01Yassine/defender-lora

LoRA Deepfake Detector

Self-contained inference package for the LoRA-augmented deepfake detector.

• Backbone: facebook/dinov3-vith16plus-pretrain-lvd1689m (LoRA on attention projections)
• Input: RGB image, letterboxed to 512×512
• Output: prob_fake in [0, 1] — label fake if prob_fake >= 0.5

Checkpoints

File	Training step	Notes
lora_75k.pth	75,000	Default
lora_115k.pth	115,000	Later checkpoint

Both were trained with HUGE_LORA_aug (augmentation on, 512px).

Setup

pip install -r requirements.txt

You must accept the DINOv3 license and be logged in:

huggingface-cli login
# or: export HF_TOKEN=...

Usage

CLI

python infer.py path/to/image.jpg
python infer.py path/to/image.jpg --weights lora_115k.pth --device cuda

Example output:

fake (0.8734)

Python

from lora import Model

model = Model.load(device="cuda")
result = model.predict("photo.jpg", device="cuda")
print(result)  # {"label": "fake", "prob_fake": 0.87}

Model architecture

1. DINOv3 ViT-H+ with LoRA (r=16, alpha=32) on q/k/v/o_proj
2. MMFuser — multi-layer feature fusion
3. Head — L2-normalized CLS + mean patch tokens → linear → binary softmax

Preprocessing uses aspect-preserving letterbox to 512 (no stretch).

Files

• lora.py — Model class and preprocessing
• infer.py — command-line inference script
• lora_75k.pth / lora_115k.pth — fine-tuned weights (state dict)
• config.json — metadata

Citation

DFKI / GAI Joint project.