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

Explanations from AI security systems are unreliable when features are correlated: The paper demonstrates through formal theorem and experiments that when input features are highly correlated with each other, the explanations provided by popular tools like SHAP and LIME become unstable and non-identifiable—meaning the same model can produce different explanations for the same decision depending on minor variations in the data.

This fragility affects all major types of AI models used in cybersecurity: Testing across linear models, tree-based systems, kernel methods, and neural networks revealed that multicollinearity-induced explanation instability is a widespread problem, not limited to specific model architectures, raising concerns about the trustworthiness of explanations across the cybersecurity AI landscape.

Current benchmark datasets used to evaluate security AI contain this hidden vulnerability: The UNSW-NB15 dataset, a representative benchmark for intrusion detection systems, exhibits multicollinearity that destabilizes explanations—suggesting that AI systems trained and validated on such datasets may be deployed with false confidence in their interpretability.

Two practical solutions can stabilize explanations with different trade-offs: CAA-Filtering groups model attributions to stabilize explanations without retraining, while SHARP adds a regularization penalty during training to reduce explanation instability—offering organizations options depending on whether they can modify their training process.

This work highlights a critical gap in responsible AI practices: The paper reveals that security-critical AI systems can appear trustworthy and explainable while actually producing fragile, unreliable explanations, emphasizing the need to evaluate explanation robustness alongside predictive performance before deployment.