Banking, card and payment-related fraud has been growing steadily over the last five years, climbing to $533 billion globally in 2024 - around $148 billion more than in 2021 - according to estimates by research company Celent.
In its latest report, 'Mitigating Fraud in the AI Age', Celent explores the underlying reasons for this worrying trend, including how the 'bad actors' are using AI and automation technology to devise more sophisticated ways to defraud their victims in greater numbers.
The report, which was commissioned by IBM, goes on to examine the advanced detection and prevention techniques that financial institutions are using to combat fraud and how the industry could potentially cut $190 billion off annual fraud losses by running all transactions through AI models using the new IBM z17 with its on-chip and off-chip AI acceleration capabilities.
Why is fraud on the increase?
Celent's analysis suggests banking, card and payment fraud has grown in sophistication over the last five years, with fraudsters using automation and machine learning to commit fraud in multiple ways. "Every step in the financial services value chain, from account opening onwards, and every product - including retail and corporate payments, loans, and cards - is targeted by fraud," the report explains.
Examples of fraudulent activities that are on the rise include invoice fraud (scammers sending fake invoices), business email compromise, confidence tricks and other social engineering scams in which fraudsters pose as relatives or friends or impersonate businesses or government departments to trick victims into transferring money.
These and other trends have contributed to an annual growth of 11.1% in banking fraud between 2021 and 2024 - and 14.2% in cards and payments fraud.
How the industry is fighting back
Of course, financial institutions have been doing their part to stop the scammers. They've been evolving machine learning models and developing ways to use transformer technologies such as generative AI and quantitative foundation models to ramp up their efforts in fraud detection and prevention.
For example, using deep learning, a subset of machine learning, it's possible to analyze transactional data at scale to detect fraudulent activity, including exposing new, previously unseen types of fraud. Recurrent neural networks (RNNs) a type of deep learning model, offer another new approach that has been shown to improve fraud detection, as have advanced techniques that use multiple machine learning models simultaneously for greater accuracy.
By expanding the range of data they analyze, organizations have been able to improve the ability of AI models to detect fraud. As well as the core transaction data, they're using customer data, biometric data, and consortium data (pooled across the industry from other financial institutions) to spot signals that indicate fraudulent behaviour.
Quantitative foundation models - also called large transaction models (LTMs) - are being used to analyze quantitative transactional data to predict fraud, with the insights then being used by machine learning models to enhance fraud detection.
Despite all this effort, however, Celent suggests that fraud detection will only be able to fully benefit from the advances in machine learning and transformer models once financial institutions achieve the "holy grail" of being able to run these complex models in real-time on ALL transactions.
Where IBM Z comes in
Globally, Celent estimates that 70% of transactions by value run on IBM. But until recently, most large financial institutions running mission-critical transaction processing on Z were having to take transaction data off the mainframe and send it to anti-fraud solutions that resided in the cloud - or on another platform - for real-time analysis.
Sending the data off IBM Z like this introduces latencies and impacts SLAs for processing transactions, which has meant most large organizations have only been able to send a small sample of their transaction data for real-time fraud analysis. This increases the risk of some fraud slipping through unnoticed.
IBM went some way to addressing this problem with the introduction of the Telum processor for the IBM z16 mainframe in 2022. Telum included an innovative AI accelerator to enable AI inferencing directly on the mainframe. With this, banks could run complex models to check for fraud on all their transactions in real-time, avoiding the latency and throughput issues created by having to send data off the platform.
But, of course, technology never stands still. As the fraudsters have got smarter, anti-fraud systems are becoming even more complex, running larger models, running models simultaneously and using wider internal and external data sets. Plus Celent believes that LLMs and other transformer models will eventually be used for real-time inferencing, becoming integral tools for transaction fraud detection. So, there is the prospect of even heavier compute demand on the systems running the models.
This is why Celent suggests in its report that the new IBM z17 will likely be a game changer for anti-fraud applications.
Enter the IBM z17 with next-gen AI acceleration
The new mainframe incorporates the next-generation Telum II processor with an on-chip AI accelerator, delivering eight times the AI processing units of the IBM z16. It is estimated that if it's used to run a deep learning model applied to credit card fraud detection, the IBM z17 can process up to 5 million inference operations per second (up to 450 billion inference operations per day, with 1 millisecond or less response time).
Plus, IBM has developed an off-chip accelerator for the z17: the IBM Spyre Accelerator with 32 accelerator cores. AI processing power can be shared across the chipset with both Telum II and Spyre tuned to handle LLMs for both generative and predictive use cases.
The Telum II and Spyre together give the IBM z17 enough compute power to run larger and more complex models, to run multiple models simultaneously, and to run LLMs directly on the mainframe to improve the accuracy of real-time fraud detection. By allowing organizations to run these models on the mainframe, it helps safeguard the data and the intellectual property of the models.
The numbers
According to Celent's estimates, if all banking, cards, and payments institutions currently running on IBM Z were to put all transactions through advanced AI models running on the platform - benefiting from the additional compute power provided by the IBM z17 - it could mean an additional $190 billion in fraud captured globally. This would comprise a $165 billion reduction in bank fraud losses and a $25 billion cut in cards and payments fraud. In the US alone, bank fraud could be reduced by $51 billion and cards and payments fraud by $7 billion.
Keeping pace with the bad actors
Celent's report highlights how fraud has grown in both scale and complexity in recent years, with automation and advanced AI technologies empowering the bad actors to target every stage of the transaction process at scale.
The financial institutions have no choice but to keep pace with the fraudsters and fight fire with fire by adopting increasingly sophisticated AI-powered anti-fraud systems. The IBM z17, with its next-generation AI accelerators, is a significant step forward in this war; it has the potential to transform fraud detection, paving the way for significantly cutting global fraud losses.
This blog was originally published on the IBM Community.