At this year’s Credit Scoring and Credit Control Conference in Edinburgh, my colleague Peter Szocs and I presented on an increasingly hot topic: How federated learning could help banks tackle fraud.
Everyone knows that fraud has escalated sharply in recent years. It has grown both in scale and in sophistication. Cifas reported more than 217,000 cases to the National Fraud Database in the first half of 2025 alone, the highest number ever recorded. Within that, over 118,000 cases were identity fraud — a category that increasingly includes synthetic IDs — while misuse of facility filings rose by 35%, signalling a growing use of mule accounts.
|
➡More on this here: What the latest Cifas Fraudscape update tells us. |
Together, these trends are putting pressure on banks to detect fraud earlier, while regulators continue to expect minimal disruption for genuine customers.
Collaboration is needed, but privacy rules, security concerns and operational barriers slow data sharing. And that’s where federated learning has potential. Here’s why…
Why traditional data sharing is limited in fraud prevention
We’ve summarised this in a few key points:
#1. Fraud spans across banks
Fraud rarely stops at the boundaries of a single institution. Synthetic identities are often used to open accounts in more than one bank, while mule networks rely on quickly moving money through multiple accounts across different firms. Despite this reality, most fraud detection systems are still set up to work within their own walls.
#2. Privacy slows down collaboration
Of the barriers to data sharing, privacy laws create the biggest challenge. They are important, but they slow information exchange to the point where fraudsters can take advantage of the delay. Security concerns and operational constraints add further friction, which makes collaboration between banks even harder in practice.
#3. The impact of fragmented data
The result is that each bank sees only part of the activity. A transaction that looks legitimate in isolation can look very different once it is combined with data from other institutions. When sharing is restricted, those connections are missed, and the chance to intervene earlier is lost. Fraudsters are well aware of this and use it to their advantage.
These barriers don’t make collaboration impossible. They highlight the need for a different model. One that allows banks to combine insight while protecting sensitive customer data. This is the space where federated learning has the potential to add real value.
A hybrid approach: Data sharing and federated learning for fraud prevention
One way forward is a hybrid approach that combines the strengths of traditional data sharing with the benefits of federated learning.
Traditional data sharing still has a role. Aggregated information can be exchanged between institutions to support compliance, transparency and industry-wide monitoring. It helps create the baseline view of activity across the market.
But federated learning then takes this further. Sensitive, customer-level data does not need to leave the institution, but it can still contribute to a model that is trained collaboratively across banks. Each organisation trains the model locally, fine-tunes it on their own data, and then feeds confirmed fraud cases back into a shared database.
This creates what we described in Edinburgh as a privacy-preserving method of collaboration. It runs in parallel with existing data sharing frameworks, and together they form a hybrid strategy that maximises effectiveness while staying compliant.
The result? A stronger fraud and risk insights with less regulatory friction.
Why federated learning is a must-have in fraud prevention
Federated learning brings several clear advantages for banks:
- Privacy by design: sensitive customer-level data stays within each bank.
- Scalability: models can be trained across many institutions at once.
- Adaptability: the approach can adjust as fraud tactics change.
These benefits are valuable, but the bigger issue is exposure. No single bank sees enough of the problem on its own:
- One institution might be dealing with synthetic identities.
- Another could be facing more mule accounts.
- A third might be seeing an increase in cross-border scams.
Brought together, they can. That is where federated learning excels: it allows banks to combine intelligence without moving or exposing raw data.
Importantly, federated learning does not require banks to rip out existing fraud systems. It can be layered on top of current defences, running alongside traditional data-sharing frameworks. This makes it a privacy-preserving method of collaboration that strengthens what banks already have, rather than asking them to start again.
Lessons from testing federated learning in practice
Proof of concept work is essential when looking at new techniques such as federated learning. The aim is not only to test whether the method functions, but also to understand the challenges of applying it in practice.
Our own exercise achieved both. It showed that federated learning can be implemented effectively and, just as importantly, it gave insight into the issues that need to be addressed along the way.
These challenges span three areas:
- Technical: ensuring data is consistent in format and quality across institutions.
- Operational: integrating federated models alongside existing fraud systems.
- Cultural: overcoming resistance to information sharing, and the perception that federated learning will only work if benefits are reciprocal.
None of these areas can be ignored, but all of them can be overcome. This mirrors wider industry studies, such as those carried out by SWIFT and IBM, which found that federated models can perform at levels close to centralised ones. Taken together, these projects show that the method is workable and deserves further exploration across the industry.
Federated learning: The future of financial crime prevention?
We think so, yes. Proof-of-concept work has shown that federated learning can be implemented effectively and, just as importantly, that it highlights the challenges that need to be addressed along the way. Honestly, none of these areas can be ignored. Data has to be consistent in format and quality, models need to sit alongside existing fraud systems, and collaboration has to be seen as reciprocal. The encouraging part is that all of these barriers can be overcome.
That is why I believe federated learning can move from proof of concept into mainstream use over the next three to five years. I strongly feel it can be a difference maker, particularly as fraudsters continue to adapt their tactics. Retail banking, payments and cross-border monitoring all stand out as potential areas of adoption, but in the near term my view is that telecommunications may be the strongest opportunity given the growth of fraud in that sector.
Fraud is evolving too quickly for any single institution to tackle on its own. Federated learning gives banks a way to collaborate without breaking privacy rules, and that makes it one of the most promising approaches for the future of fraud prevention.
Keen to hear more from the Edinburgh Credit Conference? Take a look at this blog: Re-evaluating loss forecasting: Why affordability could be the missing link