Demo: PhishSense: A LLM-Enhanced Multimodal Framework for Phishing Website Detection

Phishing attacks pose persistent cybersecurity threats. In this paper, we propose PhishSense, a GPT-enhanced multimodal framework designed for detecting phishing websites. It integrates both visual and textual information through large language model (LLM) calls. Experimental results show that the LLM-augmented image branch reduces false negatives from 45.8% to 28.0%, while the LLM-augmented text branch lowers false negatives from 45.7% to 31.9%. Moreover, the multimodal approach improves overall detection precision by up to 75% compared to unimodal methods. These improvements enable our multimodal model to substantially outperform unimodal baselines, underscoring the potential of conditional LLM integration in improving phishing detection accuracy.