LLMAntiPhish: A Hybrid Intelligence Framework for Real-Time Phishing URL Detection Using Multi-modal Deep Learning and Large Language Models

The escalating sophistication of phishing attacks threatens the security of ubiquitous multimedia systems and advanced communication networks, necessitating defense mechanisms that transcend traditional rule-based methods. To address this critical challenge at the intersection of cybersecurity and cognitive computing, this paper introduces LLMAntiPhish, a novel hybrid intelligence framework that synergizes multi-modal deep learning with Large Language Models (LLMs). Our architecture integrates a Bidirectional LSTM (BiLSTM) with attention mechanisms for deep URL sequence analysis, comprehensive structural feature engineering, and LLM-based contextual reasoning for enhanced explainability. The system employs a sophisticated attention-based fusion network that processes BiLSTM probabilities, rule-based scores, and LLM predictions in parallel before feature fusion and final classification. This core detection engine is augmented by real-time threat intelligence from PhishTank and OpenPhish, creating a multi-layered defense suitable for securing wireless, mobile, and ubiquitous multimedia environments. Extensive evaluation on a real-world dataset of 105 URLs demonstrates the framework’s superior performance, achieving 98.10% accuracy, 95.83% recall, and 97.87% F1-score, significantly outperforming individual component models and traditional ensemble approaches. The efficient neural fusion architecture, with attention-based feature weighting, learns optimal combinations of detection modalities automatically. LLMAntiPhish represents a substantial advancement in phishing detection, demonstrating that hybrid cognitive systems can effectively balance state-of-the-art performance with the interpretability required for trustworthy deployment in modern Ubi-Media infrastructures.