Real-Time Phishing URL Detection Using a Hybrid Stacking Ensemble: Gradio and Browser Extension Deployment

Abstract

Phishing attacks remain a prevalent and rapidly evolving cybersecurity threat, leveraging deceptive Uniform Resource Locators (URLs) and fraudulent websites to steal sensitive user data, financial credentials, and personal information. Traditional detection mechanisms, such as blacklist-based and heuristic approaches, struggle to mitigate zero-day phishing threats due to their reliance on static, manually updated databases. While machine learning and ensemble techniques have enhanced detection accuracy, existing literature predominantly focuses on offline evaluations using static datasets, offering limited support for real-time deployment, adaptability to evolving attack patterns, and continuous monitoring. To bridge this gap, this paper introduces a hybrid machine learning framework for real-time phishing website detection. The proposed architecture integrates an Artificial Neural Network (ANN) and a Bagging K-Nearest Neighbors (Bagging-KNN) classifier through a Logistic Regression-based stacking ensemble, combining their complementary learning capabilities to maximize classification performance while minimizing prediction error. Developed using the PhiUSIIL Phishing URL Dataset, the framework implements a leakage-free machine learning pipeline encompassing automated data cleaning, a train/test split performed prior to any preprocessing, feature engineering, StandardScaler-based normalization, SMOTE-based class balancing, and SelectKBest feature selection, all splitting-dependent steps fitted exclusively on the training partition. The framework’s efficacy is validated using Accuracy, Precision, Recall, F1-score, ROC-AUC, and Confusion Matrix analysis. Beyond offline validation, the model is operationalized through two real-time deployment channels: a Gradio-based web interface for on-demand URL analysis, and a Chrome browser extension that automatically screens the active browser tab using a combination of rule-based checks, live queries to the deployed model, and a local heuristic fallback. By unifying ensemble learning, a leakage-conscious preprocessing pipeline, and dual real-time deployment tools, the proposed framework provides an effective, transparently evaluated solution for real-world phishing detection, while explicitly discussing the boundaries within which its strongest offline results should be interpreted.