A study on malicious URL detection based on BiGruCNN-MHA-FTW

Malicious URLs disguised as legitimate links pose growing threats to cybersecurity. We propose BMFTW (BiGruCNN-MHA-FTW), a deep learning model for malicious URL detection that integrates character-level index mapping, Word2Vec embeddings, multi-scale dilated convolutional neural networks, BiGRU layers, and residual multi-head attention to capture both structural and contextual URL features. We propose a hybrid FTW loss combining Focal Tversky and Weighted Binary Cross-Entropy to address class imbalance. L2 regularization, dropout, and exponential learning rate decay enhance generalization, while Bayesian optimization tunes hyperparameters using the negative F1-score. Experimental results on the Malicious and Benign URLs dataset show that our model achieves 98.28% accuracy, 94.25% precision, 98.71% recall, and a 96.43% F1-score in binary classification, as well as a 94.93% F1-score in multi-class detection. These results significantly outperform baseline models with statistical confidence (P < .05). The findings demonstrate the effectiveness and robustness of combining deep learning and NLP-based strategies in malicious URL detection, offering practical value for real-world cybersecurity applications.