1. Work 01
  2. Craft 02
  3. Research 03
  4. About 04
  5. Writing 05
  6. Contact 06
Email LinkedIn GitHub CV
All work

05 / 12 · AI · Forensics

DeepTrace / Deepfake Detector

Exploring how synthetic media can be identified, analyzed and understood.

01

92% fake

The first version I put online looked at a real webcam clip and called it 92% fake. I had built the thing I was worried about: a confident number that anyone could believe.

02

Starting again

I swapped in two DFDC-winning EfficientNet-B7 models and a Community Forensics ViT, measured it on 163 videos and 205 photos, and taught it to say inconclusive instead of guessing. It still wrongly flags 2 of 88 real videos, and this page says so.

03

Who decides

That raised a question code could not answer: who acts on a probabilistic verdict, and on whose authority? It became forty days of research in Malaysia and Singapore, and an essay.

04

On device

The models now also run on a Mac or iPad, converted to Core ML, so a photo never has to leave the device to be checked.

DeepTrace: main screen
Category
AI · Digital forensics
Stack
Python · Streamlit · PyTorch
Models
EfficientNet-B7 (DFDC winner) · Community Forensics ViT
Hosting
Hugging Face Spaces

What it does

  1. Checks videos and photos up to 200MB for face swaps, re-enactment and AI-generated images.
  2. 16 frames are sampled evenly across a clip; YuNet finds every face and each person is scored separately.
  3. Two EfficientNet-B7 networks from the winning Deepfake Detection Challenge entry score each face; the Community Forensics detector checks for AI generation.
  4. The file's own metadata is read for AI labels, generator settings and Content Credentials.
  5. Three answers: no manipulation found, likely manipulated, or inconclusive. A face too dark or blurry to judge gets no verdict.
  6. Heatmaps show where on the face the evidence sits, and every case exports as an HTML evidence report.
  7. Measured before release: 2 of 88 real videos wrongly flagged, 45 of 75 fakes caught. The full evaluation is public.
  8. It began as a Streamlit prototype on MesoNet-4, which called ordinary webcam footage fake and had to be replaced.

The detector was the starting point. It raised a question I could not answer with code: who acts on a probabilistic verdict, and on whose authority? That question became the comparative governance research further down this page.