Optimizing Phishing URL Detection via Feature Selection and Deep Learning model

Abstract

Phishing URL detection is a critical cyber-security challenge, as adversaries continuously evolve attack patterns to defeat blacklist and rule-based defences. This paper presents a comprehensive multi-model deep learning framework trained on a curated set of 20 discriminative features extracted from the publicly available dataset_B_05_2020 benchmark (11,430 balanced URL samples). Five architectures—Deep Neural Network (DNN), Feedforward Neural Network (FNN), Autoencoder-based classifier, Graph Convolutional Network (GCN), and TabTransformer—are trained, evaluated, and compared under a unified protocol. The DNN achieves the best performance with 95.0% accuracy, 95.2% precision, 94.8% recall, and 95.0% F1-score. All models are deployed in an interactive Flask web application with SQLite-backed user authentication and real-time URL prediction with confidence scoring. Results demonstrate that compact feature engineering combined with deep learning provides robust, low-latency phishing detection suitable for practical deployment.