FIDO testing experts address deepfake threat and biometrics certification changes

Reports from Biometric Update and Goode Intelligence forecast market growth of 42 and 99 percent CAGR from 2026 to 2028, respectively, for deepfake fraud detection and injection attack detection (IAD). A session at the FIDO Alliance’s Authenticate 2026 conference last week in California delved into the practical implications of this situation for relying parties and more broadly for those across the identity industry. University of North Carolina – Charlotte Professor and CITeR Director Stephanie Schuckers and Ingenium Biometrics Laboratory CTO and Co-founder Chris Allgrove co-presented the Wednesday session.

They began by outlining how biometric presentation attack detection (PAD) and IAD need to be understood and often addressed as different things, even though Schuckers says they both belong under the umbrella term “liveness.” 3 attack points and 6 methods Deepfakes and less sophisticated spoofs can be injected at three different points in the biometrics capture flow, Allgrove explained. They take place in the camera feed, within the app or device itself or between the app and the remote orchestration, integration or decision. Vendors do not even take consistent views of what they are responsible for in this flow, Allgrove said at the end of the presentation in response to an audience question.

While some might consider the data beyond their scope once it has been processed within the app, while others “provide a different security paradigm, almost.” Allgrove divides injection attack methods into six different buckets: software and hardware virtual cameras, external video capture cards and mobile device emulators all involve the use of a modified or fake capture device. In function hooking and man-in-the-middle attacks, captured data is replaced with the injection attack instrument. But there is a lot of variance between these methods.

Emulators, for instance, “mostly don’t work” in terms of beating IAD systems, according to Allgrove. Allgrove also discussed how different attack methods and instruments are detected. Reflections on the windows of the soul reveal deepfakes Schuckers shared an example of an IAD method based on “eye specular reflections.” A shape is shown on the user’s device which reflects off of the eye and is captured along with the rest of the image.

The shapes can be randomized, and the order of their appearance must match. “From an attacker point of view, now they have to have a deepfake that would not only capture the specular reflection in the pupil, but also be able to randomly switch in real time not anticipating which shapes are shown.” This “really raises the bar” for attackers, Schuckers says, emphasizing that the solution is just one of many available that appear to be effective. Face swap fails Schuckers shared work being conducted in her lab into face swapping apps.

A pair of examples show the range of performance, with one matching every frame of the target video, while another looked convincing enough at a glance, but was unable to replicate any of the subject’s movement in real-time, so none of the video frames matched. “Why are we doing that? Because we found out that the benchmark datasets that are out there for training systems, most of them didn’t match the target individual,” Schuckers says.

Aligning FIDO biometrics certification with 800-63 Schuckers also emphasized the importance of lab certification (like FIDO’s) to the value of tests those labs perform, and ran through the requirements for FIDO’s biometric component and face verification certification tests. For the latter certification, FIDO has added requirements to version 5.0 of its certification requirements for tighter alignment with the NIST SP 800-63 Digital Identity Guidelines. That means more subjects in Level 2 tests and lower thresholds for both false acceptance rate (FAR) and false rejection rate (FRR).

The benefit of this change goes two ways, Schuckers explains, as 800-63A specifies that biometrics must be independently evaluated.