- Amex has been historically creating machine learning models that examine fraud risks during point-of-sale transactions, such as when a client uses a credit card in a store.
- The new ML model can evaluate most current data in real-time and then factor it while calculating its most recent performance as it adjusts.
Using Artificial Intelligence (AI) and Machine Learning (ML) to combat credit card fraud is nothing new for American Express (Amex). The firm has been utilizing AI to automate billions of decisions about fraud risk for years, even as hundreds of Amex data scientists work on AI and ML models linked to fraud risk.
James Lee, vice president of global fraud risk at Amex, says, “It is certainly a key focus for us. We’re totally vigilant to make sure that we defend against those risks.”
But account login fraud is a complicated problem that is only becoming more significant. With the introduction of chip-and-pin cards and online one-time passwords, fraudsters seek novel ways to steal credit card information.
Amex ML model detects fraudulent account logins
One way they accomplish this is by accessing a customer’s online account to change crucial demographic information, ordering replacement cards, gaining access to OTPs, or disabling spend/fraud alerts, and then using the customer’s card to make fraudulent transactions. They might even use membership rewards currencies to buy digital gift cards.
To eliminate login fraud, Amex recently designed an end-to-end ML modeling solution that, at the account login level, can determine whether the login is from a legitimate client to detect login fraud. While low-risk logins enjoy a seamless online experience, high-risk logins demand incremental authentication. As a result, there will be little to no impact on the good consumers as bad logins will be quickly detected.
According to Lee, the next-step-up authentication is high in friction for legitimate customers. “There was a strong push from our leadership team to make sure that we evaluate the risk of the individual logging in, leveraging the vast amount of data and history we have on that customer’s activities,” he added.
Fraud rates have been dropping over time with the iteration of the ML model for real-time prediction of account login risk. According to him, “With the first iteration versus now, the model is stronger-performing than most other models in the marketplace provided by third-party vendors.”
Stopping login scammers in real time
A 60-person machine-learning team dedicated to tackling all types of fraud is led by Abhinav Jain, VP of global fraud decision science at Amex. According to him, a significant project objective over the past few years has been to develop an ML model to manage the danger of login fraud.
According to him, Amex has been historically creating machine learning models that examine fraud risks during point-of-sale transactions, such as when a client uses a credit card in a store.
However, as online robberies and account hacking increased, Amex saw the need to eliminate login fraud “so that we can block the bad actors upfront and not wait for them to transact,” he said.
The first problem Jain’s team could overcome was integrating logins into an ML platform whose model had been trained using past customer data. “Each login needs to get scored by the model in real-time,” he said.
The second challenge came from learning how to spot fake logins. He said, “When we build a transaction or point-of-sale model, we reach out to customers, or customers reach out to us, so we know which transactions are fraudulent or not. But “it becomes tricky because we don’t go back and ask customers” with account login fraud.
Instead, Amex had to create a learning strategy for the ML model. Based on the customer’s prior online login activity, it determines which logins are fraudulent, valid, and unsure.
ML model from Amex provides a feedback loop
“It’s really about that feedback loop,” said Lee, who added that the machine learning model incorporates new data and assesses whether specific signals and features represent false positives or are legitimate predictions of future fraudulent activity.
He said, “There was always a rules-based structure to determine the low versus moderate versus high risk.” However, that was more of a static output than the new ML model, which can evaluate most current data in real-time and then factor it while calculating its most recent performance as it adjusts.
He continued, “That has allowed us to strengthen the hit rate for high-risk prediction. It is what enables us to have the industry’s leading fraud reduction rates relative to any networks or competitor issuers in the marketplace.”