New Amex ML Model Helps Reduce Account Login Fraud

For American Express (Amex), using AI and machine learning (ML) to tackle credit card fraud is nothing new. The company uses artificial intelligence to automate billions of fraud risk decisions over years as hundreds of Amex data scientists work on AI and ML models related to fraud risk.

“It’s definitely an important focus for us,” James Lee, VP of global fraud risk at Amex, told VentureBeat. “We are fully vigilant to ensure that we defend against those risks.”

However, account login fraud is a particularly thorny challenge that is only growing in importance. With the advent of chip-pin cards and online one-time passwords, fraudsters are looking for more unconventional ways to commit credit card fraud.

Amex ML model locates account login fraud

One way they do that is by logging into a customer’s online account to change key demographics, order replacement cards, access OTPs or disable spend/fraud alerts – then fraudulent transactions to be performed on the customer’s card. They can even access membership reward currencies and try to exchange them for digital gift cards.

To detect login fraud, Amex recently developed an end-to-end ML modeling solution that can predict, at the account login level, whether the login belongs to a real customer. High-risk logins are required to add incremental authentication, while low-risk logins get a seamless online experience. This ensures that bad logins are logged in real time, while good customers are minimally impacted.

Next-step-up authentication has a lot of friction for real customers, Lee explained. “There was strong pressure from our leadership team to make sure we evaluate the risk of the person’s login, leveraging the massive amount of data and history we have about that client’s business,” he said.

Now, with the iteration of the ML model for real-time prediction of account login risks, fraud rates have decreased over time. “With the first iteration versus now, the model outperforms most other models on the market offered by third-party vendors,” he said.

Stop login fraudsters in real time

Abhinav Jain, VP of global fraud decision science at Amex, leads a 60-person fraud machine learning team working for Amex worldwide on projects related to fraud of all kinds. He says building an ML model to address the risk of login fraud has been a key project goal in recent years.

Traditionally, he explained, Amex has developed machine learning models that analyze fraud risks in the transaction at the point of sale, such as when a customer uses a credit card in a store.

But as the activity of login fraud increased with online takeovers and account hacking, Amex saw the need to prevent login-level fraud, “so we can stop the bad guys beforehand and not wait for them to transact,” he explained.

The first challenge Jain’s team was able to solve was integrating logins into an ML platform that had trained the model on historical customer data. “Every login has to be scored by the model in real time,” he said.

A second challenge was figuring out how to identify fraudulent logins. “When we build a transaction or point of sale model, we contact customers, or customers contact us, so we know which transactions are fraudulent or not,” he said. But with account login fraud “it gets tricky because we don’t go back and ask customers.”

Instead, Amex had to develop a logic to learn the ML model. It uses the customer’s previous online login behavior to identify which logins are fraudulent, which are good and which are insecure.

Amex ML model provides a feedback loop

“It’s really about that feedback loop,” said Lee, explaining that the machine learning model contains new information and determines whether certain signals and characteristics translate into false positives or are actually accurate predictions of future fraud behavior.

“There was always a rules-based structure to determine the low versus moderate versus high risk,” he said. But that was more of a static output, as the new ML model can review all the latest information in real time and then factor that into the most recent performance as the model calibrates itself.

“That has enabled us to strengthen the success rate for high-risk predictions,” he added. “It’s what allows us to have the industry’s leading fraud reduction rates compared to any network or competitive publisher in the market.”