Permutation-Based Feature Selection with Multi-Model Deep Learning for Phishing Attack Identification

Abstract

 Phishing is a serious danger to safety. It is employed to obtain confidential data from people by luring them into divulging it to a bogus site. In this article, I will be explaining how ML can be leveraged to identify fake emails. It also possesses DL, advanced feature selection and explainable intelligence. FFNN, DNN, TabNet, and Wide and Deep models all use a variety of ways to figure out which traits of phishing are the most important. These include measures of association such as chi-squared analysis, mutual information and permutation importance. To improve the stability of the system an ensemble vote classifier is used with the RF and Bagging with the DT. The SMOTEENN resampling method is used to make the classes more equal. It achieved a 98.7% success rate in the phishing websites dataset and 98.5% success rate in web page phishing dataset, which was more accurate than the individual DL models.The results are based on two public phishing experiments. To explain predictions and show the effect of features like LIME and SHAP, methods like Explainable Artificial Intelligence are used. That's why they are helpful to be open and to build trust. In the real world it is used on the Flask platform in this framework. It provides a safe way for users to register and log in, as well as an online, interactive interface. Users can type in any URL to be scanned and the system will immediately inform them if the URL is a phishing or not. This can lead to a more reliable identification of phishing in online security and making it easier to understand.