Developing A Deep Learning Based Phishing Website Detection Model Master’s Thesis

Abstract

Phishing attacks continue to be a serious risk to online security since attackers are using more advanced techniques to trick users and steal private data. Conventional detection techniques such as whitelist, blacklist, and rule-based, are unable to keep up with the constantly changing phishing tactics. Existing researches often suffer a problem of class imbalance and poor model generalization because it relies on single dataset, which make it difficult for accurately detecting phishing websites. In this study by using a diverse of datasets and cutting-edge deep learning algorithms we overcomes these constraints. The aim of this study is to develop a Deep Learning based phishing website detection model. To achieve this objective, we used an experimental research approach. To improve detection accuracy, we present a novel hybrid technique that makes use of CNN-BiLSTM and CNN-RoBERTa architecture. We used data balancing techniques like SMOTE-Tomek, SMOTE, and RUS to address class imbalance. With an accuracy of 99.91 %, the result shows that hybrid deep learning model CNN-BiLSTM architecture outperform noticeable better than single models when trained on diverse dataset. Based on the result of the models, the hybrid approach of deep learning and transformer models trained on combined datasets the CNN RoBERTa model achieved an accuracy of 99.65%. While thorough preprocessing and optimization approaches help to increase generality and efficiency in phishing website detection, SMOTE-Tomek and other balancing techniques effectively address class imbalance. Considering the encouraging of the result of this investigation, it was highly suggested to real time detection mechanism if the hardware as better to plugin. This can entail looking into an application in IT services to keep the security of internet users in addition to blocking or warning users from phishing website attempts