Video Deepfake Detection

Bynn Video Deepfake Detection

The Bynn Video Deepfake Detection model identifies face swaps in video: footage in which a real person's face has been replaced, frame after frame, by a face-swapping tool. Built by Bynn for identity verification, it returns a clear decision, a probability, per-frame scores and the face-swapping method the content most resembles.

The Challenge

Video deepfakes represent an escalation in manipulation sophistication. Real-time face-swapping tools maintain facial consistency across frames, match lighting and expression, and follow head movement, so the result blinks, turns and follows liveness challenges like a real person.

For KYC and identity verification, this undermines the premise of video-based authentication. A manipulated feed can show a stolen document's face on an attacker's moving head. Presentation attack detection looks for printed photos, masks and screens, so it sees a real, moving face and lets it through.

Model Overview

The Bynn Deepfake Detection model is Bynn's own face-swap detector, built for identity verification: it is designed for the selfie frames and videos real verification sessions produce, and it is kept current as new face-swapping tools appear. It returns a probability, a clear decision and the face-swapping method the content most resembles.

Achieving 99.8% accuracy, with 99.99% of face-swap attacks detected at a false reject rate of 0.32% on held-out verification sessions, it provides robust protection against face-swap attacks on identity verification and authentication flows.

How It Works

• Frame extraction: 32 frames per minute of video are taken evenly through the clip
• Face analysis: the face in each frame is located and analyzed for the traces a face swap leaves behind
• Per-frame scores: every frame receives its own fake probability, so you can see which frames carry the manipulation
• One decision: the frame scores are combined into a single probability and a clear real or fake decision for the whole input
• Attribution: the response names the face-swapping method the content most resembles, for investigation and reporting

Request

Send a video with video_url or base64_video (MP4, MOV, AVI, WebM or MKV, up to 100MB). 32 frames per minute of video are taken evenly through the clip and combined into one decision, and the per-frame results in the response show which frames carry the manipulation. Send the file as it was recorded, without re-encoding or downscaling it first.

Billing: per minute of video, with a partial minute counted as a full minute.

Response Structure

The API returns a structured JSON response containing:

• label: String, "fake" or "real", or "unverifiable" when no face could be analyzed
• is_fake: Boolean, true when a face swap was detected; null when unverifiable
• is_real: Boolean, the inverse of is_fake; null when unverifiable
• fake_probability: Float (0.0 to 1.0), probability that the content is a face swap
• confidence: Float (0.0 to 1.0), the model's confidence in the returned label
• frames_total and frames_scored: Integer, frames received and frames in which a face was analyzed
• swapper: String, the face-swapping method the content most resembles, "real" for real content
• frame_probabilities: Array of floats, the fake probability of each frame in order, null where no face was found

Example Response:

{
  "label": "fake",
  "is_fake": true,
  "is_real": false,
  "fake_probability": 0.998,
  "confidence": 0.998,
  "frames_total": 32,
  "frames_scored": 31,
  "swapper": "inswapper_128",
  "frame_probabilities": [0.998, 0.997, 0.996, null, 0.998, 0.995, 0.997, 0.998, ...]
}
Code
Copy

Note: an "unverifiable" result means no face could be analyzed. Treat it as content to retry or review, not as evidence of a real person.

Performance Metrics

Use Cases

• Video KYC: check a selfie video or liveness session for a face swap before matching the face to the document
• Video Onboarding and Meetings: flag a participant who joined a call through a face-swapping tool
• Content Platforms: screen uploaded clips in which a real person's face was replaced
• Media Verification: verify footage of real people before publication

Known Limitations

• Face swaps only: the model detects a real face that was replaced. Fully AI-generated faces are covered by AI-Generated Image Detection, and printed photos, masks and screen replays by Face Liveness Detection. Use them together for full coverage.
• One face per frame: the most prominent face is analyzed. Crop first if a specific face in a group matters.
• Image quality: very small, blurred or partially visible faces reduce reliability.
• Generator evolution: new face-swapping tools appear constantly; the model is updated as they do, and detection effectiveness should be monitored.

Disclaimers

This model provides a probability and a decision, not definitive proof of manipulation.

• Screening Tool: use it as part of a multi-layered verification strategy together with liveness, document checks and face matching, not as the sole decision factor
• Human Review: route flagged high-stakes decisions to a reviewer
• Not Legal Evidence: a probability is not a finding of fraud