This repository contains code for the above-mentioned paper. Specifically, we provide a script for using a pre-trained TetraLoss model and MAD model to obtain comparison/detection scores from the corresponding models.
To use this repository, ensure you have the following installed:
- PyTorch
- Numpy
- scikit-learn
The repository was tested using Python 3.8.16 and specific versions outlined in the requirements.txt file.
The implementation was tested on Linux using NVIDIA A100 Tensor Core GPUs.
This repository includes a pre-trained TetraLoss model and a Morphing Attack Detection (MAD) model for evaluation on bona fide and morphed images.
To run this code, you can use the provided pre-trained models:
-
TetraLoss AdaFace ResNet100 model:
Link: https://cloud.h-da.de/s/Qx8BCQrsG7DpcwM
Password: CzYb6TyABKjoYa -
MAD model:
Link: https://cloud.h-da.de/s/Nj6bABtScEqP6q7
Password: mmREmj6g4DjqRA
Before running the model, extract face embeddings using the adaface_ir101_ms1mv2.ckpt file from the AdaFace repository. Ensure you use the specific version of AdaFace based on ResNet100 (ir_101), trained on MS1MV2. Follow the general inference guidelines from the AdaFace repository to extract and save embeddings from your dataset. The code expects that the embeddings are numpy arrays saved using np.savetxt and are loaded as numpy arrays in the code (see utils/helper.py).
To test the model, a script, test.py, is provided. It takes as input a reference (suspected) and a probe image and produces the following:
- TetraLoss comparison score
- MAD score
- Combined score from the TetraLoss FR model and the MAD system
The resulting scores are written to the standard output.
The script can be run as follows:
python test.py --checkpoint_tetra <path_to_tetra_checkpoint> \
--checkpoint_mad <path_to_mad_checkpoint> \
--path_original_reference_emb <reference_embedding_file_path> \
--path_original_probe_emb <probe_embedding_file_path>For more information about the parameters of the test script, see python test.py --help
If the same images are used in multiple comparisons, it is recommended to first extract the embeddings and save them (e.g., to disk) and then calculate the similarity scores.
Information to obtain the checkpoints are provided in the Pre-trained models section.
Part of the code in this repository is adapted from the Self-restrained Triplet Loss repository. The used MAD algorithm is based on the approach described in the paper Deep Face Representations for Differential Morphing Attack Detection.
If you find this useful for your research, please consider citing our paper:
TetraLoss: Improving the Robustness of Face Recognition against Morphing Attacks:
@inproceedings{Ibsen-TetraLoss-FG-2024,
Author = {M. Ibsen and L. J. Gonzalez-Soler and C. Rathgeb and C. Busch},
Booktitle = {{IEEE} Intl. Conf. on Automatic Face and Gesture Recognition ({FG})},
Title = {TetraLoss: Improving the Robustness of Face Recognition against Morphing Attacks},
Year = {2024}
}
Additionally, consider citing the following works:
Self-restrained triplet loss for accurate masked face recognition:
@article{Fadi-SelfRestrainedTLForMaskedFR-PatternRecognition-2022,
title = {Self-restrained triplet loss for accurate masked face recognition},
journal = {Pattern Recognition},
year = {2022},
author = {F. Boutros and N. Damer and F. Kirchbuchner and A. Kuijper},
volume = {124},
issn = {0031-3203}
}
Deep Face Representations for Differential Morphing Attack Detection:
@article{Scherhag-FaceMorphingAttacks-TIFS-2020,
Author = {U. Scherhag and C. Rathgeb and J. Merkle and Christoph Busch},
Journal = {{IEEE} Trans. on Information Forensics and Security},
Title = {Deep Face Representations for Differential Morphing Attack Detection},
Year = {2020},
volume={15}
}