deepfake-detect

Deepfake Model Card

EfficientNet-B4, fine-tuned on a merged real/fake face manifest: StyleGAN1 (manjilkarki/deepfake-and-real-images), StyleGAN3 (troykueh/real-vs-fake-faces-stylegan3), diffusion (mohannadaymansalah/stable-diffusion-dataaaaaaaaa).

Real photos are pooled across sources (content-hash deduped); fakes are kept per-generator so cross-generator generalization can be measured directly instead of inferred from pooled accuracy.

Held-out test AUC, by generator family

Each generator family is scored as its own fakes against the pooled real photos, so the numbers are directly comparable and a fake-only source (diffusion) is measurable at all.

• stylegan1: 0.997
• stylegan3: 0.998
• diffusion: 1.000
• _pooled: 0.997
• _mean_per_source: 0.998

Selection criterion during training: best mean-per-source validation AUC (not pooled accuracy), so the largest source dataset can't dominate checkpoint selection.