A deepfake of a world leader appearing on a live news broadcast hasn't happened yet. When it does, every network that doesn't have real-time detection will scramble. V100 has it built into the relay layer today.
Three Models, One Score
V100's deepfake detection runs a 3-model ensemble on sampled frames (every 5th frame at 30fps = 6 inferences per second):
- Facial Artifact Detection (ONNX) — A neural network trained on millions of real and synthetic faces identifies pixel-level artifacts that GANs leave behind: inconsistent skin texture, misaligned facial geometry, unnatural eye reflections
- Temporal Consistency Analysis — Compares consecutive frames for jitter that doesn't match natural head movement. Deepfakes produce frame-to-frame facial geometry changes that violate physics — micro-teleportations that human vision can't detect but algorithms can
- Frequency Domain Analysis — 2D FFT of face regions reveals spectral signatures of GAN upsampling. Transposed convolutions in neural network generators create checkerboard patterns at specific spatial frequencies invisible to the eye but unmistakable in Fourier space
Composite score: 0.5 * facial + 0.3 * temporal + 0.2 * frequency. Range: 0.0 (definitely real) to 1.0 (definitely fake).
Alert System
Three alert types fire automatically:
- Threshold — Single frame exceeds configured score (default 0.7)
- Spike — Score jumps by >0.3 within 1 second (someone swapped a feed)
- Sustained — Score stays above threshold for >5 seconds continuously
Alerts dispatch via Redis pub/sub (internal) and webhook (external) simultaneously. A news network's master control can receive the alert, hold the feed, and switch to a backup source — all within seconds.
Why At the Relay Layer?
Most deepfake detection tools are post-hoc — you upload a video and get a report hours later. That's useless for live broadcast. V100 runs detection at the relay layer because that's where every video frame passes through anyway. There's no additional latency, no separate upload step, and no way for a compromised upstream feed to bypass the check. If it reaches viewers, it was scanned.
The Business Case
News networks won't pay for deepfake detection today because there hasn't been a headline incident yet. But the technology to create convincing real-time deepfakes is advancing faster than the technology to detect them. V100 offers detection as a built-in feature — not an add-on — which means networks get it as part of their broadcast infrastructure. When the incident happens, V100 customers are already protected.
POST /api/v1/deepfake/enable — Enable on any ingest feed
GET /api/v1/deepfake/{id}/scores — Rolling average + per-frame data
POST /api/v1/deepfake/thresholds — Configure alert sensitivity
GET /api/v1/deepfake/{id}/alerts — Alert history