Metadata driven malicious URL detection using RoBERTa large and multi source network threat intelligence

In this study, we address the critical security challenge of detecting malicious URLs, which serve as primary cause for website phishing, malware attacks, and website defacement. Detection of malicious links has remained active research in Artificial Intelligence as machine learning (ML) models and deep learning (DL) are applied in this regard. The use of sequential models captures character-level patterns but remain vulnerable to novel adversarial manipulations. This study proposes to apply state of the art large language model which are dual-mechanism, RoBERTa-Large transformers that combine contextualized subword embeddings with lightweight metadata signals via attention layers. Fine-tuned on a balanced dataset encompassing benign, defacement, phishing, and malware URLs, the model achieves 98% overall accuracy substantially outperforming ML and DL models. To enhance interpretability, SHAP and LIME explainability analysis, confirming that features such as URL length, slash depth, and entropy drive predictions, while the transformer's attention heads isolate subtle lexical anomalies. The empirical analysis-based findings demonstrate that integrating metadata attention with a masked language model yields state-of-the-art performance and transparent decision-making for real-world malicious URL detection.