Visa on using advanced AI such as unsupervised learning to fight fraud

The thing about fraud is that it’s constantly changing — looking at a past attack doesn’t guarantee the next attack will look the same or target the same kind of victim — and defenders have to continuously adapt. Visa utilizes artificial intelligence to analyze all of the transactions that go across the network and track large-scale transactional changes as part of its fraud detection efforts, Melissa McSherry, Visa’s senior VP and global head of data, security, and identity products, said at VentureBeat’s Transform 2021 virtual conference on Monday.

Visa scores all of the transactions that go across the Visa network, which allows the company to define a set of behaviors that would be considered “normal.” The team is “constantly” updating the model’s view of history and updates the model itself to reflect the data on a fairly regular basis, McSherry said.

“The fraudsters do not stand still. And they’re always looking to innovate,” McSherry said in a conversation with Jana Eggers, the CEO of synaptic intelligence company Nara Logics.

Being able to detect changes in the data is useful for authentication, McSherry said. A single phone and email address pair is likely associated with a legitimate transaction, especially if that same pair has been used for a lot of transactions. The next transaction that comes through with the pair will also likely be tracked as legitimate. But if that one phone number is associated with 500 email addresses, it is more likely that all the email addresses are associated with compromised accounts and that the phone number is not associated with a real identity, either.

“It is absolutely the case that the data is constantly evolving, but we take advantage of the velocity of the enormous amount of data that we get and try to evolve our perspective with it,” McSherry said.

Detecting fraud and weaknesses

Everyone knows Visa has an enormous amount of data, and McSherry was able to shed some light on how Visa uses it. Visa has been using neural networks for fraud detection since the 1990s. Eventually, the self-learning technology updated the frame of what was “normal” to identify big deviations in model distributions. More recently, the company is using convolution neural networks (CNN) and recurrent neural networks (RNN) to improve pattern recognition across the network. They aren’t just used as models — but also around the models to identify areas that require more scrutiny or highlight changes that need to be made to the model, McSherry said.

Visa uses generative adversarial networks (GAN) to create virtual fraudsters and pit them against the anti-fraud tools to identify gaps in the fraud-detection models, McSherry said. The gaps can also be in tools provided by partners or in the business logic.

“My experience has been that sometimes things don’t work exactly the way that you think that they’re going to work the very first time that you use them,” McSherry said. Consequently, new methods will be used in parallel, or just for monitoring until they are better understood.

Making the commitment

Incorporating AI requires commitment and follow-through, McSherry said. Visa consistently sees a 20% to 30% lift for advanced AI methods over “more garden-variety technologies,” but this requires heavy investment. Stakeholders need to remain engaged and focused because the first few attempts may not work exactly as planned. Experimentation and patience are key.

“The first and most important thing is just making sure that the problem itself will benefit from those kinds of lifts,” McSherry said. “It’s just really helpful if everybody understands that the value on the other side [of the implementation] is really worth quite a lot.”

Having personnel learn about newer techniques will allow businesses to get the most out of machine learning. While it makes sense to hire new people with strong AI backgrounds, Visa also gave existing employees who understood the business the opportunity to experiment and learn new techniques.

“I think that when people who really have the business context and long-term pride in the quality of the product [are combined] with a really good understanding of AI techniques, that’s when you get something really special,” McSherry said.