Demographic Bias in Deep Face Recognition in The Wild

This is the official repository of the paper entitled "Demographic Bias in Deep Face Recognition in The Wild", published in: IEEE Journal of Selected Topics in Signal Processing (Volume: 17, Issue: 3, May 2023)

The original paper is available here.

We provide a Pytorch toolbox for Face Image Degradation (1) and Face Recognition training and testing (2).

1. The Image Degradation module provides a training part for both GANs and their Discriminators and a use/evaluation part that:

• Degrades given datasets

• Shows the results of the degradation in relation to other types of algorithmic disturbances and to original images acquired in unconstrained environments

2. The Face Recognition module provides a training part with various SOTA Face Recognition backbones and heads and an evaluation part that:

• Provides evaluations of the given model(s) in order to obtain metrics like ROC curves, AUCs and EERs.

• Provides metrics as FRR@FARs (FNIR@FPIRs) Degree of Bias across multiple factors (e.g.: gender and ethnicity and their combinations) and imposed security thresholds.

Requirements

• Python >= 3.7
• PyTorch >= 1.10.0
• DeepFace == 0.0.72
• MatPlotLib == 3.5.2
• Scikit Learn == 1.0.1
• Scipy == 1.7.1
• Seaborn == 0.11.2

In order to install all the necessary prerequisites, you can simply execute the following command:
pip install -r requirements.txt

Degradation Module

GAN Training

See README in src/image_degradation_rq1/training_module folder

GAN Evaluation

See README in src/1_image_degradation/evaluation_module folder

Image Degradation

See README in src/1_image_degradation/degradation_module folder

Face Recognition Module

Model Training

See README in src/2_face_recognition/training

Dataset Preprocessing and combined models training

See README in src/2_face_recognition folder

Model Evaluation

See README in src/2_face_recognition/experimental folder

Contributing

This code is provided for educational purposes and aims to facilitate the reproduction of our results and further research in this direction. We have done our best to document, refactor, and test the code before publication.

If you find any bugs or would like to contribute new models, training protocols, etc, please let us know.

Please feel free to file issues and pull requests on the repo and we will address them as we can.

License

This code is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.

This software is distributed in the hope that it will be useful, but without any warranty; without even the implied warranty of merchantability or fitness for a particular purpose. See the GNU General Public License for details.

You should have received a copy of the GNU General Public License along with this source code. If not, go the following link: http://www.gnu.org/licenses/.

Acknowledgements

This work is an extension of Explaining Bias in Deep Face Recognition via Image Characteristics, which has been published in the International Joint Conference on Biometrics (IJCB 2022) proceedings. (Available here)

@ARTICLE{10054108,
  author={Atzori, Andrea and Fenu, Gianni and Marras, Mirko},
  journal={IEEE Journal of Selected Topics in Signal Processing}, 
  title={Demographic Bias in Low-Resolution Deep Face Recognition in the Wild}, 
  year={2023},
  volume={17},
  number={3},
  pages={599-611},
  doi={10.1109/JSTSP.2023.3249485}}