Abstract

Phishing is the most widespread cyberattack in the world, and email remains the most frequently used medium. Traditional machine learning models are one method for phishing detection, and many studies have been conducted to identify phishing in customer service emails; however, no one has looked into combining the Transformer and Graph Neural Network (GNN) to identify fraudulent emails sent to customer service. This study proposes a hybrid model that combines a GNN for graphical features with RoBERTa for textual features. The experiments were performed with the CEAS-08 phishing email dataset, an openly available benchmark released as part of the 2008 CEAS Live Spam Challenge, partitioned into 70% for training and 30% for validation and testing. The hybrid RoBERTa + GNN model developed here attained an accuracy of 98.40% with the frozen RoBERTa base on Google Colab Free. In comparison with other traditional classifiers like SVM, KNN, and Naive Bayes, the new hybrid model provides more contextualized knowledge with RoBERTa and relational knowledge with GNN. Future work may investigate fine-tuning the RoBERTa base to further improve performance, opening a new research pathway that offers deep contextual and structural insights beyond what conventional models achieve.