URL-Based Phishing Detection Using a BERT-LSTM Model

The rising prevalence of phishing websites presents substantial cybersecurity threats by deceiving users into revealing sensitive information through malicious URLs. This study aims to enhance phishing URL detection by introducing a deep learning model that combines Bidirectional Encoder Representations from Transformers (BERT) with Long Short-Term Memory (LSTM). In this framework, BERT is fine-tuned on a phishing URL dataset and utilized as a contextual embedding to represent URL tokens, while Bayesian Optimization is employed to identify optimal hyperparameter settings during model training. Experimental results demonstrate that the BERT-LSTM model achieves impressive detection performance, with a precision of 0.9299, recall of 0.9795, F1-score of 0.9540, accuracy of 0.9756, and ROC-AUC of 0.9962. The model consistently outperforms embedding-based methods such as Word2Vec, FastText, and GloVe, as well as a classical baseline model using Logistic Regression with TF-IDF features. These findings suggest that the contextual embeddings generated by BERT effectively capture structural patterns in URLs, leading to more accurate phishing detection and providing a promising approach for enhancing cybersecurity systems.