Reviewed by Mayer Hyman, Payments Specialist | Reviewed for accuracy July 2026
Key Takeaways
- Fraud protection evolved from physical bank security to rule-based systems in the 1980s to machine learning today, each generation trading speed for accuracy gains.
- Rule-based systems required hundreds of static rules and struggled with instant-payment speed requirements; neural networks trained on historical transaction data catch far more fraud in a fraction of the review time.
- Global card fraud losses were roughly $33.4 billion in 2024, underscoring why faster, smarter detection matters more every year.
- Quantum computing is the next frontier for both fraud prevention and fraud itself — still years from mainstream deployment, but worth watching.
The war waged between financial institutions and thieves began long ago. It started with metal bars on bank teller windows, three-foot thick doors, and steel-reinforced walls to protect vaults. Then, as money and transactions turned digital, cyber warfare called for the weaponizing of mathematics by creating complex algorithms. What’s next in the fight against fraud?
Tamsin Crossland, an architect for the fintech Icon Solutions, gave insights into the history of fraud management in an interview with InfoQ. She explained how fraud protection has transformed over the years and how machine learning-based solutions seek out patterns and prevent fraud before it occurs. Her framework is a useful lens for understanding where fraud management has been — and where it’s headed next.
From Branches to Binary
Throughout most of the twentieth century, banking services meant visiting a brick-and-mortar branch. If you wanted to open an account or cash a check, you went to the building on the high street — and so did anyone trying to steal from the bank, typically armed with a plan and little else.
In the 1960s, financial institutions became more sophisticated as computers took over much of the manual accounting work. These systems were still secure. One of the earliest computers was the IBM 360 series, which used a closed network system: data stayed contained in the mainframe and could only be accessed by bank employees, so fraud stayed minimal.
The Need for Speed
Connecting banking networks to the internet exposed customer data to attack for the first time. Closed networks opened up, and cloud computing and mobile payments pushed that data further into reach of hackers looking for any opportunity.
Digital payments need protection, but scanning networks for malware or identifying suspicious patterns takes time — and time is the crux of the problem. Customers want speed with minimal friction, and merchants want a seamless payment process that keeps customers coming back.
Instant payments are expected to happen instantly. If John wants to send Peter an instant payment, John’s bank has to confirm sufficient funds, check sanctions lists, verify the transaction doesn’t look suspicious, and confirm the details are correct — all within about five seconds. That’s not much time for the painstaking work fraud checks require.
Fraudsters also collaborate on the dark web, where a technique that works against one bank gets shared and applied against others. The problem is time, and the need for speed is precisely why machine learning has become so central to modern fraud management.
Rule-Based Systems
In the 1980s, before machine-learning algorithms, rule-based systems were the first real attempt at automated fraud detection. The technology took the knowledge of fraud experts and turned it into rules. For example, if a credit card transaction ran ten times larger than a customer’s average, the rules would flag it as an anomaly and trigger a review, usually by a human.
Rule-based systems designed by experts eventually became unwieldy because they were difficult to build and maintain. It’s hard to encode an expert’s career of learned judgment into a fixed set of rules, especially since fraud tends to occur precisely in the anomalies nobody anticipated. Add dark-web collusion into the mix, and the number of edge cases becomes nearly impossible to keep up with. According to Crossland, legacy rules-based systems often have to apply around 300 different rules before approving a single transaction.
Neural Networks
Neural networks are interconnected nodes with inputs and outputs, loosely modeled on the human brain, that can be trained to recognize patterns — an idea introduced by McCulloch and Pitts back in 1943. The principle: neurons connect with other neurons in a network to encapsulate knowledge.
Each node carries a numeric value that adjusts as the network trains. Feed a system labeled images of cats and dogs, and it adjusts the weights between nodes until it can reliably tell the two apart. Apply the same logic to fraud: banks have years of transaction data, much of it labeled fraudulent or legitimate. Load that data with its labels, and the network adjusts its weights to learn the difference.
Crossland cites a case from NetGuardians as an illustration. NetGuardians collected 10 million payments over a 12-month period and trained a neural network on those transactions. Where a prior rule-based system could only meaningfully review around a third of payments, the neural network could review 100% of them — with a 93% reduction in fraud investigation time. The neural network also caught every fraud case the rule-based system had caught, plus another 18% it had missed entirely.
Crossland doesn’t suggest rule-based systems should be abandoned outright. In practice, the strongest fraud programs combine machine learning, rules-based checks, and human review. Machine learning is excellent at detecting patterns at scale, but some scenarios still need a human eye.
Quantum Computing: The Next Era of Fraud Management
Even as machine learning has raised the bar, fraudsters keep working to beat new algorithms. The technology arms race continues, and the stakes keep climbing — global card fraud losses reached roughly $33.4 billion in 2024, per the Nilson Report, a scale that keeps fraud prevention a permanent priority rather than a one-time fix.
The next technology likely to reshape this fight is quantum computing, which can increase the speed of complex calculations exponentially. In 2019, Google’s 54-qubit quantum processor, known as Sycamore, performed a calculation in 200 seconds that would have taken roughly 10,000 years on the fastest classical supercomputer of the time.
It’s not just speed. Quantum computing also promises to reduce false positives — the incorrectly blocked transactions that frustrate legitimate customers and cost merchants sales. Classical systems mistakenly flag legitimate activity more often than anyone would like; quantum computing’s greater predictive power could meaningfully cut that error rate.
How Long Until Merchants See Quantum Computing in Payments?
Major tech companies — including Alibaba, Amazon, Google, IBM, Intel, Lockheed Martin, and Microsoft — are actively experimenting with quantum computing. Most industry projections still put practical, widespread deployment in payments several years out. In the meantime, machine learning remains the most mature and accessible tool merchants have for fighting fraud today.
As Walt Manning, CEO of the Techno-Crime Institute, has put it: quantum computing could be a game changer for fraud detection and prevention, and artificial intelligence’s role in cybersecurity will only grow more essential.
You don’t need steel vaults or digital walls alone to protect your payments — you need fraud tooling that keeps pace with how fraud itself is evolving. Cartis works with merchants to integrate modern fraud management tools directly into existing payments infrastructure.
FAQ
Is machine learning fraud detection better than rule-based systems?
Generally yes for scale and speed — machine learning models can review a much larger share of transactions and adapt as new fraud patterns emerge. Most effective fraud programs still combine machine learning with rules-based checks and human review rather than relying on any single method alone.
Is quantum computing already used in payment fraud detection?
Not yet in any mainstream sense. Major tech companies are actively researching quantum computing applications, but practical deployment in payments infrastructure remains several years away according to most industry projections.
How much does fraud cost the payments industry each year?
Global card fraud losses were about $33.4 billion in 2024, according to the Nilson Report — a figure that has generally trended upward over the past decade despite improvements in detection technology.
