Multicollinearity Threatens AI Explainability in Cybersecurity
A new paper from arXiv reveals that multicollinearity in benchmark datasets like UNSW-NB15 destabilizes AI explainability tools such as SHAP and LIME used in intrusion detection systems. The authors prove mathematically that correlated features inflate attribution variance, making explanations non-identifiable. Experiments across linear, tree-based, kernel, and neural models confirm the fragility. They introduce the Explanability Fragility Score and two mitigation methods using VIF and correlation thresholding.
Key facts
- Paper ID: arXiv:2605.22529
- Investigates multicollinearity-induced instability in AI explainability for intrusion detection
- Formal theorem states multicollinearity inflates attribution variance
- Experiments on UNSW-NB15 benchmark dataset
- Evaluates linear, tree-based, kernel, and neural models
- Proposes Explanability Fragility Score metric
- Two novel mitigation methods using VIF and correlation thresholding
- Published on arXiv
Entities
Institutions
- arXiv