Adaptive Sequence-Embedded Gradient Boosting for Real-Time Phishing Website Detection

Abstract

Phishing threats keep changing, exploiting the properties of static URLs and the dynamic aspects of the website to defraud victims, making the traditional methods of detection inadequate. In this paper, a new hybrid phishingdetection model that combines sequence-aware embeddings with Gradient Boosting is proposed, which outperforms in terms of predictive accuracy and flexibility. Embedded scripts and URLs are coded as sequential patterns by a lightweight embedding layer, and structural and temporal patterns are recovered, which are not usually used by feature-based methods. The embeddings are then fused with traditional features, such as domain attributes, SS/TLS certificate metadata, and HTML indicators, resulting in a complete representation of the behaviour of websites. An adaptive Gradient Boosting classifier is applied to the integrated feature set and ranks features according to their dynamically changing usefulness to overcome already developed phishing techniques. Experimental analysis on various real-life datasets, the classification accuracy is above 99%, outperforming conventional methods such as XGBoost, Random Forest, and SVM methods, while also maintaining low false-positive rates. The feature importance analysis offers interpretability for cybersecurity professionals, enabling them to take appropriate actions accordingly. The proposed system can be deployed at scale in browser extensions, enterprise security systems, and automated threat intelligence platforms, providing proactive protection against increasingly sophisticated phishing attacks.