Enhancing URL Classification with Hyperparameter-Tuned Convolutional Neural Networks

Phishing attacks, which exploit deceptive URLs to steal sensitive information, pose a significant cybersecurity threat. This paper presents a novel framework for phishing URL detection utilizing Convolutional Neural Networks (CNNs) optimized through advanced hyperparameter tuning. The proposed method involves preprocessing URLs by tokenizing them at the character level and padding sequences to ensure uniform input dimensions. A CNN model is constructed with embedding layers, convolutional operations, max pooling, and dense layers. Hyperparameter optimization is performed using the Hyperband algorithm from Kera's Tuner, systematically exploring various configurations including filter counts, kernel sizes, pooling strategies, and learning rates. Evaluated on a large dataset of phishing URLs, the model achieved a test accuracy of 80.50%, demonstrating the efficacy of CNNs for phishing detection and the benefits of thorough hyperparameter tuning. This study contributes to the field of cybersecurity by offering a robust, data-driven approach to phishing detection and providing insights into optimizing deep learning models for practical applications.