Fraud Detection in ATM and Payment Environments Using Machine Learning

Fraud patterns never stay the same. Criminal syndicates invest significant resources in finding new ways to siphon off money from vulnerable organizations and individuals. These organized crime units are consistently scanning and testing the limits of organizational controls, lurking in the shadows and waiting for the right moment to launch a new attack. For this reason, fraud prevention is a never ending activity, forcing us to always be steps ahead.

EMW Webinar Image

Over the years, we have seen fraud patterns increase significantly in numbers and complexity. Organizations have responded by implementing tool upon tool, many of which make use of rule engines to catch the type of fraud that is surfacing at a point in time. Unfortunately, the issue with this is that we are only able to pick up fraudulent transactions that match the rule – new fraud patterns go by completely undetected. The other major issue is that setting these rules either means having a large number of rules for different customer types and fraud patterns, which is in itself an administrative burden, or having a limited number of broader rules, which reduces the administrative burden, but results in high false acceptance rates (FAR) or false rejection rates (FRR). Finding the right balance is a challenge….

Enter Machine Learning. Simplistically, Machine Learning allows us to build behavioural models for each and every individual customer that transacts or every device that acquires transactions within a payment environment. As transactions are carried out, they are compared in real-time, against models that have been created by statistically analyzing past transactions and then passing this onto a scoring engine. When transactions are out of pattern for a specific customer or device, or matches known fraud patterns, an alert is generated.

For Machine Learning to be effective in detecting fraud, both Supervised and Unsupervised Machine Learning capabilities must be harnessed. Unsupervised Machine Learning is used to build individual models and detect anomalous transactions, while Supervised Machine Learning is used to detect known fraud patterns, once transactions have been labelled as fraud. Together, these two capabilities offer the widest ability in detecting new fraud types, while at the same time, improving detection accuracy, thereby eliminating the guesswork and limiting the number of cases that Fraud Analysts have to deal with on a day to day basis.

Machine Learning is a complex area of study, spanning many different constructs, algorithms and technologies. INETCO seeks to make it easier for our customers to harness this exciting area in an already packaged form. If you have any questions, please do not hesitate to reach out to us at

Leave a Reply