MM-TFDX: a multimodal graph neural network framework for dynamic cross-border tax fraud detection with regulation-aware explainability

Cross-border tax fraud exploits jurisdictional asymmetries and complex multi-entity structures, costing governments hundreds of billions of dollars annually. Existing detection methods either ignore the relational context of trade networks or fail to integrate heterogeneous data sources and temporal dynamics. We propose MM-TFDX, an end-to-end multimodal graph neural network framework that jointly addresses three challenges: heterogeneous data fusion, continuous-time graph modeling, and domain-grounded explainability. The framework encodes tabular transaction features, textual customs declarations, and temporal event sequences through modality-specific sub-networks, merging them via adaptive gating and cross-modal attention within a heterogeneous graph structure. A continuous-time memory module captures fine-grained temporal signatures without snapshot discretization, while a regulation-aware explanation generator produces subgraph-level rationales anchored to predefined regulatory risk templates. Experiments on two cross-border trade datasets show that MM-TFDX achieves F1 scores of 0.891 and 0.862, outperforming seven baselines spanning traditional classifiers, static graph neural networks, and dynamic graph methods. Ablation analysis confirms that temporal modeling contributes the largest individual gain, followed by multimodal fusion, and that the explanation module simultaneously improves interpretability and detection accuracy through implicit structural regularization.