Face Morphing Embedding & Evaluation Toolkit

This project provides a toolkit for extracting face embeddings from morphing datasets and evaluating morphing attack detection using state-of-the-art face recognition models.

Features

• Model Selection: Easily switch between ElasticFaceArc, ElasticFaceCos, and CurricularFace models.
• Flexible Dataset Support: Process various morphing datasets (e.g., BonaFide_aligned, FaceMorpher_aligned, etc.) in the SYN-MAD22 directory.
• Batch Embedding Extraction: Efficiently extract and save embeddings for large image sets.
• Morphing Attack Evaluation: Compute metrics such as MMPMR and IAPAR for morphing attack detection.

Project Structure (May be adjusted, models and images are not being shared!)

main.py                  # Main script for embedding extraction and evaluation
backbones/               # Model backbone definitions
embeddings/              # Output folder for embeddings and IDs
models/                  # Pretrained model weights (.pth files)
SYN-MAD22/               # Morphing datasets (aligned images)
triplets/                # Triplet files for evaluation

Quick Start

1. Install Requirements

   • Python 3.8+
   • PyTorch
   • OpenCV
   • tqdm
   • numpy

   You can install dependencies with:

   pip install torch opencv-python tqdm numpy

2. Download Pretrained Models

   • Place the required .pth model files in the models/ directory.

3. Prepare Datasets

   • Place your morphing datasets in the SYN-MAD22/ directory.

4. How to Run

   • Run the evaluation script:

     python main.py --dataset <name_of_the_dataset_you_want_to_evaluate>

   • This will evaluate all the models on the given dataset at an FMR of 0.01 and 0.001.
   • The evaluation can also be specified for a specific FMR and model like this:

     python main.py --dataset <name_of_the_dataset_you_want_to_evaluate> --fmr <fmr> --model <name_of_the_eval_model>

   • The datasets are named just like in the SYN-MAD22 directory (e.g. MorDIFF_aligned).
   • Embeddings and IDs will be saved in the embeddings/ folder.
   • To generate our post-processed morphs run:

     python <name_of_the_attempt_script>.py

   • An example would be:

     python post_process_morphs.py

   • The post-processed datasets will be saved in the SYN-MAD22 directory.

   Running the Detector

   • To run the morphing attack detector, use:

     python detector.py --dataset <name_of_the_dataset_you_want_to_evaluate> --output_directory <optional_output_dir>

   • An example would be:

     python detector.py --dataset FaceMorpher_aligned --output_directory out

   • The detector will output detection results and metrics for the selected dataset.
   • The default output directory is the root directory

5. Results

   • The evaluation script prints MMPMR and IAPAR metrics for the selected model and dataset.
   • The detector gives an output csv of all assessed images and it also prints the EER.

License

This project is for research and educational purposes only. See LICENSE for details.