DEVELOPMENT OF A TEXT-BASED MODEL FOR DETECTING AND PREVENTING PHISHING ATTACKS USING DEEP LEARNING

Abstract

Phishing attacks are cyber threats that have persisted over the years, and are often disguised as legitimate messages. Different studies have been focused on the detection of attempted phishing attacks, and significant milestone has been reached. However, previous studies rely mostly on URL-based phishing detection, with machine learning and deep learning models, but limited studies have focused on text-based phishing detection, and adoption of hybrid deep learning and transformer based models. Addressing these gaps, this study focused on the development of a text based phishing detection modes using a hybrid of CNN-LSTM with attention mechanism, and a transformer-based BERT model. Deep learning based models were train separately on a publicly available dataset collected from Kaggle. The dataset was pre-processed and used to train the two models. The model was evaluated using standard evaluation metrics. The experimental result shows that the CNN-LSTM-Attention model outperformed the BERT model across all evaluation metrics used, achieving  92.77% accuracy, 93.00% precision, 92.00% recall, 92.00% F1-score, and 93.00% AUC, while the BERT model had average performance with 76.55% accuracy, 74.00% accuracy, 75.00% precision, 73% recall, 75.00%, F1-scores, and 81.70% AUC. This study was compared with the baseline study that used the K-Nearest Neighbour (KNN) on URL-based features. Our study, CNN-LSTM-Attention model demonstrated superior performance. This study shows that hybrid deep learning with attention mechanism approach is highly effective in the detection of phishing attacks. Future research can focus on exploring multilingual capabilities, integratingbehavioural feature, and real-time model deployment to enhancecyber security.