From building analytical systems to detect fraudulent payments to exploring privacy-preserving AI, the finance professional is focused on a persistent problem: fraud can cross systems and institutions faster than the information used to detect it.

As banks and insurers turn increasingly to artificial intelligence to combat financial fraud, a finance professional and researcher is focusing on a harder question: how can financial institutions learn from one another’s fraud patterns without exposing the sensitive customer information they are required to protect?

Ugochukwu Daniel Ofurum is among the researchers examining that problem through federated learning, a form of machine learning designed to allow institutions to collaborate on model development while keeping their underlying datasets within their own environments.

His interest in fraud analytics began with a problem that was anything but theoretical. Earlier in his career, Ugochukwu worked in finance functions involving variance analysis, budgeting, forecasting, and enterprise reporting. During this period, he helped his organisation discover vulnerabilities and control gaps within its payment and reconciliation processes that were being exploited to create fraudulent transactions resembling legitimate company outflows. He assessed weaknesses in the existing payment and reconciliation processes, redesigned key controls, and developed analytical models to identify unusual patterns and duplicate transactions. According to finance leaders in the organisation, the enhanced control framework uncovered approximately ₦500 million in fraudulent transactions and financial leakage. His models and methodology prompted the finance team and relevant counterparties to investigate the activity and initiate recovery efforts, while the redesigned controls strengthened the organisation’s ability to detect and prevent similar losses.

The experience gave him a view of fraud that has continued to shape his work: warning signs can exist without being useful if the information needed to connect them is fragmented. “That experience changed the way I thought about fraud,” Ugochukwu said. “The individual transactions did not necessarily tell the whole story. Once we started connecting information and looking at the pattern, the problem became much clearer. It showed me that sometimes the challenge is not a lack of data. It is that the right pieces are not being seen together.”

More than seven years of experience across banking, asset management and insurance have since broadened that perspective. Ugochukwu’s work now sits at the intersection of finance, artificial intelligence and financial crime, with a particular focus on how institutions can use data to identify sophisticated fraud without creating new risks around privacy, explainability and accountability.

The technology has changed since that earlier incident, but Ugochukwu believes the underlying problem has not. Modern financial crime can move across banks, insurers, payment platforms and other organisations, while the information used to detect it often remains confined within institutional boundaries. A bank may see one transaction, another institution may see the next, and a payment platform may see another part of the same movement. Each organisation can improve its own fraud controls and still miss a pattern that becomes obvious only when those pieces are connected.

“You can build a very sophisticated fraud model and still have a visibility problem,” Ugochukwu explained. “If part of the pattern sits somewhere else, the model cannot learn from information it never sees. The question I am interested in is whether institutions can overcome that limitation without simply pooling sensitive customer data.”

That question has become central to Ugochukwu’s work on privacy-preserving artificial intelligence. Rather than treating better fraud detection as simply a matter of collecting more data or building a more complex algorithm, he is interested in whether financial institutions can expand what their systems learn while retaining control of the underlying information entrusted to them.

One approach Ugochukwu has explored is federated learning, a form of machine learning that allows organisations to contribute to the development of a shared model while keeping their underlying datasets within their own controlled environments. Instead of moving customer records into one central repository, participating institutions can train a model locally and exchange protected learning updates that improve the shared model.

For financial services, the appeal is straightforward. Fraud can cross institutional boundaries, but customer and transaction data cannot simply move between institutions with the same freedom. Privacy obligations, security concerns and regulatory requirements create legitimate limits on how that information can be shared. Federated learning offers a possible way to separate the sharing of learning from the sharing of raw data.

Ugochukwu’s interest in the approach has also extended into published research. He has co-authored work examining federated learning for privacy-preserving fraud detection in digital banking, as well as research addressing AI governance, automated compliance and explainability in financial services. He does not, however, view the technology as a ready-made answer.

“Privacy is only one part of the problem,” Ugochukwu said. “Two institutions can be looking at the same type of fraud and still collect, label or structure the relevant information differently. Keeping the data local does not make those differences disappear. If collaborative AI is going to work in practice, those differences have to be dealt with.”

Ugochukwu’s focus on fraud detection is closely tied to a second concern: whether the people responsible for acting on an AI-generated alert can understand why the system produced it. A model might identify an unusual relationship between accounts, repeated documentation patterns, suspicious payment chains or connections between apparently unrelated cases. Those are leads an investigator can examine. A risk score with no understandable explanation is far less useful.

“AI performance cannot be the only objective,” Ugochukwu said. “If a system flags an account or transaction, an investigator needs to understand what caused the alert. The technology should help people decide where to look. It should not become a black box that makes consequential decisions simply because the model produced a high score.”

That distinction is particularly important in regulated financial services, where fraud controls can affect customers, businesses and investigations. For Ugochukwu, the strongest use of AI is not to remove people from the process, but to help them identify relationships and anomalies that would otherwise be difficult to find.

Ugochukwu argues that the next step should not be a sweeping industry-wide deployment. He favours small, measurable pilots involving a limited number of institutions and one clearly defined fraud problem. Participating organisations could first establish how well their existing institution-specific models perform, then compare those results with a collaborative approach.

The test would be deliberately practical. Did the collaborative model identify relationships that the institutions missed individually? Did it improve the quality of fraud referrals? Did it create too many false positives? Could investigators understand the reason behind an alert? If the collaborative system performs no better than the existing approach, Ugochukwu says that result should be accepted rather than forcing adoption.

“The objective should be evidence, not technology for its own sake,” he said. “If the model does not give investigators useful information they did not have before, then it is not solving the problem. But if it consistently reveals patterns that no institution could see alone, then we have learned something important.”

Ugochukwu believes the same visibility problem can appear elsewhere in financial services. Insurance fraud, for example, can involve claimants, providers, repair facilities or other actors whose activity touches several insurers. Each carrier may see an individual claim without having enough information to recognise a broader relationship. Payments and digital financial platforms can face a similar challenge when suspicious activity moves rapidly from one institution to another.

That is why Ugochukwu sees his earlier experience as connected to the problem he is pursuing today. In the earlier case, the fraud became easier to detect when separate pieces of financial information were connected. The challenge he is now exploring is what happens when those pieces sit not merely in different records or systems, but in different organisations that cannot simply combine everything they know.

“The scale is different, but the underlying question is familiar,” Ugochukwu said. “Fraud often succeeds not because the warning signs do not exist, but because the information needed to connect them is fragmented. What interests me is how we can use AI to connect the learning without requiring institutions to surrender the underlying data.”

For Ugochukwu, that question sits at the centre of a broader effort to make artificial intelligence in financial services more useful, explainable and privacy-conscious. As financial crime becomes more connected, he believes the systems designed to detect it will eventually need to learn beyond the boundaries of any single institution.