Robust and Accurate Phishing Detection Using Enhanced DistilBERT: A Transformer-Based Approach

Abstract

Phishing scams are still a big problem in email, constantly changing and finding ways to sneak past regular filters by using tricky language and structures. To tackle this, we developed an improved system for spotting these fake emails. Our approach leverages an enhanced version of DistilBERT, fine-tuned and optimized specifically for phishing detection. The proposed “Enhanced DistilBERT” was trained on a comprehensive Email Phishing Dataset—a carefully curated collection of around 82,500 emails drawn from multiple trusted sources, including Enron, Ling, CEAS, Nazario, Nigerian Fraud, and Spam Assassin. This dataset provides a balanced mix of 42,891 phishing attempts and 39,595 legitimate messages, with a strong focus on both the subject lines and email body text. Such diversity enables robust text analysis and helps the model effectively learn the subtle distinctions between phishing and genuine emails. When evaluated, Enhanced DistilBERT outperformed other leading models, including RoBERTa and the original DistilBERT, by achieving perfect scores in accuracy, precision, recall, and F1-score, alongside a near-perfect ROC AUC of 99.76%. The model not only excelled in standard evaluations but also demonstrated resilience against noisy and adversarial data, maintaining high accuracy and low false-positive rates under challenging conditions. These results confirm the superiority and practicality of our approach, establishing Enhanced DistilBERT as a reliable and ready-to-deploy tool for combating phishing threats in today’s dynamic digital landscape.