Real-time browser-integrated phishing uniform resource locator detection via deep learning and fuzzy matching

Abstract

Phishing attacks through deceptive URLs remain a critical cybersecurity threat, particularly in financial transactions and online payment systems. This study evaluates multiple deep learning (DL) models on the PhiUSIIL dataset of 235,795 URLs, with bidirectional gated recurrent unit (BiGRU) achieving the best performance—99.82% accuracy at a 60:40 split, along with high precision, F1-score, and the lowest test loss. To further improve detection of obfuscated URLs, an enhanced BiGRU variant is proposed using an expanded 366-character vocabulary. For real-time deployment, a Chrome extension is developed, integrating exact and fuzzy matching via the Ratcliff–Obershelp algorithm with cloud-based whitelist and blacklist checks. When fuzzy matching is inconclusive, the BiGRU model performs the final classification. By combining an adaptive browser-side tool with a robust DL backend, the proposed system ensures high accuracy, scalability, and efficiency for phishing detection in practical web environments.