Deepfakes can be injected at three different points in a biometric check, experts told the FIDO Alliance’s Authenticate 2026 conference. Their account also details tougher requirements in FIDO’s version 5.0 face-verification certification, including more subjects in Level 2 tests and lower false-acceptance and false-rejection thresholds.
Watch Desk analysis
What happened
Stephanie Schuckers of the University of North Carolina at Charlotte and Chris Allgrove of Ingenium Biometrics Laboratory discussed injection attack detection at the conference. Biometric Update reports that they identified three possible points of attack: the camera feed, the app or device, and the link between an app and remote processing or decisions.
Allgrove grouped injection methods into six categories, including virtual cameras, video capture cards, emulators, function hooking and man-in-the-middle attacks. He said methods vary, and that emulators mostly do not defeat injection-attack detection systems.
Schuckers described one detection technique that shows changing shapes on a user’s device and checks whether their reflections appear in the person’s eyes in the expected sequence. She said this raises the bar for a deepfake that would need to reproduce those reflections in real time. The report also says FIDO’s version 5.0 face-verification certification requirements align more closely with NIST’s SP 800-63 digital identity guidelines.
Why it matters
A biometric check is not just a camera asking whether a face looks familiar. Attacks may target the capture device, the software handling the image or the connection to later processing, so security responsibilities can fall between organisations. The experts said vendors do not always agree on where those responsibilities begin and end.
The certification changes matter because they affect how face-verification systems are tested: Level 2 tests require more subjects, with lower thresholds for both false acceptance and false rejection. That makes the testing standard, not just the marketing claim, part of the trust question.
Our read
The useful distinction is between presentation attacks, such as a spoof shown to a camera, and injection attacks that feed altered data into the process. Calling both “liveness” may be convenient; treating them as the same security problem is less so. The eye-reflection technique is an example, not a guarantee that deepfakes are solved.
What to watch
- How FIDO’s version 5.0 certification requirements are applied in practice.
- Whether vendors clarify responsibility across the capture, app and remote-processing stages.
- How detection methods perform against different injection techniques and real-world deepfakes.
Discussion spark: Should biometric security standards require vendors to secure the whole capture-to-decision chain, or is responsibility better divided between device makers, apps and service providers?
Sources and evidence
- Source update (29 October 2026, 18:16 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.