Biometric image forensics
Is this face photo manipulated?
Upload a face photo. A neural network trained on face-swap and GAN faces returns a probability, shows the exact crop it scored, and refuses to guess when the photo is too small or blurry.
The image is analysed in memory and is not stored. Typical time: a few seconds, longer on the first request.
Prototype. Not a security guarantee and not proof. A probability from a model trained on one manipulation method and a small set of synthetic faces. Read the method and limits before relying on any result.
What it measured
On held-out test images from its training sources, video-level AUC 0.946 and 1.5% of real faces wrongly flagged. See the full table on the method page.
What it does not do
It catches only about 72% of face-swap fakes and 36% of StyleGAN faces at its strict threshold, and it was not tested on modern AI image generators. Treat it as a screening aid.
How it is built
Xception network, MTCNN face detector and a compression-matching step, all running as ONNX models in a Python function on Vercel. Method and accuracy