When most people think about AI-powered identity fraud, they picture deepfakes. A criminal feeds a few photos of their victim into a tool, generates a convincing fake video, and fools the system into thinking it’s looking at the real person.
This story is compelling – and not exactly wrong. But it misses the more important threat, the one that security researchers have been discussing for the past few years and that the industry has been slow to act on.
In this article, we break down why a device’s camera is the main point of vulnerability in identity fraud and what organizations need to do to safeguard their identity verification flows.
Why Your Camera Is the True Weak Spot
Remote identity verification has become a standard part of how banks, fintechs, insurance companies, telecoms, and others onboard customers online. The process typically works in two steps: the user scans their ID document, and then the system runs a face recognition check to confirm that the person holding the phone matches the photo on the ID.
To make that check harder to fool, most solutions add liveness detection, technology that asks the user to blink, open their mouth, or turn their head, and also scans the camera feed for signs of a printed photo, a 3D mask, or a video playing on a screen. On paper, it sounds robust.
The problem is that all of that security assumes the camera is honest. In a remote verification scenario, the user’s phone is entirely under their control. A motivated attacker does not need to hold a deepfake video up to the camera and hope the artifact detection misses it.
Instead, they can intercept the camera feed at the software level and replace it with their attack video before the security systems ever see it. The system receives what looks like a perfectly normal camera stream. The liveness check passes. The artifact detection finds nothing suspicious. The face recognition runs against the victim’s stolen photo and confirms a match.
Researchers have demonstrated this is possible using tools that are freely available and, in some cases, require only basic technical knowledge to use. One fraud scheme using a variation of this method cost a government system an estimated $76 million over two years.
This is the attack the industry underestimates. Not the deepfake itself, but the method used to deliver it.
How To Secure the Identity Verification Process
To understand what this means for organizations that rely on remote identity verification today, and what they should actually be doing about it, we sat down with Lovre Persen, our Director of Documents and Fraud.
Lovre has spent over 30 years working at the intersection of document verification and financial fraud. He advises clients on identity verification security and sees firsthand how fraud tactics are evolving. We asked him to cut through the noise on injection attacks and what they mean for the industry.
You work with banks and fintechs every day on identity verification. When you mention injection attacks in those conversations, what’s the typical reaction?
Many of our customers read a lot about deepfakes and AI-generated content, but they are not fully aware of how these attacks actually land in their systems, that is, what happens somewhere between the camera and their application.
Most organizations invest heavily in liveness detection and deepfake detection. Why isn’t that enough?
The new ETSI TS 119 461 standard is a great step in this direction, as it specifies the controls organizations should have in place to detect this type of fraud. Beyond that, there is a whole subset of fraud detection tools based on various signals. Depending on your attack surface and risk appetite, these tools help identify fraudulent activity even before it reaches the camera. There is no silver bullet. Today’s sophisticated fraud landscape requires multiple layers of protection working together.
How much technical skill does an attacker actually need to pull off an injection attack today?
Much less than people think. Most fraudsters and fraud rings have little to no technical expertise and rely on readily available tools rather than building attacks themselves. Thankfully, many of these attacks are still relatively simple and should be detectable through the combination of measures we discussed earlier. So, the short answer is: almost none.
Where should organizations be focusing their security efforts if the camera itself is the weak point?
Today, identity fraud prevention is like an onion. It requires multiple layers, and each layer serves a purpose. It’s also important to look beyond marketing claims when selecting a vendor. There is a difference between a vendor simply displaying a certification on their website and a vendor that can clearly demonstrate the scope of that certification across web, Android, and iOS platforms. Some vendors offer injection attack detection protection for biometric flows but not for document verification flows. So before choosing a vendor, read carefully.
As digital identity wallets roll out across Europe under eIDAS 2.0, does this threat get better or worse?
For me, there are two elements to consider. First, we are moving away from simple one-time identity verification toward continuous identity verification. The EUDI Wallet is, in theory and by design, intended to provide the highest level of assurance and protection.
However, in this mindset of “we have the best technology,” we often overlook a very old attack vector that has existed for as long as fraud itself: does the person actually have the right to use that identity? Because of that, I believe various forms of social engineering will become the biggest threat to the EUDI Wallet. Unfortunately, machines are still not very good at detecting these types of attacks.
Final Thoughts
The conversation with Lovre points to something the broader industry debate tends to miss. The focus on deepfakes is understandable because they are visible, dramatic, and easy to explain. But deepfakes are only the payload.
The delivery mechanism, an attacker who has already taken control of the camera before your security systems start running, is the part that current defenses largely cannot stop.
As a QTSP under eIDAS, Evrotrust operates at a level of security and legal accountability that goes well beyond standard identity verification. In a space where the stakes are this high, that distinction matters.
Want to understand what this means for your organization’s onboarding security? Get in touch with our experts.
