PhisNet: Intelligent Detection of Phishing

Abstract

Phishing and fraud attacks continue to be common in the cybersecurity environment, as criminals use URLs and email messages. Here we conduct a side-by-side evaluation of two transformer-based machine learning techniques to identify phishing. BERT-LSTM model that focuses on spotting email phishing. Combined RoBERTa and Attension model that aims to detect URL phishing. With email phishing detection, the proposed BERT-BiLSTM model with an attention mechanism achieved 98.7% accuracy by efficiently utilizing linguistic metadata and structural properties of emails that extract and combine discrim- inative content from emails and focus on key details required to complete the classification. So, for detecting the URL, the hybrid- RoBERTa model was achieved 93% accuracy. On the other hand, confirming our hypothesis that semantic patterns in URLs are crucial to detection. Furthermore, it should be recognized that transformer models outperformed all traditional machine learning models in every domain, exhibiting incredible recall superiority for advanced phishing strategies. Further analysis of feature importance indicated URL entropy and email sentiment features as the prominent discriminators. These results lay down the foundation for layered active systems to thwart phishing attacks by guiding the implementation of RoBERTa hybrids for web traffic filtering and a motion-controlled BERT-LSTM operation.