Towards Trustworthy Keylogger detection: A Comprehensive Analysis of Ensemble Techniques and Feature Selections through Explainable AI

High detection accuracy is achievable: The study demonstrates that keylogger detection can reach 99.76% accuracy using AdaBoost with Fisher Score feature selection, suggesting that technical solutions can effectively identify this security threat in most cases.

Explainability matters for trustworthy AI: The research emphasizes using Explainable AI techniques (SHAP and LIME) to show why the system flags something as a keylogger, not just that it is one—this transparency is crucial for users and security teams to understand and trust automated detection decisions.

Feature selection improves efficiency without sacrificing performance: By using approaches like Fisher Score, the study reduced computational complexity while maintaining near-perfect accuracy, meaning trustworthy AI systems don't need to be resource-intensive or difficult to audit.

Ensemble methods outperform single models: Combining multiple machine learning approaches (Stacking, Blending, Voting) produced better results than individual models, suggesting that responsible AI systems benefit from diverse analytical perspectives rather than relying on a single decision-making approach.

Multiple evaluation metrics reveal the full picture: The study used accuracy, precision, recall, F1 score, and AUC rather than relying on a single metric, demonstrating that comprehensive evaluation is necessary to ensure AI systems perform reliably across different scenarios and don't hide blind spots.