A Large Language Transformer Approach to Phishing Email Categories Classification
Phishing emails remain a common and evolving cyber threat, which also employs sophisticated social engineering mechanisms to evade regular filtering methods. While rule-based and classical machine learning approaches have gained wide usage, their reliance on static characteristics limits their efficacy against fresh and advanced attacks. In this work, the application of large language transformer (LLT) models, specifically BERT and its variants, is investigated for phishing email classification. This research suggest a robust methodology for fine-tuning these models on an extensive dataset comprising both human-written and LLM-generated phishing and authentic emails. Our findings indicate that fine-tuned transformers dramatically surpass a set of conventional baseline models, such as Support Vector Machines (SVM), Random Forests, and LSTMs, with state-of-the-art accuracy and F1-scores. Furthermore, we provide a comparative assessment of model performance on specific social engineering tactics, authority, urgency, and persuasion, and demonstrate that the transformers possess a stronger ability to detect subtle malicious intent. We also examine the zero-shot capability of the BART-large-mNLI model, and we describe the performance disparity between general and domain-specific approaches. Our findings conclusively establish that fine-tuned LLTs represent a paradigmatic shift in phishing detection, from keyword matching to deep contextual understanding, and represent an effective, scalable countermeasure for modern email security systems .RoBERTa-base is the best performer with an accuracy of 94%.
Download Paper