A BERT-Enhanced Framework for Semantic-Aware Phishing Email Detection and Real-Time Analysis

The problem of phishing emails is still one of the most serious cyber threats. It calls for smarter and more reliable detection mechanisms. This paper proposes a deep learning-based approach using Bidirectional Encoder Representations from Transformers for improving the understanding of contextual information within email text. The paper first constructs and preprocesses a phishing e-mail dataset, showing in detail the process of data cleaning and how to address class imbalance. In an attempt to establish a baseline, several traditional machine learning models, namely Logistic Regression, KNN, SVM, Random Forest and XGBoost were evaluated using TF-IDF features. Among these models, Logistic Regression and SVM achieved full precision, while Random Forest attained the maximum accuracy of 99.56%. Despite this strong performance, it is observed that TF-IDF cannot capture deeper semantic patterns which motivated shifting towards a transformer-based model. The BERT model demonstrated superior language understanding and achieved 99.47% accuracy and a precision score of 0.989. It was useful in practical real-time phishing e-mail detection to prepare the trained BERT model integrated into a RESTful API implemented in Flask and connected with a user-friendly web-based interface for live analysis of e-mails. Results show that though traditional models are effective, BERT’s contextual capabilities make it a more suitable choice for practical phishing email detection.