Abstract

This paper develops a fraud detection solution using graph neural networks (GNN) enhanced by large language models to address the increasingly severe issue of phishing fraud on the Ethereum platform. As new phishing attack methods continue to emerge, posing significant threats to user financial security, traditional detection models, often limited to superficial feature matching, struggle to effectively counter these complex and variable fraud patterns. This study innovatively integrates large language model dynamic predictions with graph neural networks to establish a multi-layered, high-dimensional predictive framework. Leveraging the strong semantic understanding and generative capabilities of large language models, it provides a dynamic and diverse perspective on graph structures. Within this framework, the enhanced GNN delves deeper into the hidden patterns of the network, effectively distinguishing between normal transactions and fraudulent activities. Experimental results demonstrate that the system proposed in this paper surpasses existing fraud detection models in terms of detection accuracy, recall rate, and adaptability to new fraud patterns.