Development of a hybridized CNN-BiGRU Framework for detection of website phishing attacks

Abstract

Phishing remains a major cybersecurity challenge, with attackers using deceptive tactics to trick users into disclosing confidential data. Traditional detection systems, which often rely on fixed features or predefined rules, struggle to keep up with rapidly evolving phishing strategies. This research introduces a deep learning-based solution that combines Convolutional Neural Networks (CNN) and Bidirectional Gated Recurrent Units (BiGRU) to improve phishing website detection. The CNN component is responsible for learning spatial patterns from web data, while the BiGRU layer captures sequential relationships, providing a more complete understanding of the underlying threats. The framework involves meticulous preprocessing steps such as data cleaning, normalization through MinMax scaling, and optimal feature selection using the SelectKBest, CNN and BiGRU methods. The model was trained and tested on large scale, publicly available datasets from IEEE Data Port and Mendeley, consisting of over 250,000 URL entries. Through train-test split and cross-validation techniques, the model consistently achieved outstanding results: 99.96% accuracy, 99.92% precision, 100% recall, and a 99.92% F1 score. When compared to existing solutions, this hybrid approach sets a new performance benchmark, underscoring the power of combining spatial and temporal deep learning methods in defending against phishing threats.