The new fraud frontiers

In the late 19th century, the French police officer Alphonse Bertillon had a radical idea: What if, rather than relying on imprecise techniques of interrogation and investigation, law enforcement could identify criminals through unique physical characteristics? Although flawed, “Bertillonage” laid the foundation for fingerprint identification, and thus the modern science of biometrics⁴.  

  The ubiquity of smartphones with high-quality cameras and touch sensors, combined with advances in computing power and artificial intelligence, mean that biometrics are now an everyday part of the user experience for countless businesses. From logging into an account to completing a payment, users can verify interactions via their unique biological markers⁵. The field of biometrics is evolving fast, and its cutting edge now encompasses everything from facial and voice ID to palm recognition and long-range iris scans. 

The strength of biometrics as a security measure, however, has paradoxically created a blind spot for businesses. When a biometric test is passed, it is tempting to accept that at face value. But as we have seen, generative AI can produce convincing deepfakes, and these can fool facial and voice checks—even ones enhanced to detect signs of “liveness”. Breaking biometric security can allow fraudsters to commit identity theft, onboard to other products such as authenticator apps, or simply carry out fraudulent transactions. 

Techniques to help businesses combat deepfakes are emerging. One is the use of residual neural networks, which are deep machine learning models that have an aptitude for detecting falsified elements on images and videos. Another involves “challenges” that aim to expose deepfakes by exploiting the limitations of the models behind them. Nasir Memon, the Dean of Computer Science at New York University (NYU) Shanghai, has created a tool, as yet largely untested, that involves cascading challenges. These can be active (requiring the participant to do something such as poking their cheek, say, or holding an object in front of their face), or passive (such as changing the scene illumination by projecting structured light patterns onto the subject). Deepfakes can struggle to pass such tests. “If I don’t know whether you’re a deepfake, I can probe you in some way, send some data in your direction and see how things come back,” he says. Ideas like this can not only help prevent biometric failures,  but could also thwart the kind of social engineering attacks that led to the Hong Kong fraud.  

An alternative avenue of research focuses on using secondary factors to support the biometric test. This might involve inspecting aspects of the input for consistency. If the facial scan purportedly comes from a phone whose camera offers a particular resolution, for example, it would clearly be a red flag if the system was looking at a face with a different resolution. Another possibility is to involve secret knowledge; for example a voice-based system might ask the user to say a phrase that only they know. 

Since the ongoing evolution of technology threatens to make many anti-deepfake techniques almost obsolete as soon as they see the light of day, that kind of holistic thinking is essential. “Deepfakes are a greater concern when a business is only assessing a limited number of factors for risk. Good fraud detection should involve assessing a bunch of different factors in aggregate,” says Mike Frost, Senior Product Specialist at J.P. Morgan. He also cites changing payment instructions, use of a new phone number, a change of device or an account being accessed at atypical times of the day as examples of factors that can be evaluated to help build a detailed picture of suspicious activity. 

“The tactics that we’ve tried to use are less related to technology and more about practices,” agrees his colleague Steven Bufferd, Managing Director of Trust & Safety Products at J.P. Morgan. “How do we validate it? Are there sources we can trust? Is the change of payment instruction legitimate? Is that a real person? Can we validate based on the information you’ve given us? Trying to ascertain whether a person is real should not just be a technological endeavor, but one in which technology assists and augments human-based best practices.” 

To protect data, businesses typically encrypt it. But data then has to be decrypted before it can be used—including for the detection of fraud itself—at which point it potentially becomes vulnerable. Enter homomorphic encryption (HE), which has its origins at the Massachusetts Institute of Technology (MIT) in 1978. HE enables computation to be performed on data while it remains encrypted.  

Vinod Vaikuntanathan, an MIT Professor of Electrical Engineering and Computer Science,  who is the co-inventor of most modern fully homomorphic encryption systems explains the concept as follow: “You can think of traditional encryption as putting your data inside a locked box. You send that locked box over a channel and the person at the other end has a key. They can unlock the box and get the answer,” he says. “Homomorphic encryption, on a basic level, is being able to take a locked box and manipulate the contents within without opening it up.” 

While this has obvious advantages for data security—businesses could ask a third-party provider of powerful cloud computers to work on data without introducing risk, for example—there is a specific use case for HE in detecting fraud, because HE has the potential to facilitate safe data sharing. In financial services, multiple entities could securely work together to identify suspicious activity. “If you think of two or more banks, they each have their own data about their customers’ transactions. But fraud often operates across many banks, often with small transactions that flow across multiple accounts,” says Vaikuntanathan. “It’s very hard to detect, if your visibility is limited to your own network. So, can we enable people to compute across all that data, and identify suspicious operations? That is something that is being worked on.” 

HE is not yet a widespread technology. It’s secure, and simple computations often take just seconds to execute, but when extrapolated to large and complex datasets, especially when paired with equally complex algorithms, there is a significant computational “overhead.” This makes scale implementations often impractical in a real-world financial setting¹³. However, this may change. “That will either happen through finding ways to rewrite your computations so they’re more amenable to homomorphic approaches, which could be assisted by machine learning or via better hardware. There is lots happening with hardware—although unlike with AI, it’s not the compute that’s the bottleneck, it’s memory.” 

If the problems can be solved, HE has an opportunity to proliferate—perhaps to the stage where we never need to decrypt encrypted data. “I think we have a reasonable shot of making it in, let’s say, a decade,” says Vaikuntanathan. “I’m putting all my time and energy where my conviction is, which is in inventing new techniques to make this faster.”