Because cross-border trade fraud involves multiple types of entities, multiple business relationships and complex interactive structures, it exhibits high heterogeneity and strong concealment, which has brought significant challenges to the traditional risk identification methods. Aiming at the problem that existing methods struggle to balance the ability of structural modeling and classification performance, this paper proposes a cross-border trade fraud detection framework based on heterogeneous graph neural network (HGNN) and gradient lifting tree model XGBoost. Firstly, the cross-border trade system is modeled as a heterogeneous graph of multi type entities and multi relationship interactions, and HGNN is used to learn the high-order structural semantic representation of entities in complex trade networks; Then, the graph embedding features and statistical features are input into XGBoost to achieve high-precision classification of fraud. The experimental results on the real cross-border trade data set show that the AUC of the proposed model on the test set reaches 0.966, which is 18.7% and 3.4% higher than using XGBoost and HGNN alone, and significantly improves the recall rate of fraud samples in a variety of typical fraud scenarios. Ablation experiments further verified the key role of heterogeneous relationship modeling, attention mechanism and integration strategy in performance improvement. The above results show that HGNN–XGBoost integration framework has good detection performance and engineering application potential in complex heterogeneous scenes.