecurity AI: The Impact and Mitigation of Multicollinearity in Public Benchmark Datasets

2026-05-21 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Expert, quick

Summary

This paper identifies multicollinearity as a significant, unexplored vulnerability impacting AI explainability in intrusion detection systems (IDS). It formally proves that multicollinearity inflates attribution variance, rendering explanations and feature importances non-identifiable. Comprehensive experiments on the UNSW-NB15 benchmark dataset validate this theorem across linear, tree-based, kernel, and neural models, using both full and pruned feature sets. The research introduces the "Explanability Fragility Score" and proposes two novel mitigation methods: CAA-Filtering, which stabilizes explanations by grouping attributions of trained models, and SHARP, a training-time regularization framework that penalizes attribution instability for monotonic improvement. The findings, quantified using Kendall's τ for instability, support stable predictive performance and have direct implications for the trustworthiness and reproducibility of XAI in security-critical contexts.

Key takeaway

For AI Security Engineers deploying XAI in intrusion detection systems, you must address multicollinearity in your datasets. Ignoring this vulnerability inflates explanation variance, compromising the trustworthiness and reproducibility of your AI's insights. Incorporate multicollinearity mitigations like CAA-Filtering or SHARP into your IDS pipelines to ensure stable and reliable explanations, thereby enhancing the integrity of your security decisions.

Key insights

Multicollinearity in cybersecurity datasets inflates AI explanation variance, making attributions non-identifiable and impacting XAI trustworthiness.

Principles

• Multicollinearity inflates AI attribution variance.
• Explanations are non-identifiable under multicollinearity.
• XAI trustworthiness requires multicollinearity mitigation.

Method

The paper proposes CAA-Filtering to stabilize explanations by grouping attributions of trained models, and SHARP, a training-time regularization framework that penalizes attribution instability.

In practice

• Evaluate explainability fragility with Explanability Fragility Score.
• Apply CAA-Filtering for post-hoc explanation stabilization.
• Implement SHARP for training-time attribution regularization.

Topics

• AI Explainability
• Multicollinearity
• Intrusion Detection Systems
• Cybersecurity AI
• UNSW-NB15 Dataset
• Attribution Stability

Best for: Research Scientist, AI Scientist, AI Security Engineer, Machine Learning Engineer

Related on AIssential

• 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 — Integrating XAI with ML-based intrusion detection enhances transparency and trust for critical infrastructure cyber governance.
• Context-Fractured Decomposition Attacks on Tool-Using LLM Agents: Exploiting Artifact Provenance Gaps — Context-Fractured Decomposition (CFD) attacks exploit untracked artifact provenance in LLM agents, enabling delayed, multi-step jailbreaks.
• Cognitive Threat Intelligence and Explainable Federated Security Analytics for distributed Infrastructure Systems — The framework uses FL, XAI, and cognitive analytics for privacy-preserving, collaborative threat detection in distributed systems.
• 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 — An XAI-driven framework enhances cyber risk analytics and intrusion detection for critical infrastructure by improving model transparency and reliability.
• When Fine-Tuning Changes the Evidence: Architecture-Dependent Semantic Drift in Chest X-Ray Explanations — Fine-tuning medical image models can alter visual evidence (semantic drift) even with stable diagnostic accuracy.
• The Query Channel: Information-Theoretic Limits of Masking-Based Explanations — Masking-based explanations are fundamentally limited by information-theoretic rate-capacity tradeoffs, dictating recoverability.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.