Today, the internet is extensively utilized across numerous domains. With indispensable applications ranging from education to healthcare, and from banking systems to e-commerce, it also attracts the attention of malicious actors. In the first quarter of 2024 alone, approximately 10 million attacks were recorded. Therefore, the detection and prevention of internet-based attacks is an increasingly critical issue that demands resolution.

In this study, we propose three Deep Learning (DL) models, namely Deep Neural Network (DNN), Densely Connected Deep Neural Network (DenseNet), and Convolutional Neural Network (CNN), for the detection of URL-based phishing attacks. Additionally, we examine the impact of Word Encoding (WE) and Character Encoding (CE) approaches on the performance of these models. The results demonstrate that the WE approach yields superior performance on large-scale datasets. Conversely, the CE approach achieves better results on smaller datasets that are insufficient for effective model training. In all experiments, the CNN model got the most successful, achieving an accuracy of 0.99732 on the first dataset and 0.83447 on the second dataset.