Smart Phishing Detection via URL Characteristics: Machine Learning and Deep Learning Techniques

In today's interconnected world, phishing remains one of the most pervasive threats to online security, deceiving users into revealing sensitive data such as account credentials, identity documents, and payment details by mimicking legitimate websites that are visually indistinguishable from their authentic counterparts. To address this challenge, we propose a URL-based phishing detection framework that leverages both classical machine learning and deep learning techniques. From each URL, 22 hand-crafted features are extracted and standardized, ranging from basic statistics (lengths, character counts, Shannon entropy) to structural and content-based indicators (subdomain count, suspicious word presence, protocol flags). On the small_dataset of 520285 URLs-assembled by merging the original train $(70 \%)$, validation $(20 \%)$, and test $(10 \%)$ splits into a single CSV file-we train and evaluate five traditional classifiers (Logistic Regression, K-Nearest Neighbors, Decision Tree, Support Vector Machine, and Random Forest) and three neural architectures (fully-connected ANN, 1D-CNN, and RNN with LSTM units) using 5-fold cross-validation. Among the classical methods, Random Forest achieves the highest accuracy of $\mathbf{9 4. 8 2 \%}$, while the leading deep models, ANN and RNN (LSTM) achieve an accuracy of $\mathbf{9 4. 4 7 \%}$ and $\mathbf{9 4. 5 9 \%}$. These results demonstrate the superiority of ensemble tree methods for phishing URL classification and highlight the competitive performance of sequence-based neural networks in capturing URL dynamics. The proposed system offers a robust and reproducible approach to phishing detection in real time and can serve as a foundation for more advanced security analytics.