Cross-Bank Fraud Detection, Federated Learning + Mixture-of-Experts
Mar 2026 to present
Four-person UTS research project on privacy-preserving fraud detection across simulated banks. I initiated and architected it, wrote the core federated engine and the triage layer, and am first author on the six-page conference manuscript. Mixture-of-Experts gates over per-bank gradient-boosted experts and FL-trained global models beat the FL family by +0.0115 AUPRC (p = 0.030).
- FL algorithms
- 6
- Datasets, 3 non-IID levels
- 5
- AUPRC over FL (p = 0.030)
- +0.0115
Bank 1 · Bank 2 · Bank Nlocal data onlyFederated clientsDP noise, gradient clippingAggregation serverFedAvg, FedProx, FedNova, SCAFFOLD, Per-FedAvg, FedAvgMGlobal model
Per-bank GBDT expertsXGBoost, LightGBM, CatBoostMoE gateCalibrated scoresTriageauto-clear, flag, defer to analyst
- Six FL algorithms across five transaction datasets (including IBM AML and synthetic AML) at three Dirichlet heterogeneity levels
- Differential privacy, gradient clipping and Byzantine-client filtering evaluated per algorithm
- Triage layer: per-expert calibration, per-bank conformal miss-rate guarantee, cost-optimal threshold, budgeted deferral under analyst capacity
- AUPRC instead of ROC-AUC because the fraud base rate is 0.17 percent (492 frauds in 284,807 transactions)
- Five-seed reproducibility with Wilcoxon signed-rank tests and an order-independent partitioning fix
- Retracted a headline result that failed to reproduce under the corrected pipeline rather than publishing it
- Python
- PyTorch
- Federated Learning
- Differential Privacy
- XGBoost
- LightGBM
- CatBoost
- Streamlit
