Detecting Phishing Websites Using Hybrid Feature Extraction and Classification

Phishing websites are a huge cyber security threat because they deceive people into divulging sensitive information, most of the time without being detected by traditional cyber security measures. The basic problem with phishing is that it can get very dynamic and change sometimes, with lexical tricks, bad content, or the deceiving characteristics of hosting, making it rather difficult to determine correctly and fairly if it's phishing or not. This research aims to present a hybrid extraction and classification framework to detect the occurrences of a phishing containing lexical, content-based, domain level (hybrid pattern) and behavioral features. Specifically, this research entails data pre-processing from the UCI Phishing Websites Dataset, hybrid feature engineering with dimensionality reduction, as well as ensemble methods, which include Random Forest, Gradient Boosting, Support Vector Machines, and a stacking meta-classifier. The quantitative results that are presented show that the hybrid model is 98.2% accurate compared to the current state-of-the-art methods such as CNN-LSTM and deep neural networks. Hybrid methods show a great impact in the accuracy in detection and still effective on emergent threats. In summary, the presented research identifies scalable and practical framework for real-time detection in regards to phishing and future work will be focused on adaptive deep learning models and explainable AI.