Official implementation of the paper "Face Morphing Attack Detection with Denoising Diffusion Probabilistic Models" ¹.

MAD-DDPM is a one-class learning model that uses a reconstruction-based measure to determine whether the input images are bona fide or face morphs. At the core of the technique is a two-branch reconstruction procedure that uses denoising diffusion probabilistic models (DDPMs) learned over only bona-fide samples as the basis for the detection tasks. The first branch models the distribution on bona-fide samples directly in the pixel-space (for low-level artifact detection), while the second captures the distribution of higher-level features extracted with a pretrained CNN.

For more information please refer to the paper available here.

The model is implemented using PyTorch. The full list of used libraries can be found in requirements.txt.

pip install -r requirements.txt

The datasets you are using for training or testing should be placed in the directory called data. Datasets should have the following directory structure:

Dataset_name
├── images
│   ├── bonafide
│   │   ├── raw
│   │   │   └── bonafide_img_1.png
│   │   │   └── bonafide_img_2.png
│   │   │   └── ...
│   ├── morphs
│   │   ├── morphing_method_1
│   │   │   └── morph_img_1.png
│   │   │   └── morph_img_2.png
│   │   │   └── ...
│   │   ├── morphing_method_2
│   │   ...
├── features_scale_1
│   ├── bonafide
│   │   ├── raw
│   │   │   └── bonafide_img_1.pt
│   │   │   └── bonafide_img_2.pt
│   │   │   └── ...
│   ├── morphs
│   │   ├── morphing_method_1
│   │   │   └── morph_img_1.pt
│   │   │   └── morph_img_2.pt
│   │   │   └── ...
│   │   ├── morphing_method_2
│   │   ...
├── features_scale_2
│   ...

Images are expected to have one of the following image extensions: '.jpg', '.jpeg' or '.png'. Their corresponding pre-extracted feature maps should be saved with the same name in '.pt' format (PyTorch tensors). features_scale_1 is the root directory of features extracted from cropped face images of size 224x224 pixels (tensor size is 1024x14x14), while features_scale_2 contains features of images of size 448x448 pixels (tensor size is 4x1024x14x14).

MAD-DDPM is trained and tested on preprocessed datasets, where faces were first detected with RetinaFace, then cropped out with a margin of 5% of the detected bounding box height. Corresponding feature maps are extracted with a pretrained WideResNet. For more details please refer to the paper.

To train the image branch MAD-DDPM on your dataset, run the following:

python train.py --train-set ./data/CASIA-WebFace/ --config configs/casia_webface.json --branch 1

To train with evaluation on a testing set after each training epoch run:

python train.py --train-set ./data/CASIA-WebFace/ --config configs/casia_webface.json --branch 1 --test-every 1 --test-set ./data/FRGC/

To train the branch for features, set the value of the argument branch to 2:

python train.py --train-set ./data/CASIA-WebFace/ --config configs/casia_webface.json --branch 2

Checkpoints are by default exported to the directory named output.

To test a pretrained MAD-DDPM model run the following:

python test.py --test-set ./data/FRGC/ --image_branch_checkpoint output/experiment/checkpoints/model_image.pth --features_branch_checkpoint output/experiment/checkpoints/model_features.pth --config configs/casia_webface.json

If you find this code useful or you want to refer to the paper, please cite using the following BibTeX:

@INPROCEEDINGS{ivanovska2023mad_ddpm,
  author={Ivanovska, Marija and Štruc, Vitomir},
  booktitle={2023 11th International Workshop on Biometrics and Forensics (IWBF)}, 
  title={Face Morphing Attack Detection with Denoising Diffusion Probabilistic Models}, 
  year={2023},
  doi={10.1109/IWBF57495.2023.10156877}}}

This code is largely based on k-diffusion.