Phishing attacks are a serious cybersecurity trouble just one successful attack can lead to major fiscal losses and exposure of sensitive information, and although machine literacy models have come relatively effective at detecting phishing websites, numerous studies overlook important aspects similar as explainability, robustness to noisy data, and real- world usability; this study addresses those gaps by erecting on the work of Awasthi and Goel( 2022) and using the UCI Phishing Websites dataset, which contains over 11,000 samples and 30 features, where six bracket algorithms Logistic Retrogression, Support Vector Machines, Decision Trees, K- Nearest Neighbors, Random Forest, and XGBoost are estimated with hyperparameter tuning for ensemble models and10-foldcross-validation for dependable performance, and among all models, XGBoost performs the stylish, achieving 96.70 delicacy and a 97.06 F1- score; the main benefactions of the study are threefold first, explainability, where SHAP( SHapley Additive Explanations) identifies anchor URL, HTTPS operation, and website business as the most important features in detecting phishing websites; second, robustness, where the model is tested with noisy data up to 20 and XGBoost still maintains a strong delicacy of 87.38; and third, deployment, where the XGBoost model proves to be effective with an conclusion time of 13.11 milliseconds and a compact size of just 501 KB; overall, the study demonstrates that it's possible to achieve high delicacy while also icing interpretability, robustness, and practical usability, showing that featherlight ensemble models, particularly XGBoost, give an effective and deployable result for phishing discovery in real- world scripts.

Index Terms: Phishing Detection, Machine Learning, Ensemble Learning, XGBoost, Explainable Artificial Intelligence (XAI), SHAP, Robustness Analysis, Deployment Analysis, Cybersecurity, Classification.