Stabilising Explainability Fragility in Cybersecurity AI: The Impact and Mitigation of Multicollinearity in Public Benchmark Datasets
This paper introduces a novel metric, the Explanability Fragility Score, to quantify instability in AI explanations within cybersecurity intrusion detection systems. The research demonstrates that multicollinearity, a statistical issue with correlated features, can significantly inflate explanation variance and make feature importances non-identifiable. To address this, the paper proposes two mitigation methods, CAA-Filtering and SHARP, aimed at stabilizing AI explanations and improving trustworthiness in security-critical applications. AI
IMPACT Introduces methods to improve the trustworthiness and reproducibility of AI explanations in security-critical systems.