Privacy-Preserving Federated Fraud Detection in Payment Transactions with NVIDIA FLARE
Authors: Holger R. Roth, Sarthak Tickoo, Mayank Kumar, Isaac Yang, Andrew Liu, Amit Varshney, Sayani Kundu, Iustina Vintila, +13 more
Organizations: NVIDIA, Santa Clara, USA · J.P. Morgan, New York, USA · DeepTempo, San Francisco, USA · Royal Bank of Canada, Toronto, Canada · Bank of New York, New York, USA
Fraud-related financial losses continue to rise, while regulatory, privacy, and data-sovereignty constraints increasingly limit the feasibility of centralized fraud detection systems. Federated Learning (FL) has emerged as a promising paradigm for enabling collaborative model training across institutions without sharing raw transaction data. Yet, its practical effectiveness under realistic, non-IID financial data distributions remains insufficiently validated. In this work, we present a multi-institution, industry-oriented proof-of-concept study evaluating federated anomaly detection for payment transactions using the NVIDIA FLARE framework. We simulate a realistic federation of heterogeneous financial institutions, each observing distinct fraud typologies and operating under strict data isolation. Using a deep neural network trained via federated averaging (FedAvg), we demonstrate that federated models achieve a mean F1-score of 0.903 - substantially outperforming locally trained models (0.643) and closely approaching centralized training performance (0.925), while preserving full data sovereignty. We further analyze convergence behavior, showing that strong performance is achieved within 10 federated communication rounds, highlighting the operational viability of FL in latency- and cost-sensitive financial environments. To support deployment in regulated settings, we evaluate model interpretability using Shapley-based feature attribution and confirm that federated models rely on semantically coherent, domain-relevant decision signals. Finally, we incorporate sample-level differential privacy via DP-SGD and demonstrate favorable privacy-utility trade-offs...
Figures & tables
Figure 1: Federated learning for collaborative fraud detection.
Figure 2: Data distributions across sites. A, Site-level fraud rates and aggregate amount statistics. B, Amount distributions for normal vs fraudulent transactions. C–D, Anomaly-type distributions for training and evaluation partitions.
Algorithm
Mean F1-score
Central
0.925
FedAvg
0.903
FedOpt
0.905
FedProx
0.903
Local
0.643
Table 1: Mean binary F1 after 20 rounds across all clients and fraud types.
Figure 3: Central vs. Local vs. FL: convergence and precision–recall performance. A, Mean binary F1-score across clients and anomaly types over federated rounds (shaded bands: s.d. across clients). B, Precision–recall curves evaluated across client test sets for each checkpointed model; solid lines show the mean and bands indicate ± 1 s.d. across clients. The no-skill baseline is included, and AUPRC values are reported in the legend.
Figure 4: Confusion matrices of the final FedAvg global model for each participating bank.
Figure 5: Feature importance plot based on Shapley-value–based attribution (Captum GradientShap).
Figure 6: FedAvg vs. DP-FedAvg: F1-scores for all fraud types averaged across clients. The bands indicate the std. dev. of the metrics.
Algorithm
Privacy
Mean F1-score
FedAvg
no DP
0.920
FedAvg
ϵ=50.0
0.889
FedAvg
ϵ=10.0
0.879
FedAvg
ϵ=1.0
0.791
Table 2: Mean F1-score after Round 100 with DP-SGD. Mean binary F1 across all clients and fraud types for different privacy budgets.
Anomaly type
Description
Example
1: Location-based
Significant geographical deviation between expected and actual transaction location
Expected location: 500 miles from NYC; origin: Azerbaijan
2: Account-age-based
High-value transactions from disproportionately new accounts
Account age: 5 hours; amount: $15,000
3: Inactivity-based
Unusual transactions following extended periods of account inactivity
Last activity: 6 months ago; sudden high/low-value transaction
4: Unusual activity for account type
Excessive transaction frequency over a short time window
50 transactions in 10 minutes for a CHECKING or SAVINGS account
Table 3: Anomaly types used across client sites. Structured anomaly families injected into the synthetic payment dataset.
Feature
Type
Description
Data type
FRAUD_FLAG
Target
Indicates whether the transaction is anomalous (historically labeled FRAUD ).
int
DEBITOR_GEO_LATITUDE
Type 1
Latitude of debitor’s expected location.
float
DEBITOR_GEO_LONGITUDE
Type 1
Longitude of debitor’s expected location.
float
DEBITOR_TOWER_LATITUDE
Type 1
Latitude of cell tower used in transaction.
float
DEBITOR_TOWER_LONGITUDE
Type 1
Longitude of cell tower used in transaction.
float
PAYMENT_INIT_TIMESTAMP
Type 2
Timestamp when payment was initiated.
datetime
Table 4: Features of interest and associated anomaly type. Only a subset of the 60+ features is shown (most relevant for the anomaly rules).
School of Computer Engineering, KIIT Deemed to be University, Bhubaneswar, India · School of Electronics Engineering, KIIT Deemed to be University, Bhubaneswar, India
Department of Computer Science, University of Bucharest · Luxembourg Institute of Science and Technology (LIST), Luxembourg · Research Institute of the University of Bucharest (ICUB), Bucharest, Romania
Department of Advanced Computing Sciences, Maastricht University, Netherlands · Department of Cognitive Science and Artificial Intelligence, Tilburg University, Netherlands