Deepfakes are increasingly used by attackers to target biometric processes. Reports from Biometric Update and Goode Intelligence project significant growth in deepfake fraud detection and injection attack detection (IAD) from 2026 to 2028. At the FIDO Alliance's Authenticate 2026 conference, experts highlighted a crucial point: presentation attack detection (PAD) and IAD are distinct challenges, even though both fall under "liveness."
This checklist is for Identity Governance Administrators responsible for biometric verification flows. Each item is actionable, with clear indicators of completion.
Prerequisites
Before you begin, ensure you have:
- An inventory of workflows using biometrics for enrollment, authentication, or account recovery.
- Vendor documentation for each biometric component, including PAD or IAD test reports or certifications.
- Access to your SIEM and biometric component logs.
- A designated owner for biometric risk, separate from the vendor relationship owner.
Checklist
1. Document PAD and IAD as separate controls.
PAD deals with what is presented to the sensor, like masks or replayed videos. IAD involves what is fed into the system, such as virtual cameras or tampered data streams. Ensure your control register lists them separately, with distinct owners or evidence sources.
Good looks like: Independent PAD and IAD controls with unique test evidence for each.
2. Map every biometric capture point to its attack surface.
Identify where capture occurs in each workflow: native mobile app, mobile browser, desktop webcam, or kiosk. Injection risks vary greatly across these.
Good looks like: A table of capture points with noted injection and presentation exposures, reviewed with any flow changes.
3. Ask vendors which attack class each claim covers.
"Liveness detection" is vague. Require vendors to specify if results cover presentation attacks, injection attacks, or both, and under what conditions.
Good looks like: Vendor claims in your files are tagged PAD, IAD, or both, with cited test sources.
4. Check for deepfake-specific evidence.
Conventional PAD testing may not account for AI-generated faces and voices. Verify if the vendor's evaluation includes deepfake scenarios and current techniques.
Good looks like: Recent test reports reflecting current attacker tools, with a retesting schedule.
5. Confirm what the certification actually covers.
Biometric certification is evolving. Don't assume a previous certificate covers injection attacks. Review the scope statement and test criteria version.
Good looks like: A certification tracker listing scope, criteria version, and expiry for each component, with renewal reviews scheduled.
6. Add injection defenses at the capture layer.
Where you control the client, use device integrity signals, camera source verification, and signed capture data to reject untrusted streams.
Good looks like: The server rejects unsigned or unattested data by default, demonstrated with a test.
7. Use AI defensively, with human-reviewed tuning.
AI can detect synthetic artifacts and inconsistencies. Treat these models as controls needing regular tuning and monitoring.
Good looks like: Documented detection thresholds, scheduled reviews of false outcomes, and controlled model updates.
8. Don't let biometrics stand alone for high-risk actions.
Combine biometric checks with a second signal for account recovery, credential reset, and privilege elevation. Phishing-Resistant Authentication using FIDO2 credentials is a strong complement.
Good looks like: No high-risk workflow relies solely on a biometric match.
9. Feed biometric events into your SIEM.
Log liveness failures, anomalies, retries, and device changes. Injection attempts often show patterns.
Good looks like: Correlation rules for retry bursts and shared artifacts, with alerts triaged by your team.
10. Run a tabletop for a deepfake-enabled takeover.
Simulate an attacker using synthetic media to pass remote identity verification. Decide who revokes credentials, reviews entitlements, and detects entitlement drift.
Good looks like: A written playbook with named roles and completed exercises, with findings tracked to closure.
11. Include biometric-enrolled identities in Certification Campaigns.
Revisit accounts enrolled through weak remote verification. Flag identities enrolled before IAD controls and consider step-up re-verification.
Good looks like: A defined cohort of pre-control enrollments with a re-verification plan and completion date.
Common Mistakes
- Treating "liveness" as one box to tick. A vendor may excel in PAD but not IAD. The umbrella term can obscure gaps.
- Trusting a certificate without reading its scope. Criteria evolve, and older certifications may not address injection or deepfakes.
- Testing only with physical artifacts. If validation doesn't involve virtual cameras or stream tampering, IAD hasn't been tested.
- Relying on the vendor to monitor. Detection tuning and alert triage should be your team's responsibility, with vendor input.
- Using biometrics as the sole gate for recovery. Attackers target recovery processes to bypass primary authenticators.
Next Steps
Start with items 1 through 3. They're cost-effective and reveal your real exposure. Then, focus on workflows with the weakest capture environments and highest-value outcomes, like remote enrollment and account recovery. Revisit item 5 with each certification update, and schedule regular reviews to ensure your controls address current threats, not outdated assumptions.
If you can't identify which controls address presentation attacks and which address injection attacks, that's your first finding.
