Fraud does not respect the boundaries of any one institution. A scheme that fails at one bank moves to the next, tries a different card, opens an account somewhere else, and reuses the same underlying tactics across a dozen targets that never compare notes. Each institution defends its own perimeter with its own data, and each sees only the fraction of the pattern that touches it directly. The result is a defense that is fragmented by design against an adversary that is coordinated by nature.
Adarsh Naidu, a Senior Solutions Architect in the Enterprise Banking group at Amazon Web Services, has spent much of his career inside that asymmetry. Across more than two decades at financial services firms, including American Express and Assurant, and now at AWS, he has worked on fraud and dispute systems in banking. The structural weakness he keeps returning to is not any single model’s accuracy. It is that the institutions fighting fraud cannot easily learn from one another.
Why the Data Stays Locked Inside Each Bank
The obvious fix, pooling everyone’s transaction data so models can learn the full picture, runs straight into a wall. Raw financial data is among the most tightly held information a bank has. Privacy law, banking secrecy rules, and internal controls make sharing customer records across institutions difficult, and for good reason. The same caution that protects customers also keeps each bank’s view of fraud narrow.
That leaves the industry in an awkward position. The signal needed to detect cross-institutional fraud exists, but it is scattered among competitors who cannot legally share it with each other. A fraud pattern that would be obvious if you could see it play out across ten banks stays invisible when each bank sees only its own slice.
“Fraud moves across institutions, but our defenses usually stop at the bank’s own walls. The question is how banks learn from each other without ever handing over their customers’ data,” says Naidu.
The question is not rhetorical. It defines the boundary of what is currently possible in collaborative fraud detection, and it is the boundary that Naidu’s more recent research is aimed at.
Learning Together Without Sharing the Data
The direction he points to is federated learning, an architecture that inverts the usual assumption. Instead of moving data to a central model, the model travels to the data. Each institution trains on its own records locally, and only the resulting model updates, not the underlying transactions, are combined into a shared system. Layered on top is a technique called differential privacy, which adds mathematical noise to those updates so that no individual customer’s information can be reconstructed from what is exchanged.
Applied to fraud, this changes what is possible. Banks could contribute to a shared understanding of emerging fraud patterns without their customers’ raw data ever leaving their systems. The collective model would learn from the whole network’s experience, but no institution would ever expose its records to a competitor or a central authority.
This is the premise of research Naidu has in preparation: a federated approach to cross-institutional fraud, which he has described under the working name FedSynFraud. It sits alongside other manuscripts in development, including work on LLM-powered dispute adjudication and privacy-preserving analytics for insurance claims. The through-line across all of it is the same constraint: extract shared intelligence from data that cannot be shared.
A Governance Problem as Much as a Technical One
Federated learning is not a finished solution waiting to be switched on. Coordinating model training across rival institutions raises questions of governance and trust that are at least as hard as the mathematics. Who operates the shared model, how contributions are validated, and how institutions agree to participate are open problems, and Naidu treats them as such rather than as details to be settled later.
That posture is consistent with the rest of his work. He has authored a Responsible AI framework for AWS that embeds governance into each phase of system development rather than treating it as a final review, and he has presented this approach at US industry events between 2023 and 2025. Any cross-institutional system would also have to meet the growing expectation that automated decisions affecting customers can be explained and audited, a bar that only rises when several banks share one model.
The fraud problem, no single bank can solve alone, is, in the end, a test of whether competitors can build shared defenses without surrendering what they are obliged to protect. Naidu’s bet is that the architecture to do it already exists, and that the harder work is getting institutions to trust it. On current evidence, that is the part still to be proven.
