xade-deepfake-detector

XADE Deepfake Detector

EfficientNet-B4 model trained for deepfake detection as part of the XADE (eXplainable Automated Deepfake Evaluation) thesis project at Jönköping University, 2026.

Model Details

• Architecture: EfficientNet-B4 with custom two-layer classifier head
• Task: Binary classification (real vs. fake faces)
• Training: Progressive mixed training across 4 manipulation types
• Final checkpoint: Run 4 (140k + CIPLAB + FF++ + Celeb-DF)

Cross-Dataset Performance (AUC-ROC)

Training Details

• Base dataset: 140k Real and Fake Faces (StyleGAN-generated)
• Additional training data: CIPLAB (~960 images), FF++ derived (~1500 frames), Celeb-DF v2 (~1500 images)
• Training samples: 100,000 per run (sampled from combined pool)
• Epochs: 10 (early stopping patience 7)
• Optimizer: AdamW with differential learning rates (backbone: 1e-4, classifier: 1e-3)
• Batch size: 64
• Validation accuracy: 98.51%

Architecture

EfficientNet-B4 (ImageNet pretrained, last 30% unfrozen)
└── Custom classifier head:
    Dropout(0.5)
    Linear(in_features → 512)
    ReLU
    BatchNorm1d(512)
    Dropout(0.4)
    Linear(512 → 2)

Usage

import torch
from huggingface_hub import hf_hub_download
from torchvision.models import efficientnet_b4
import torch.nn as nn

# Download model
model_path = hf_hub_download(
    repo_id="viktorahnstrom/xade-deepfake-detector",
    filename="best_model.pt"
)

# Load checkpoint
checkpoint = torch.load(model_path, map_location="cpu", weights_only=False)
print(f"Trained for {checkpoint['epoch']} epochs")
print(f"Classes: {checkpoint['class_names']}")  # ['fake', 'real']

Citation

@misc{xade2026,
  author = {Viktor Ahnström and Viktor Carlsson},
  title = {XADE: Cross-Platform Explainable Deepfake Detection
           Using Vision-Language Models},
  year = {2026},
  institution = {Jönköping University},
  howpublished = {\url{https://huggingface.co/viktorahnstrom/xade-deepfake-detector}}
}