Privacy-Preserving Financial Fraud Detection: Challenges and Solutions with Generative Models, Lifetime-Aware Detection, and Federated Boosting

While privacy regulations prohibit direct data sharing among institutions, improving fraud detection performance requires collaboration across banks. To mitigate this limitation, we have conducted a real-world case study on privacy-preserving financial fraud detection (FFD) in the South Korean banking sector. During the research, we have identified four major challenges in practice: (C1) the degradation of tabular generative models under extreme class imbalance and sparsity, (C2) the lack of utility–privacy joint evaluation methodology, (C3) the inability of detection models to capture irregular active lifetime of fraudulent activity, and (C4) the absence of robust federated gradient boosting under dynamic participation. In this work, we introduce two novel approaches: (i) Graph-theoretical Generative Models (GGMs), which leverage graph theories to generate high-utility synthetic tabular data; and (ii) Active Lifetime-Aware Fraud Transaction (ALAFT), which adjusts fraud scores by defining and modeling active lifetime of fraudulent patterns. Across two private banking datasets and a public benchmark, GGMs consistently outperform seven baselines, while ALAFT outperforms significant gains over six representative detectors, reducing false positives during high-risk periods. Finally, we outline our ongoing work, fraud scenario-aware and similarity-based FedXGBBagging with KakaoBank, TossBank, and KBank to enable secure collaboration and support nationwide anti-fraud efforts.