The Application of the BERT Transformer Model for Phishing Email Classification

Phishing attacks pose serious challenges to users, commercial institutions and governments alike including identity theft, companies and government secrets. The weakest link in any computer security system is the people and the success of phishing attacks is substantially based on manipulating human emotions. Therefore, we cannot solely rely on humans to detect phishing attacks as phishing attacks are becoming more and more sophisticated and can bypass filters set by anti-phishing techniques. More effective and automatic phishing detection mechanisms are required and many detectors have been proposed. However, the high number of phishing emails and attacks urges additional efforts. In this study, the effectiveness of word embedding in classifying phishing emails is investigated. Pre-trained transformer model BERT (Bidirectional Encoder Representations from Transformers) is fine-tuned to execute the task of classifying emails into phishing and non-phishing. The results show that transformer technology is accurate enough to improve phishing emails detection and can complement existing classifiers.