GNN's for financial fraud detection

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Financial fraud, including stolen-card use, identity theft, fake loan applications, and organised group fraud, causes large economic losses and weakens trust in digital financial services. Detectors that score each transaction or customer independently miss the relationships through which fraudsters operate: shared phone numbers, addresses, devices, and coordinated activity. Graph Neural Networks (GNNs) address this limitation by representing customers, accounts, or transactions as graph nodes and learning from both their features and their connections, which turns fraud detection into node classification. This paper reviews the fundamentals, architectures, and limitations of GNN-based fraud detection and critically analyses four recent studies: GraphFA (industrial loan fraud with edge-level camouflage detection), a GCN/GAT/RGCN study on credit-card transactions, FFDM-GNN (IEEE-CIS data), and the proposed real-time Autonomous GNN (AGNN). The strength of evidence differs sharply. GraphFA is supported by a large real dataset, gradient-boosting baselines, and an ablation study, whereas the other studies show probable data leakage after oversampling, conclusions that contradict their own tables, or performance claims without experimental detail. A conceptual framework is outlined that combines leakage-safe data preparation, justified graph construction, camouflage-aware edge filtering, an imbalance-aware relational GNN, strong baselines with honest metrics, and explainability with monitoring. The study is literature-based; the framework has not been implemented or tested