Skip to content
SecTepeMit Sicherheit gut aufgestellt!
DE
Glossary

AI Biometrics: Securing Biometric Authentication

AI-based biometrics verify faces, voices or fingerprints. Which attacks (deepfakes, presentation attacks) matter and what GDPR and the AI Act require.

Definition: AI-based biometric authentication checks who a person is by their body or behavior. Typical traits are face, fingerprint, voice or typing rhythm. Neural networks turn the trait into a math template. They then compare it with a stored reference. Biometric traits are not secret, and you cannot replace them. So security depends mainly on two things: spotting fake input and protecting the templates.

How AI-driven biometrics work

A biometric system works in two phases:

  • Enrolment: The system captures the trait. A model turns it into a feature vector, and the system stores it as a template.
  • Verification: A new sample goes through the same steps. The system compares it with the template. If the match is above a set threshold, the person gets in.

Keep two cases apart. Verification is 1:1 and asks "Is this person X?", as when you unlock a laptop. Identification is 1:n and asks "Who is this?", as when you search a database. Login at work almost always means verification. That is much less sensitive, both in legal and in technical terms.

You judge a system by its error rates:

  • The false acceptance rate (FAR, or FMR) shows how often the system lets in impostors.
  • The false rejection rate (FRR, or FNMR) shows how often it turns away the right users.

Both depend on the threshold. If you lower one, the other goes up.

Attacks on biometric systems

  • Presentation attacks: The attacker holds a fake up to the sensor. This can be a photo, a video on a screen, a silicone mask or a copied fingerprint.
  • Injection attacks: The attacker does not fool the sensor. Instead, they feed fake data straight into the data stream, for example through a virtual camera or an emulator. Generative AI makes deepfake videos and cloned voices. These make such attacks look real, above all in remote ID checks such as video identification or voice hotlines.
  • Morphing: Two face images are blended into one. The result is accepted for both people. This is a known problem with passport photos.
  • Adversarial examples: Inputs that are changed on purpose, such as printed glasses frames, have fooled face recognition models in research.
  • Template theft: In some cases, attackers can use stolen templates to rebuild a trait or to attack other systems. A password can be reset. A face cannot.

Fairness matters too. If the training data was unbalanced, models can make more errors for some groups of people. That is an ethical issue. It is also a problem of access when the system keeps turning away the same staff.

Practical safeguards

  1. Treat biometrics as a factor, not a secret: Combine biometrics with something the user owns, such as a device with a security chip. FIDO2 and passkeys are a proven pattern. The match happens on the device and only unlocks a cryptographic key. The template never leaves the device.
  2. Require liveness detection: Look for presentation attack detection (PAD) tested against ISO/IEC 30107. For remote use, also look for protection against injection attacks.
  3. Protect templates: Store them encrypted, at best spread out or in secure hardware. Prefer template protection schemes (see ISO/IEC 24745) that let you revoke a template.
  4. Set thresholds on purpose: Pick FAR and FRR to fit how much protection you need. Check the error rates for each group of users on its own.
  5. Secure the fallback path: When biometrics fail, users fall back to a PIN, a helpdesk call or a spare card. This path is often the weakest link. Reset processes need their own ID checks.
  6. Never rely on voice alone: For payment approvals by phone, always call back on a known number or use a second channel.

Legal framework

GDPR: Biometric data used to single out one person is a special category under Art. 9 GDPR. You may not process it unless an exception applies. At work, it is hard to show that consent was given freely. So offer a non-biometric option of equal value. A data protection impact assessment usually makes sense for biometric processing. In Germany, the works council also has a say when you bring in such systems (Section 87(1) no. 6 of the Works Constitution Act, BetrVG).

EU AI Act: Biometric verification that only confirms a claimed identity is not on the high-risk list in Annex III. Other uses are high-risk where they are not banned outright. These are remote biometric identification, biometric categorization and emotion recognition. After the Digital Omnibus, the Annex III high-risk duties apply from 2 December 2027. Art. 5 bans some practices. These include emotion recognition at work (except for medical or safety reasons). They also include scraping face images from the internet at random to build databases.

Conclusion

AI makes biometric login easy and accurate. But generative AI also makes fakes cheaper. Biometrics only become secure when you add cryptographic factors, tested liveness detection and solid fallback processes. SecTepe can help you assess them and build them into your identity and access management.

Frequently asked questions

Does a fingerprint-unlocked passkey mean the employer processes biometric data?

With FIDO2 passkeys, the match happens on the device. The employer never gets any templates. This makes the data protection review much simpler. Still, agree on the exact rating with your data protection officer.

Is voice recognition enough to approve payments?

No. Voice clones have become far too real. Voice recognition can be one helpful signal. But always combine it with a separate, independent factor.

More terms in “AI & Machine Learning”

All terms in “AI & Machine Learning” →