Implementation of a Hybrid Autoencoder-CatBoost Model for Real-Time Phishing Site Detection Based on URL and HTML Analysis

Aziz, Muhammad Hilmy (2026) Implementation of a Hybrid Autoencoder-CatBoost Model for Real-Time Phishing Site Detection Based on URL and HTML Analysis. Undergraduate thesis, UPN Veteran Jawa Timur.

	Text (Cover) 22081010324.-cover.pdf Download (3MB)
	Text (Bab 1) 22081010324.-bab1.pdf Download (1MB)
	Text (Bab 2) 22081010324.-bab2.pdf Restricted to Repository staff only until 4 September 2028. Download (8MB)
	Text (Bab 3) 22081010324.-bab3.pdf Restricted to Repository staff only until 4 September 2028. Download (9MB)
	Text (Bab 4) 22081010324.-bab4.pdf Restricted to Repository staff only until 4 September 2028. Download (19MB)
	Text (Bab 5) 22081010324.-bab5.pdf Download (523kB)
	Text (Daftar pustaka) 22081010324.-daftarpustaka.pdf Download (1MB)

Abstract

Phishing is one of the dominant cybersecurity threats involving fraudulent attempts to obtain users' sensitive information by imitating the appearance and flow of legitimate services. Conventional detection methods such as blacklisting and whitelisting are static in nature and therefore struggle to recognize new (zero-day) phishing attacks. This research aims to build a hybrid Autoencoder-CatBoost model to detect phishing websites in real time based on URL and HTML feature analysis, to evaluate its performance using classification evaluation metrics, and to implement it as a browser extension. The Autoencoder is used to learn a latent representation from 49 features (31 HTML features and 18 URL features), which are subsequently classified using the CatBoost algorithm. Testing results show that the model configuration with a latent dimension of 36 and an 80:10:10 data split ratio provides the best balance between accuracy and computational efficiency, achieving an accuracy of 94.68%, an F1-score of 93.53%, and an inference time of 0.4445 ms, or about 2.2 times faster than the baseline CatBoost model. The selected model was successfully implemented as a Chromium-based browser extension integrated with a backend API, and all black-box testing scenarios were declared successful. These results indicate that the hybrid Autoencoder-CatBoost model can provide an accurate, adaptive, and efficient phishing detection solution suitable for real-time deployment.

Item Type:

Thesis (Undergraduate)

Contributors: