Liveness Detection: Stopping Deepfakes at Onboarding

A selfie matched against an ID photo answers one question: do the faces look alike? It does not answer whether a real person sat in front of the camera. Attackers exploit that gap with printed photos, replayed video, silicone masks, and increasingly, AI-generated faces injected directly into the verification stream. Liveness detection exists to close it.

Two Attack Classes, Two Defenses

Fraud against face verification splits into two categories, and they require different controls.

• Presentation attacks put a fake artifact in front of a genuine camera: a photo on a phone screen, a high-resolution print, a video loop, or a 3D mask. Defenses look for texture, depth, reflections, moiré patterns, and micro-movements a flat or rigid surface cannot produce.
• Injection attacks bypass the camera entirely. Using a virtual camera driver or a compromised SDK, the attacker feeds a pre-recorded or synthetically generated frame straight into the app. No physical artifact exists to detect, so the defense shifts to device integrity, stream provenance, and signals that the frames originate from real hardware.

Many teams deploy strong presentation-attack detection and leave injection wide open. As deepfake tooling gets cheaper, injection is the faster-growing threat.

Active vs Passive Liveness

Liveness checks come in two operational styles. Active liveness asks the user to perform an action: turn their head, blink, follow a dot, or read numbers aloud. It produces a clear challenge-response record but adds friction and can be defeated by responsive deepfakes that animate on command.

Passive liveness analyzes a single capture or short burst without instructions, inspecting skin texture, depth cues, lighting consistency, and sensor noise. It is faster and less abandonment-prone, but harder to explain to an auditor because the decision is model-driven.

A practical configuration combines them: passive analysis on every attempt, with an active challenge escalated only when the passive score is ambiguous or the risk profile is elevated. This mirrors step-up logic elsewhere in onboarding and keeps friction proportional to risk.

When evaluating vendors, ask for certification against ISO/IEC 30107-3, which measures Attack Presentation Classification Error Rate (APCER) and Bona Fide Presentation Classification Error Rate (BPCER). A single accuracy number tells you nothing; you need both the miss rate on attacks and the false-reject rate on genuine users, tested by an accredited lab against defined attack instrument levels.

Privacy Engineering for Biometric Capture

Liveness introduces biometric data, which carries heightened obligations under regimes such as the EU GDPR (Article 9 special categories), Illinois BIPA, and Brazil's LGPD. The engineering goal is to prove liveness without accumulating a biometric archive you cannot defend.

• Process, then discard. Where your risk assessment allows, treat the liveness frames as transient. Retain the pass/fail result, the model score, and a hash or audit token rather than the raw video.
• Separate storage. Keep biometric artifacts in a distinct store with its own encryption keys and access controls, decoupled from name, address, and account data.
• Set explicit retention. Configure deletion timelines per jurisdiction and document the legal basis. BIPA, for example, requires a written retention schedule and destruction when the purpose is satisfied.
• Capture consent that is specific. Generic terms acceptance rarely covers biometric processing; the disclosure should name the data, purpose, and retention period.

Fitting Liveness Into a Chat Flow

Verification increasingly happens inside messaging channels rather than dedicated apps. In a chat-based flow, the user is guided step by step, and the liveness capture becomes one gated stage: the applicant submits the required frames, the check runs, and only a minimal result is retained for the compliance record. Because the conversation controls what is asked and when, you can apply data minimization by default and escalate to an active challenge only for higher-risk sessions. The outcome is a verifiable liveness decision, an audit trail examiners can follow, and no standing pool of raw biometric footage waiting to become a liability.

This is general information, not legal advice; confirm biometric handling requirements with counsel for each market you serve.