Comparative Study of AI-Based Models for Malicious URL Detection

Abstract

Phishing and other malicious URL-based threats continue to challenge cybersecurity systems, exploiting the limitations of traditional defense mechanisms such as blacklists and heuristic detection methods. This study conducts a comparative analysis of artificial intelligence (AI)-based techniques used for malicious URL detection. The investigation focuses on both traditional machine learning models, such as LightGBM and XGBoost, and deep learning approaches like BiLSTM with attention. The research also reviews the role of metaheuristic optimization methods—particularly Particle Swarm Optimization (PSO) and Optuna—for hyperparameter tuning and feature selection. By analyzing several benchmark studies and recent contributions, the study highlights the strengths, limitations, and trade-offs associated with these models in terms of accuracy, scalability, and robustness. Special attention is given to feature engineering strategies such as lexical analysis and TF-IDF vectorization, which significantly influence performance across different classifiers. This work provides a structured theoretical foundation for future research on adaptive phishing detection systems, offering valuable insights for academic and industrial researchers interested in building scalable, explainable, and AI-driven security solutions.