Explainable and Adversarially Robust Meta-Learning Framework for Real-Time Phishing Detection

Phishing is still a large threat to cybersecurity that uses user deception via realistic email messages and imitated web sites to get access to critical information. Blacklist and Rule Based forms of protection have historically been adequate for this type of attack; however, they now fall short of the mark and there has been difficulty in supporting independent machine learning models as they do not generalize well to detect sites that are highly similar to known good sites. In response to these issues, this research presents a meta-learning (stacked) approach that employs multiple models, to include Artificial Neural Networks, K-Nearest Neighbors, and a Logistic Regression-based Meta-Classifier (i.e., Meta-Classified Model), as well as a Convolutional Neural Network to analyze the visual content of a webpage; while using adversarial training to increase robustness against future attacks. In addition, the use of SHAP will provide for model explainability. The experimental results on phishing datasets (2019 – 2025) yielded impressive results with accuracies reaching 99% and exhibiting high precision as well.