Dynamic Multi-Stage Phishing Detection Using Machine Learning and LLMs

As phishing attacks become more sophisticated, conventional detection methods are easily bypassed. In particular, the widespread use of Large Language Models (LLMs) has made creating and deploying phishing attacks simple. However, LLMs can also be used to detect and prevent these attacks. This paper introduces a multi-stage phishing detection system that integrates Machine Learning (ML) with multiple LLMs to improve detection performance and efficiency. The proposed system uses a robust and efficient pipeline to handle multimodal inputs for an initial high-speed URL assessment, which can result in a thorough multimodal assessment depending on the ML-based risk analysis. Experimental results using well-known datasets and sites showed that our method achieves $99.0 \%$ recall, outperforming existing approaches. Moreover, while conventional and existing approaches only provide a binary judgment, the proposed system can cost-effectively provide a detailed diagnostic report.