<![CDATA[Phishing is a significant cybersecurity threat in which malicious URLs deceive users to steal sensitive data. Traditional detection methods, such as blacklists, often fail to keep pace with evolving phishing techniques. Deep learning, particularly Convolutional Neural Networks (CNNs), offers strong potential in phishing URL classification by capturing structural and semantic character-level patterns. However, CNN training demands high computational resources and risks overfitting. This study investigates the effectiveness of early stopping as a regularization technique to improve efficiency and generalization in character-based CNN models. Using a large-scale dataset of 130.080 URLs across four classes (benign, phishing, malware, defacement), the model employed character tokenization, embedding, convolution-pooling layers, and softmax classification. Early stopping monitored validation loss with patience values of 3, 5, and 10 epochs. Results show a 51% training time reduction and accuracy improvement from 96% to 97%, confirming early stopping as an efficient and robust detection approach.]]>
Copyrights © 2026
