A Review of Feature Optimization and Explainable AI for Phishing URL Detection

Abstract

Phishing URL detection has evolved from reactive blacklists and heuristics to machine learning, deep learning, ensemble, transformer and explain-able AI tech-niques. The provided presentation outlines a valuable pipeline that includes data preprocessing, feature optimization, classification comparison, ensemble learning, XAI, and traditional evaluation metrics. This framework is strengthened by recent literature that focuses on multimodal evidence, cross-dataset validation, incremen-tal learning, adversarial robustness, and deployment efficiency. This review com-bines these directions focusing on URL and webpage features, feature selection, model families, ensemble strategies, and SHAP/LIME interpretability. Even though contextual and infrastructure evidence can help improve coverage at high-er computational cost, it is still found on the most attractive features, namely URL and lexical features, for early screening. Feature-selection studies show that sig-nificant dimensionality reduction is possible; however, feature importance varies depending on the collection time and data set. Explainability has been employed to reveal signals of importance, inform feature reduction, and has become a more common approach, but the quality of explanations is less frequently assessed than the quality of predictive accuracy. However, there are significant gaps on the fol-lowing points: generalization over time, generalization across datasets, adversarial evasion, reproducible benchmarking, inference cost and common standards of ac-curacy, robustness, efficiency, and explanation fidelity. The adaptive, lightweight and explanation-aware detection systems are proposed as research directions for future.