A hybrid stacked bidirectional long short-term memory and Bidirectional Encoder Representations from Transformers model for enhanced phishing email detection

Phishing attacks have been among the latest cybersecurity threats as they exploit people’s weaknesses and compromise critical information. Traditional methodologies of phishing detection practices based on machine learning have generally been unsuccessful in offering high performance due to limited generalizability and very high false positive rates. In this paper, a novel hybrid deep learning model is proposed that incorporates Bidirectional Encoding Representation from Transformers (BERT) with Bidirectional Long Short-TermMemory (BLSTM) to increase the detection of suspicious emails within the framework of the proposed model. The model utilizes the contextual understanding of texts from BERT and sequential learning from BLSTM to enhance classification accuracy. Experimental results using an available phishing email dataset show that the proposed approach significantly outperforms the baseline models in F1-score, recall, and precision metrics. Those results highlight the advantages of combining transformerbased embedding and recurrent architectures in phishing detection, providing avenues towards a more robust framework for email security enhancement.