Many know one of the most standard KYC methods used by fintechs: Holding your ID up to your face and allowing biometrics to confirm your identity via selfie. But as swindlers have leveraged AI to become more savvy, knowing your customer has become increasingly complicated.
Photo by Alexey Demidov on Unsplash
From injection attacks to deepfakes, biometric technology is very much under threat. Still, experts say facial recognition can remain a heavyweight champion of digital verification — as long as fintechs beef up their stack.
“Fintechs come from this ‘We can build everything we need model,’ right? And when it comes to security, that doesn't always hold right,” Nishant Kaushik, CTO of identity technology security community and research company FIDO Alliance, told This Week in Fintech. “The challenge isn't so much that biometrics are bad or that facial recognition is unable to keep up with us, it is that as the attack vectors have shifted and the technology on the biometrics and facial recognition side have gotten better.”
According to an April report by biometric tech company Aware, 56% of organizations reported revenue loss when it came to AI fraud impacting biometric security. Deloitte’s Center for Financial Services predicted fraud losses in banking could rise to as much as $40 billion in the U.S. by 2027, from US$12.3 billion in 2023, as generative AI becomes more advanced and pervasive.
Injection attacks are one of the most common ways a criminal can bypass biometrics. Using a different video and injecting it into a device’s actual live input stream, fraudsters can easily fake the identity of anyone using deepfakes or synthetic clips. Injection attacks can take place even on audio calls.
“These KYC systems are part of financial institutions. There's an economic motivation to break them,” Tom Cross, the head of threat research at GetReal Security, told This Week in Fintech. “Whenever there's an economic motivation, you end up with people very focused on breaking that.”
Cross said that in much of his work researching deepfakes, technology has become advanced enough to easily bypass many biometric services.
“We’ve tried a number of different open source and commercial facial recognition tools…and we are consistently able to bypass open source and commercial biometric technologies — face and voice — using deepfake tools that are available,” Cross said.
To hammer the point home, he turned on his very own deepfake face during the interview.
“The bottom line is that biometrics alone cannot be trusted in the AI era,” Cross said.
Some of the solutions proposed include tech built into the verification model in order to determine a camera’s “liveness,” metadata intelligence or analysis of devices used during authentication. This can often conflict with the need to ensure that customers feel the least amount of friction as banks work to confirm their identity.
“What you are trying to solve is something that your user can actually do in a very practical way,” Kaushik said. “The idea of taking a selfie is easy enough to convey to a user, rather than trying to give them something really complicated.”
Another way to strengthen security is by verifying more frequently, Cross says, and ensuring each interaction matches for every customer.
“Threat actors are interested in creating accounts at financial institutions under stolen identities, and they know that that KYC check is going to be there, and they're very focused on passing that one-time check,” Cross told TWIF. “What's important is looking at all the interactions that occur with this individual, and determining if they're consistent with each other.”
AI agents pose a new and difficult verification problem. The rise of products like Meta’s Muse, OpenAI’s Dots and Instinct are now forcing financial institutions to determine whether an AI agent is authorized to act on behalf of their customers too. And big banks are taking notice – JPMorgan announced earlier this month that it would be blocking agents like Muse from directly accessing a customer’s bank account.
“The future involves a certain amount of consumer use of authorized deepfakes, right? And then how do we get to a place where we can get comfortable allowing consumers to use those tools,” Cross said. “We need these AI-based agents to be able to prove that they have authorization to do the things that they're that they're asking to do.”
