PHISHING DETECTION WITH SELF-ATTENTION AND NEURAL ODE LAYERS USING A TENSORFLOW-BASED DEEP LEARNING APPROACH

Phishing remains the largest cyber security problem with which we have to put up with today and is one of the most common means under which users are deceived into providing their personal information (e.g., usernames, passwords, and credit card) because it can be cheaply mass-produced. The paper proposes an effective deep learning-enabled phishing detection model, which jointly applies self-attention modules and re-layered evaluated neural ordinary differential equation (ODE) blocks to achieve customizable traditional phishing detection systems. Multithread Self-Attention mechanism allows the model to attend which are the most relevant features within complex datasets, and capture elaborate relationships between categorical and numerical features. The use of Neural ODEs adds the possibility to model in continuous time, having the system getting information on dynamic and changing patterns of phishing attempts. The architecture also uses more elaborate preprocessing layers including normalization of numeric features and one-hot encoding pattern for categories, providing a powerful means to represent different sorts of data. Regularization techniques such as the weighted activation scaling (SELU) and Alpha Dropout add to the stability of the learning process by preventing over fitting. The model, which has been built in Tensor Flow, outperforms traditional models (e.g., KNN, logistic regression and SVM) with an accuracy of 96.1% and AUC = 0.99. These results demonstrate that the model generalizes well on unseen samples, which would be a scalable robust solution for phishing detection. This work contributes to advancements in cyber security by showing that self-attention and Neural ODEs can be usefully combined to meet the demands of an ever-evolving crime.