Generative AI for synthetic data in banking transactions: Balancing utility and compliance

Data scarcity in regulated banking sectors often limits the training of machine learning models for fraud detection, risk assessment, and transaction pattern analysis. This paper explores the use of generative AI for producing high-fidelity synthetic banking transaction datasets that maintain statistical fidelity while guaranteeing privacy preservation and regulatory compliance. The approach introduces a hybrid loss function combining Wasserstein distance with privacy leakage penalties, ensuring optimal trade-offs between realism and compliance with banking regulations including PCI DSS, GDPR, and PSD2. Anomaly injection techniques are incorporated to improve the robustness of downstream fraud detection models in rare-event prediction tasks. The framework is validated on synthetic payment transaction datasets from major banking institutions, achieving 94% downstream model performance retention while passing rigorous privacy audits and regulatory compliance assessments. The research presents three novel contributions: a new hybrid loss function balancing statistical fidelity and privacy leakage constraints specifically designed for financial transaction data, anomaly injection methodologies for improving rare-event fraud detection, and integrated regulatory compliance auditing within generative pipelines. Experimental validation demonstrates significant improvements in fraud detection accuracy while maintaining strict compliance with financial industry regulations and privacy requirements.