A feature-engineered dataset of benign and phishing URLs for machine learning and large language models evaluation

Area of Science:

• Cybersecurity
• Machine Learning
• Natural Language Processing

Background:

• Phishing websites pose a significant cybersecurity threat.
• Existing datasets lack features and balance, hindering model evaluation and reproducibility.
• Machine learning (ML) and large language models (LLMs) show promise for URL-based phishing detection.

Purpose of the Study:

• To introduce a comprehensive, feature-rich dataset for evaluating phishing detection models.
• To facilitate reproducible research and fair comparison between traditional ML and LLM-based approaches.
• To support the development of robust and scalable cybersecurity solutions.

Main Methods:

• Curated a dataset of 111,660 URLs (100,000 benign, 11,660 phishing).
• Enriched each URL with 22 numerical lexical and structural features.
• Included URL, domain, TLD, and label columns for interpretability and classification.

Main Results:

• Evaluated Random Forest (RF) and MiniLM-LLM with Logistic Regression (LR) on the dataset.
• Achieved >96% accuracy and >0.99 ROC AUC scores for both baseline models.
• Demonstrated the dataset's utility for robust phishing detection model evaluation.

Conclusions:

• The dataset provides a valuable benchmark for reproducible phishing detection research.
• It bridges the gap between traditional ML and LLM-based detection methods.
• Facilitates future work on adversarial robustness and scalable security models.