Privacy-Preserving Financial Fraud Detection Using Federated Heterogeneous Temporal GraphSAGE
Digital financial ecosystems face mounting exposure to fraudulent transactions that collectively account for trillions of dollars in losses each year. Existing approaches suffer from three recurring deficiencies: they represent all transaction participants within a single undifferentiated node space, rely on fixed decision boundaries incapable of accommodating evolving fraud distributions, and fail to exploit the semantic diversity among entity categories including accounts, merchants, and transaction types. This study introduces FHT-GraphSAGE, a federated heterogeneous temporal graph learning framework for privacy-preserving financial fraud detection. The architecture combines heterogeneous temporal graph construction, a relation-aware GraphSAGE encoder with sinusoidal temporal edge embeddings, and a Federated Averaging optimization scheme that enables cross-institutional collaborative learning without exposing raw transaction records. Evaluation across two complementary benchmarks, the PaySim Mobile Money Dataset and the Credit Card Fraud 2023 Dataset, demonstrates consistent superiority over ten competitive baselines. On PaySim, FHT-GraphSAGE achieves an accuracy of 0.963, an F1-score of 0.939, an AUC-ROC of 0.986, and, most importantly under severe class imbalance, an AUC-PR of 0.724. On the Credit Card Fraud 2023 Dataset, it attains an accuracy of 0.941, an F1-score of 0.913, an AUC-ROC of 0.969, and an AUC-PR of 0.721. As AUC-PR is the most informative metric for minority-class detection under extreme imbalance, these two figures (0.724 and 0.721) constitute the primary evidence of the framework’s fraud-detection capability. Results are reported as the mean over five independent runs with distinct random seeds, and the improvements over the strongest baseline are statistically significant (p < 0.01). Ablation experiments confirm the non-redundant contribution of each architectural component, and robustness evaluations demonstrate stable minority-class detection performance down to a fraud ratio of 0.25%.