Enhanced Phishing Email Classification through Multi-Level Feature Fusion

Abstract

Phishing emails remain one of the most common cyber threats. It causes serious financial damage, information theft, and security vulnerabilities for individuals and organizations. Many existing detection systems depend on traditional feature extraction methods or private datasets, which limit their effectiveness and reduce reproducibility in real-world applications. To overcome these challenges, this study proposes a hybrid text classification framework for phishing email detection by integrating TF-IDF, Word2Vec, and BERT-based contextual embeddings. This combined approach captures lexical importance, semantic relationships, and contextual meaning within email content. The model is trained and evaluated using a large publicly available phishing email dataset to ensure reliability and practical relevance. Experimental findings show that the proposed approach achieves an accuracy of 99.23%. The proposed model demonstrates improved performance compared to conventional single-feature models. The developed system provides a robust, scalable, and efficient solution for phishing email classification and supports the development of real-time applications for strengthening email security against evolving phishing attacks.