Indian Journal of Engineering Research Networking and Development (Online ISSN:3048-7676)

Phishing has become one of the most pervasive cyber threats, evolving from straightforward deceptive emails to complex social engineering attacks that exploit human trust and sophisticated obfuscation techniques. Existing detection mechanisms, such as rule-based or blacklist-oriented systems, often fail against zero-day attacks and cleverly disguised messages. To address this, we propose a Transformer- based Natural Language Processing (NLP) approach for real-time phishing detection that analyzes textual and semantic features of emails, messages, and web content. Our methodology leverages pre- trained language models such as BERT and RoBERTa, fine-tuned for phishing classification tasks. By capturing contextual embeddings and semantic intent, the model identifies phishing attempts that traditional techniques  often  miss.  Experimental evaluation demonstrates  an  accuracy  of  98.4%, surpassing conventional SVM, Naïve Bayes, and RNN-based classifiers. The proposed framework is designed  for  real-time  integration  with  email  servers  and  web  browsers,  offering  scalable, adaptive, and practical cybersecurity defense.