This paper presents research on graph fraud detection using Graph Neural Networks (GNNs), which are effective in modeling complex relationships within multimodal data. Specifically, we highlight that existing graph fraud detection methods ignore the rich semantic cues in textual information. Therefore, we propose MLED, a novel framework that utilizes large-scale language models (LLMs) to process textual information and fuse it with graph structures. MLED improves graph fraud detection by extracting external knowledge from textual information using LLMs. We design a multi-level LLM-based framework that includes type-level enhancers and relation-level enhancers to enhance the distinction between fraudsters and legitimate entities and emphasize the importance of fraudsters in various relationships. Experimental results demonstrate that MLED achieves state-of-the-art performance on four real-world datasets, outperforming existing methods.
