mahedi420/deepfake-vit-detector

Explainable Deepfake Detection — project checkpoints

Real/fake face classifiers from the thesis project Explainable Deepfake Detection using Vision Transformer. All files are PyTorch checkpoints {'epoch', 'model_state', 'optimizer_state'}; label 0 = real, 1 = fake; input = 224×224 MTCNN face crop, ImageNet-normalised.

File	Architecture	Training data
model.pt	ViT-B/16 (timm vit_base_patch16_224)	FaceForensics++ c23 + Celeb-DF-v2
freqfusion_model.pt	Dual-branch spatial-frequency ViT	FaceForensics++ c23 + Celeb-DF-v2
xception_model.pt	Xception (timm legacy_xception)	FaceForensics++ c23 + Celeb-DF-v2
ffpp_only_model.pt	ViT-B/16	FaceForensics++ c23 only
celebdf_only_model.pt	ViT-B/16	Celeb-DF-v2 only
custom/*_custom.pt	same five architectures	the file above it, fine-tuned on a custom set of AI-generated portrait re-renders (293 train crops) with replay of the original domain

Custom-data fine-tuned checkpoints (custom/)

File	Custom test acc / ROC-AUC (n=64)	Original in-domain acc (n=3000 subset)
custom/model_custom.pt	0.844 / 0.898 (was 0.500 / 0.502)	0.969 (was 0.973)
custom/freqfusion_model_custom.pt	0.797 / 0.879 (was 0.469 / 0.495)	0.969 (was 0.974)
custom/xception_model_custom.pt	0.844 / 0.934 (was 0.453 / 0.465)	0.952 (was 0.971)
custom/ffpp_only_model_custom.pt	0.875 / 0.929 (was 0.484 / 0.452)	0.961 (was 0.970)
custom/celebdf_only_model_custom.pt	0.875 / 0.977 (was 0.500 / 0.429)	0.960 (was 0.982)

custom/custom_finetune_results.json holds every metric and per-epoch history; custom/custom_manifest.csv is the identity-disjoint split of the custom set.

from huggingface_hub import hf_hub_download
import timm, torch
ckpt = torch.load(hf_hub_download("mahedi420/deepfake-vit-detector", "custom/model_custom.pt"), map_location="cpu")
model = timm.create_model("vit_base_patch16_224", pretrained=False, num_classes=2)
model.load_state_dict(ckpt["model_state"]); model.eval()