Banks have spent decades building fraud systems that see one transaction at a time. The charge may or may not seem suspicious. Fraud gangs have built their business model around this gap, spreading activity across thousands of payments using stolen cards, illicit accounts, shared devices and synthetic identities, so no single transaction can escape filtering.
Nelson Report Projects Global card fraud losses will reach $403 billion over the next decade, with the United States accounting for nearly 42% of those losses despite representing only 26% of total card volume worldwide, according to a press release.
Nvidia Artificial intelligence scheme for detecting financial fraud It was built around a different idea. Instead of asking whether a single transaction looks suspicious, the system asks whether the people, devices and accounts involved in the transaction are linked to suspicious activity elsewhere. A $47 purchase at a gas station may seem completely normal in itself. It would look different if the phone used to approve it also showed up in 60 other disputed charges across three states that week. Or the same card was opened using an address linked to a known account.
This is the blind spot on which fraud rings rely. PYMNTS INTELLIGENCE Found Fraud by an unauthorized party — driven by credential theft and account takeovers — now accounts for 71% of all fraud incidents and dollar losses at U.S. financial institutions, up from 48% in 2024. Organized rings move quickly precisely because they know the window before detection closes.
Why does transaction-level logging fail versus structured loops?
Most bank fraud systems today use a technique called gradient-enhanced modeling, which is a scoring engine that looks at the characteristics of a transaction and decides whether it resembles previous fraud. Was the purchase made in an unusual place? Was the amount outside the allowed range for this customer? Was the card used twice within five minutes in different cities? These are useful signals for catching individual bad actors.
It is much less useful against the coordinated loop. A ring using 500 stolen card numbers can keep each card’s activity within normal-appearing ranges, making individual transactions appear routine. The Nielsen report found that cardless transactions represented the riskiest category in every region of the world, precisely because they were easier to carry out at scale using stolen credentials, according to the release.
Nvidia Mapper addresses this gap by adding a layer that defines relationships across data. This technology, graph neural networks, works by building a picture of how transactions, accounts, and devices connect to each other, then looking for groups that share suspicious links. It feeds these relationship signals into the existing scoring model as additional context, so a low-scoring transaction can still be flagged on its own if it exists within a connected set of high-risk activities.
piments I mentioned The bloc’s chief risk officer, Brian Potts, has pushed banks away from reviewing fraud after the fact towards stopping it for now. “It’s one thing to find bad actors after the fact,” Potts said. “But what would be more effective would be to invest in more real-time technology.” PYMNTS INTELLIGENCE Found 68% of financial institutions have increased spending on fraud detection year over year as the problem extends beyond legacy systems.
Real-time decisions within live payment flows
The challenge in relationship-based analysis is speed. Mapping connections across millions of accounts and transactions requires significant computing power. Doing this quickly enough to stop a payment before it completes, typically within a few hundred milliseconds, requires infrastructure that most banks have not yet built.
Nelson Report male Worldwide card fraud losses reached $33.41 billion in 2024, and AI tools have helped the industry build its best anti-fraud models to date, even as organized crime continues to adapt.
Nvidia’s scheme uses its Dynamo-Triton inference server to run relationship checks at batch speed. The system produces a fraud score for each transaction along with the flags that led to it, so a fraud investigator can see not only that a transaction was flagged, but that it was flagged because the device matched three others in an active dispute group, or because a billing address was used to open four accounts in the past week. The scheme runs on Amazon Web Services and Hewlett Packard Enterprise, with support for Dell Technologies planned, Nvidia said.
For all of our PYMNTS AI coverage, subscribe to our daily newsletter Amnesty International newsletter.
