A Reflection on Payments Fraud

​The dramatic surge in card fraud during the pandemic has been well documented. More recently, market research is showing a reduction in the number of fraud cases as a result of payments organizations implementing modern fraud prevention solutions. However, while the volume of card fraud is declining, the monetary amount of fraud is not, as fraudsters continue their penetrative testing looking for easy targets. Our interview with Fabrice Daniel, Head of the AI team at Lusis briefly introduces some of the key factors that payments organizations should consider in respect of their fraud platform.
​

​Q. Why should an organization consider upgrading their fraud decisioning platform?
F.D. : There have been significant advances in AI research for fraud detection in recent years. Advanced gradient boosting decision trees, deep learning models like Fraud Memory, rules extraction methods like Anchors, or even the Dempster-Shafer theory to combine many probability outputs are only some examples. Payments fraud is an ever evolving challenge and the older fraud prevention solutions are unlikely to cope with the newer fraud attacks. Today's fraudsters are not the opportunists of old, they are highly organized and technically sophisticated. If your fraud decisioning platform and procedures are out of date then you are significantly increasing the likelihood of a serious fraud attack.
​

​Q. What should an organization look for in a new fraud solution?
F.D.: For one thing, you want to make sure that your strategic fraud solution can be deployed in the cloud and leverages the significant processing cost savings of cloud deployments. Additionally, a modern fraud solution needs to provide massively parallel processing to allow customers to deploy a comprehensive mix of rules and machine learning models without significant processing latencies or excessive cost. 

The cost of fraud screening is a legitimate concern, both in terms of computational resources but also in staff costs. Each organization is likely to set their own balance between processing costs and the breadth of screening. However, as a rough guide a large organization should expect to process 50 million transactions using 40 rules with a per transaction latency of under 10 milliseconds.  This level of performance ensures that they can deploy a sufficiently stringent fraud screening strategy in a timely manner. 

The ability to run multiple rules engines simultaneously is another important capability. Trying to rely on a single all-encompassing rules set or model leads to greater inaccuracy. It is far better to use multiple models and rules sets, each tuned to catch a specific set of fraud attributes. This strategy results in higher accuracy and easier management.

FILL OUT FORM TO READ THE ENTIRE DOCUMENT

Complete Form to Download