Advancing Phishing Detection with Voting Classifier, Hybrid Deep Models, and Flask-Based Real-Time Prediction Interface

Abstract

Phishing attacks remain a significant cybersecurity threat, deceiving users into revealing sensitive information through fraudulent websites. This study proposes a machine learning–based phishing detection framework enhanced with advanced feature selection, deep learning, and explainable intelligence. Multiple feature importance techniques, including mutual information, chi-square analysis, and permutation importance, are applied across Feedforward Neural Networks, Deep Neural Networks, TabNet, and Wide and Deep models to identify influential phishing indicators. To address class imbalance and improve robustness, SMOTEENN resampling is combined with an ensemble Voting Classifier integrating Random Forest and Bagging with Decision Trees. Experiments conducted on two public phishing datasets demonstrate superior performance, where the Voting Classifier achieved 98.7% accuracy on the phishing websites dataset and 98.5% accuracy on the web page phishing dataset, outperforming individual deep learning models. Explainable Artificial Intelligence techniques such as LIME and SHAP are incorporated to interpret predictions and highlight feature contributions, ensuring transparency and trust. For real-world deployment, the framework is implemented using the Flask platform, offering an interactive web interface with secure user signup and signin using SQLite. Users submit URLs for analysis, and the system provides real-time predictions as “Phishing website” or “Non Phishing website,” supporting reliable and interpretable phishing detection for online security.