How they work:
- Per-frame scoring: each video frame gets a manipulation probability, aggregated across the clip. NVIDIA's detector hits 99.3% on benchmarks this way.
- Audio stream analysis: Pindrop-style systems detect synthetic voices in about two seconds of live audio by checking for physiological signals like breathing.
- Liveness challenges: unpredictable prompts that pre-recorded or generated media cannot answer.
Their limits:
- Latency versus accuracy. Faster verdicts use fewer frames and miss subtle fakes. - Compression. Real-time calls are heavily compressed, which erases the artifacts detectors look for. - Adversarial adaptation. Attackers test against public detectors and tune around them. - The human factor. A detector that flags 2% of real calls as fake will be disabled by week two. Sources: CISA For real-time deepfake detection in onboarding, treat it as a speed bump for attackers, not a wall. One layer, never the only one.
Real-time deepfake detection matters most in live onboarding video, where a single session decides the account.