Fraud detection aims to identify fraudsters by analyzing vast behavioral data. Methods based on graph neural networks (GNNs) have demonstrated outstanding performance in modeling user interactions. However, their aggregation properties are highly susceptible to disruption by fraudsters through camouflage techniques. These fraudsters disguise their identities by modifying personal information and creating fake connections to obscure abnormal behaviors, making detection difficult. Considering the complexity of real-world interactions, we model from a multi-relational graph perspective. To tackle these challenges, we propose SETE-GNN, a fraud detection model based on Semantic Extraction and Topological Enhancement. First, SETE-GNN incorporates the semantic extraction module that preserves node features, distances, and relational group information. By employing a transformer-based encoder, the model effectively captures the preference of grouped features. Subsequently, we introduce a GNN with label-aware information propagation to capture global structural patterns and adapt the message passing mechanism between source and target nodes. The representations from both modules are then combined and fed into detector. Finally, to improve the model’s generalization ability, we additionally integrate the representation enhanced co-training module. Experiments on two datasets demonstrate that SETE-GNN outperforms other competitive fraud detection models.