FraudGNN: Self-Supervised Graph Neural Anomaly Detection for Real-Time Financial Fraud with Adversarial Robustness and Explainable Reasoning

Financial fraud detection systems must operate at a very large scale, under extreme class imbalance, and with strict constraints on false alarms, where small changes in attacker behavior can quickly reduce model reliability. Recent research has shown that graph learning and self-supervised objectives can improve anomaly detection by capturing relations between transactions and shared identifiers, but many practical pipelines still struggle to jointly achieve strong recall at low false-positive rates, stable performance under feature manipulation, and efficient inference for real-time screening. In this paper, we propose FraudGNN, a relation-aware graph neural network that converts transaction records into a transaction-entity heterogeneous graph and learns fraud scores using a combined objective that includes (i) weighted supervised learning for imbalanced labels, (ii) contrastive self-supervised learning across augmented graph views to improve label efficiency, and (iii) adversarial training to improve robustness to small feature perturbations. Experiments on the IEEE-CIS Fraud Detection benchmark show that FraudGNN outperforms strong tabular and graph baselines, achieving an AUC-ROC of 0.915 and an AUC-PR of 0.862, with Recall@1%FPR reaching 0.607 while also improving calibration (Brier score 0.024). Robustness evaluation further shows that FraudGNN degrades more slowly under increasing perturbation strength, maintaining AUC-PR of 0.832 at ϵ = 1.0. Ablation results confirm that self-supervision, adversarial training, and heterogeneous relation modeling each contribute to the final performance. Overall, FraudGNN provides a practical, accurate, and robust fraud detection framework that aligns with real-world deployment requirements.