Why Patterns Are Breaking Fraud Fighting’s Old Rules

Fraud evolves. Fraudsters evolve. Technology evolves, too, and is the weapon financial services firms, merchants and others use to combat the bad guys.

Technology is also the weapon the bad guys use to gain access to personal and transactional data to wreak all sorts of mayhem.

In an interview with PYMNTS, Jack Bloch, senior vice president of software engineering at ACI Worldwide, took note of the explosion in eCommerce, where 24/7 interactions and new ways of making payments have driven transactions – and especially card-not-present transactions – “through the roof.”

Localized shopping habits, and even the benefits of geolocation (in ensuring devices are wielded by the proper individuals during purchases) have fallen by the wayside, he said. “Fraud – specifically, fraud detection – is a very complex subject,” Bloch noted. “There is not ever going to be a silver bullet that fixes all problems and captures all fraudulent transactions” in the age of digital commerce.

But there exist several lines of defense that can aid the battle against fraudulent transactions – and machine learning is crucial, he said.

Patterns Over Rules

The older fraud detection methods have been based on rules, noted Bloch. These systems are “slower, and serial in nature in their execution, traversing several sets of extremely complex rules.” Thus, detecting fraud “can take a long time,” and can impact a bank’s ability to process transactions, introducing friction into the process.

“Rules have a place for certain fraud scenarios, but can be limited,” he told PYMNTS, and a rules-based fraud prevention endeavor “gets too convoluted if you try to make it too big.” He added, “The rules also find it hard to recognize patterns – which is especially important when commerce crosses borders.

Thus pattern recognition is critical, Bloch noted, offering up examples of how patterns can point to fraud in the making – such as how many times an individual seems to be trying to use a card (during a card-not-present interaction) to transact within a set period of time. As he said, “you know these quick transactions would not be physically possible for someone who is walking around with a card.” Other signs can crop up, such as how many times an email address or a phone number is repeated across the various digital and physical channels tied to a company’s sales.

The patterns are there, indeed, but they need to be recognized. They may be too subtle, or too quickly carried away into the rivers of omnichannel commerce to be detected by human observers.

“The strong point of the machine learning model is that it can recognize patterns and consumer behavior … if it is trained correctly,” explained Bloch, who noted that as high-tech as machine learning sounds, “it’s actually, when you get down to the nitty gritty, a pretty basic method of data science.”

Drilling down into the mechanics of constructing a machine learning model, Bloch explained that “you get a huge chunk of data [that is derived] from your customer base. And … in that chunk of data, there is some sort of recognized fraud” pattern that can help build – and in fact inform and shape – the machine learning process, as it builds up a history of both “good” and “bad” transactions and users.

The “brute force” approach uses massive swaths of data to assign features as several models are built, and the best-performing models are the ones put into service. Of course, the models have a limited shelf life, said the executive, and must be updated as the data changes. After all, consumer spending habits change, and can be seasonal – as witnessed in the spikes in activity that may accompany back-to-school shopping or even tax season, when refunds may make consumers feel flush.

Bloch told PYMNTS that “machine learning has traditionally looked at what is ‘bad’ in a transaction, or a set of transactions or a pattern. But if you also start to analyze what is ‘good’ – a particular person, with a particular address, has been buying and having items delivered to their house with no issues for years – that can paint a composite picture.”

According to Bloch, “if the amount of the transaction falls within reason, you are looking for a ‘white’ rather than a ‘black’ label. It’s a form of positive profiling. The chances are very high that this is a good transaction.”

Assigning “white” labels, he told PYMNTS, can help to reduce false positives that can stop transactions in their tracks, keeping merchants from inhibiting good transactions and optimizing their top lines. “We work with many merchants who are able to identify and boost transaction revenue by optimizing a positive profiling feature in the Stream Analytics Engine of our ACI ReD Shield fraud tool. This approach looks at merchants across ACI’s global consortium database to positively enhance accept rates and reduce chargeback rates.”

The Bad Guys Use Machine Learning, Too

It should come as no surprise that the bad guys are also adding machine learning to their arsenal of tricks. “They are smart people,” said Bloch, “and they are using machine learning to ‘poison the data feed’ so it looks good, but it really isn’t.” Ultimately, merchants and financial services firms are tricked into thinking transactions are in fact legitimate.

Bloch pointed out that it is necessary for those who wage war against fraud – and who utilize machine and deep learning – to deploy multiple models simultaneously, as “you have to look at a transaction from many sides. You may have two or three models looking at the same transaction in parallel.” Such an approach can help develop a broader view of an (anonymized) individual’s spending habits.

Threats From the Inside

Machine learning can also be used to safeguard a firm from those who would do harm from within, explained Bloch. “We live in a world that is connected … the person who thinks his or her company is isolated has a short-sighted way of looking at things.”

Executives must look at their perimeter defenses and at how many people access data on a daily basis. Machine learning can pinpoint behavioral patterns showing that some individuals deviate from the way they do their day-to-day jobs, and may be sneaking data out or exposing it to outsiders.

The Importance of Infrastructure

To run deep machine learning takes some high-end hardware, Bloch noted – and that hardware is expensive, existing as a pain point for some firms. But, he said, public infrastructure can enable companies to embrace machine learning that allows them to “pay as they go” – removing costs that may have been a barrier to entry. The machine learning model “must always be on and it must be running all the time,” he said. “Also, it needs to be easy to upgrade, backwards compatible and implemented with an evolutionary rather than revolutionary approach.”

As Bloch explained to PYMNTS: “The architecture plays as vital a role as the machine learning and algorithms themselves.”