This article addresses the issue of financial fraud by integrating Graph Attention Network (GAN) into graph neural networks. It combines Node Attention Networks and Semantic Attention Networks to form a Dual-Head Attention Network module, enabling comprehensive analysis of complex relationships in user transaction data. The model incorporates Gradient-Boosting Decision Tree (GBDT) to enhance fraud identification, resulting in the GBDT-Dual-channel Graph Attention Network (GBDT-DGAN). To ensure user privacy, blockchain technology is introduced, culminating in a financial anti-fraud model that integrates blockchain with the GBDT-DGAN algorithm. Experimental results demonstrate the model's accuracy at 93.82%, a significant improvement over baseline algorithms like Convolutional Neural Networks (CNN). The recall and F1 scores are 89.5% and 81.66%, respectively, and the model exhibits superior network data transmission security, maintaining a packet loss rate below 7%. The proposed model offers an efficient and secure solution for financial fraud detection, improving accuracy and ensuring data privacy.This article addresses the issue of financial fraud by integrating Graph Attention Network (GAN) into graph neural networks. It combines Node Attention Networks and Semantic Attention Networks to form a Dual-Head Attention Network module, enabling comprehensive analysis of complex relationships in user transaction data. The model incorporates Gradient-Boosting Decision Tree (GBDT) to enhance fraud identification, resulting in the GBDT-Dual-channel Graph Attention Network (GBDT-DGAN). To ensure user privacy, blockchain technology is introduced, culminating in a financial anti-fraud model that integrates blockchain with the GBDT-DGAN algorithm. Experimental results demonstrate the model's accuracy at 93.82%, a significant improvement over baseline algorithms like Convolutional Neural Networks (CNN). The recall and F1 scores are 89.5% and 81.66%, respectively, and the model exhibits superior network data transmission security, maintaining a packet loss rate below 7%. The proposed model offers an efficient and secure solution for financial fraud detection, improving accuracy and ensuring data privacy.