Implementation of a Hybrid Phishing Detection Platform Using Machine Learning and Google Safe Browsing API

Abstract

Phishing is a pervasive issue in cybersecurity, exploiting both technological weaknesses and human vulnerabilities to gain access to sensitive data. This dissertation introduces a hybrid phishing detection system using both a supervised Random Forest model and the Google Safe Browsing API to improve accuracy and adaptability to evolving attacks.

The dataset, consisting of 247,950 URLs, was processed using lexical, domain-based, and content features, and the Random Forest model was trained on an 80/20 stratified split. The framework employs a layered architecture including allowlist checking, API verification, machine learning classification, and rule-based heuristics, whose outputs are combined to produce a final decision.

Additionally, a web application was developed to provide real-time URL assessment and enhance user awareness through integrated educational features. Experimental results show that the API-only baseline achieved an ROC-AUC of 0.49, while the Random Forest model achieved an ROC-AUC of 0.993. The hybrid system significantly reduced false negatives while maintaining strong precision and recall.

These findings demonstrate that combining API intelligence, machine learning, and user-focused interventions provides a scalable and effective approach to phishing detection.