Enhancing Malicious URL Classification with Transformer-Based Embeddings

The rise in internet traffic has increased cyber threats, with malicious URLs being a major attack vector. Traditional rule-based methods are no longer sufficient, prompting the use of advanced machine learning techniques. This study evaluates transformer-based models — BERT, SBERT, RoBERTa, AlBERT, and DistilBERT—combined with LSTM, GRU, and MLP architectures for URL classification. Using standard metrics on a balanced dataset of benign and malicious URLs, AlBERT + LSTM achieved the highest accuracy (99.4%). The results confirm the strength of transformer embeddings in malicious URL detection and suggest further exploration of hybrid models for improved cybersecurity.