Federated Learning Could Let Banks Catch Fraud Without Sharing Customer Data

Banks spend billions every year hunting fraud and bad loans, and the most accurate detectors learn from as much data as possible. Belinda Unuigbokhai and six co-authors at Wellspring University in Benin City, Edo State, Nigeria, set out to show that a credit-risk model can be trained across several institutions without exchanging a single raw customer record. The result, published in the FUDMA Journal of Sciences in May 2025, applies federated learning to a loan-default problem on the public Lending Club dataset and reports an overall accuracy of 80% with 98% recall for defaulted loans.

The timing is regulatory as much as technical. The European Union's General Data Protection Regulation and comparable regimes elsewhere restrict how financial institutions may pool sensitive customer data, even when every bank in a consortium would benefit from a shared fraud model. Federated learning answers that bind: the model travels to the data, then only model updates come back to a central server. No raw records move.

For a customer, the promise is concrete. Your transaction history stays inside your own bank while the bank still contributes to, and gains from, a system trained on patterns seen across many institutions. That matters tomorrow morning because fraud and default risk do not respect institutional borders, yet privacy law forbids the simplest fix, which would be dumping everyone's records into one warehouse.

Where the paper sees the biggest gains

The authors frame the work around three uses of federated learning in finance, though each carries a distinct pressure point. In fraud detection, more institutions sharing learned patterns should mean better recognition of novel scams without exposing individual transactions. In anti-money-laundering compliance, the same mechanics let banks build detection models on diverse data, an edge when suspicious activity crosses banks and borders without breaching confidentiality.

Biometric authentication is arguably the most delicate of the three. A leaked password can be reset; a leaked fingerprint template cannot. The paper argues that federated learning strengthens these systems by decentralizing model training so biometric features never travel in raw form, which reduces the damage a single breach can do and keeps deployment aligned with privacy regulations.

How federated learning works

Think of a consortium of chefs, each guarding a secret recipe they won't share. In federated learning, every chef cooks in their own kitchen, then writes down what they learned as a set of numbers. Those notes go to a head chef, who combines them into one master recipe and sends it back. No chef ever sees another's ingredients.

Technically, the paper follows the standard federated workflow. A central server selects participating clients and broadcasts the current global model; each client trains locally on its own data; the server aggregates the resulting parameters, weights and gradients, into a new global model. Only parameters circulate. The authors then layer privacy mechanisms on top. Differential privacy adds calibrated noise through clipping and Gaussian mechanisms so an individual's influence cannot be reverse-engineered from the updates. Secure multi-party computation lets parties compute jointly without revealing private inputs, and homomorphic encryption allows computation directly on encrypted numbers, which protects the aggregation step itself.

What the study actually built

The experiment simulates a federated setting rather than deploying across real banks. The team chose a Random Forest classifier, an ensemble method that averages many decision trees, to predict loan defaults. Random Forest suits credit work because it captures non-linear patterns while remaining interpretable enough for a lending decision. The full paper, Advancements in Federated Learning for Secure Data Sharing in Financial Services, appears in the FUDMA Journal of Sciences, a peer-reviewed open-access journal run by Federal University Dutsin-Ma in Nigeria.

For data, they used the Lending Club loan dataset hosted on Kaggle, a public collection of borrower attributes and repayment outcomes that stands in for client-held banking data. The authors write that earlier work often overlooked class imbalance, model bias, and the privacy-utility trade-off that differential privacy injects. Their framework tries to confront those problems explicitly and evaluate the model on a full set of classification metrics rather than accuracy alone.

The numbers, and where they cut

The headline figures are encouraging. The model reaches 80% overall accuracy and 98% recall for defaulted loans, meaning it catches nearly all true defaulters. Precision lands at 81%, which produces an F1-score of 89%.

Then come the caveats. The area under the ROC curve, or AUC, sits at 0.69, which the authors describe as moderate discriminatory power with room for improvement in separating default from non-default cases. The reason is class imbalance. Most loans do not default, so the model tilts toward the majority class, and the authors note a potential to misclassify some non-default cases. In plain language, the system is aggressive about catching defaulters, and the price is that a share of reliable borrowers get flagged as risky. For a lender, that means rejected applications and customer friction, though it is the safer failure mode than missing a default.

What a skeptic would ask

The study's own framing invites caution. The federated setting is simulated on one public dataset, not measured across distinct banks with genuinely different customer mixes, what deployed systems call non-IID, or non-identically distributed, data. A partition of a single dataset can understate the difficulty of heterogeneous institutions.

There is also the residual privacy question the paper concedes. Federated learning provides an initial layer of protection, but model parameters can still inadvertently disclose information about training data. That is precisely why the authors add differential privacy and secure multi-party computation; those tools reduce leakage risk rather than eliminate it. Noise injection from differential privacy typically costs some model accuracy, a trade-off the paper flags without fully resolving.

"An AUC of 0.69 is better than a coin flip, but it is not the kind of discrimination a regulated lender can sign off on without stress-testing," said an independent machine-learning auditor who reviewed the publication and was not involved in the work. "The architecture is plausible. The real test is whether the model holds up across truly separate banks with different lending books." The published version does not itemize external funding or conflicts of interest.

Where this goes next

Financial institutions will keep watching this line of research because the incentives line up. Fraud detection, anti-money-laundering compliance, and biometric authentication, where biometric data are uniquely sensitive and non-resettable, all improve with cross-institution data. Federated learning is among the few architectures that can lawfully deliver that collaboration under the General Data Protection Regulation and the similar frameworks the authors cite.

Unuigbokhai's team, including Godfrey Perfectson Oise and colleagues from Wellspring University, has supplied a scaffold more valuable than any single accuracy figure. Over the next five to ten years, the natural progression is consortium fraud models across banks, cross-border anti-money-laundering detection, and biometric authentication where the template never leaves a customer's phone. The paper's own AUC of 0.69 is a reminder that the road from an experiment to a supervised production system runs through false positives, auditability, and trust.

Photo by Jonathan Cooper on Unsplash