Deepfake detection: lessons from the field

Deepfakes fail in weird edges. We built a triage system that surfaces frequency anomalies, facial landmark drift, and transcript sentiment in one dashboard.

Our best-performing stack was hybrid: MFCC features for audio, CLIP embeddings for frames, and a late-fusion classifier tuned on hard negatives.

Ground truth comes from humans. We paid for annotation rather than trusting synthetic labels—it doubled precision overnight.

Three guardrails we keep

• Every alert links to raw media slices for manual review.
• Confidence scores are conservative and paired with narrative explanations.
• We log adversarial samples for weekly red-team sessions.

Future work: streaming inference with WebRTC taps and watermark detection on-device.