Optimized BiLSTM-Based Phishing Detection System Using Attention Mechanism for Large-Scale Data Classification

Abstract

Phishing attacks have become one of the most dominant cyber threats, contributing significantly to global data breaches and financial losses. Traditional detection mechanisms, including rule-based filters and classical machine learning models, are increasingly ineffective against evolving phishing strategies. This research proposes an optimized deep learning framework based on Bidirectional Long ShortTerm Memory (BiLSTM) integrated with a Bahdanau Attention Mechanism for large-scale phishing email detection. The system is trained on a consolidated dataset of 13,565 real-world emails obtained from four independent sources, ensuring diversity and robustness. A comprehensive preprocessing pipeline involving text normalization, tokenization, and semantic feature extraction enhances input quality. The BiLSTM architecture captures contextual dependencies in both forward and backward directions, while the attention layer emphasizes critical phishing indicators within the text. Experimental results demonstrate superior performance with an accuracy of 98.2% and an F1-score of 97.8%, outperforming baseline models such as Naive Bayes, Support Vector Machines, and conventional LSTM networks. Additionally, threshold optimization improves classification balance. The proposed system provides a scalable and efficient solution for realtime phishing detection in large-scale environments, contributing to enhanced cybersecurity measures.