DeepGuard

Know when a face
has been swapped

A detector trained on 10,852 images, half of them face swaps. Upload a photo and it returns a verdict, a confidence score, and a heatmap showing which regions gave it away.

Detection you can inspect

A fine-tuned EfficientNet-B4 backbone with gradient-based explanations, measured across five generation methods.

Trained on face swaps

EfficientNet-B4 fine-tuned on 10,852 images built with inswapper, SimSwap and DeepFaceLab.

Shows its working

GradCAM heatmaps mark the regions that moved the score, so a verdict is never just a number.

Honest about limits

Measured against five generation methods, including the ones where accuracy drops off.

It works better on some fakes than others

Accuracy holds above 89% on the face-swap pipelines the model was trained on, and falls away on diffusion-generated faces it never saw. Published rather than hidden.

StyleGAN294.8%
inswapper_12891.5%
FaceSwap89.3%
DeepFaceLab85.1%
Stable Diffusion78.4%