Fairness and Face Obfuscation

Background

Supporting code for the paper: Fairness Properties of Face Recognition and Obfuscation Systems -- https://arxiv.org/abs/2108.02707. We evaluated the demographic fairness properties of several face privacy systems.

Installation

Using python version 3.8.8

git clone https://github.com/wi-pi/fairness_face_obfuscation.git
conda update -n base -c defaults conda
conda create -n fairnessfaces python=3.8
conda activate fairnessfaces
conda install -c anaconda tensorflow-gpu
pip install -r requirements.txt
pip install -e .

or

conda env create -f environment.yml

Required weights, embeddings, datasets

Download the following folders and place them in the ./data directory: https://drive.google.com/drive/folders/1xWoGESQEtRPZhIEut-9ID2Yyu25lJDaz?usp=share_link

To download the all embeddings, download embeddings/all_embeds/ and move the files to the ./data/embeddings/attributes/ folder. To download only the non-adversarial embeddings, download embeddings/embeds/ and move the files to the ./data/embeddings/attributes/ folder.

Resources

Results reported in the paper were obtained using a server with 40 CPU cores, 2 Nvidia TITAN Xp's, and 1 Quadro P6000, 125 GB Memory, Ubuntu version 16.04 LTS, CUDA 10.0, NVIDIA Driver 410.104.

Disclaimer: The code has not yet been tested on a variety of platforms.

See https://stackoverflow.com/questions/50622525/which-tensorflow-and-cuda-version-combinations-are-compatible for CUDA - TensorFlow compatibility.

API evaluation

Note that any API evaluation requires accounts, keys, and an AWS S3 bucket. Below are some links to resources helpful for setting up keys. Follow the step-by-step instructions found in the below links.

AWS S3: https://docs.aws.amazon.com/AmazonS3/latest/dev/access-points.html

AWS S3 - Public Read Access: https://aws.amazon.com/premiumsupport/knowledge-center/read-access-objects-s3-bucket/

Follow the instructions in these 3 links to obtain your public and private keys. Add them to the FaceAPI/credentials.py file.

AWS Rekognition: https://docs.aws.amazon.com/rekognition/latest/dg/getting-started.html

Azure Face: https://azure.microsoft.com/en-us/services/cognitive-services/face/#get-started

Face++: https://console.faceplusplus.com/documents/7079083

Development

Setup

./scripts/create_directories.sh

Creates the necessary subdirectories. Code creates subdirectories in data/new_adv_imgs.

Generation

./scripts/fairness_attack.sh

Generates perturbations on a single image-class pair. Code generates perturbations on faces, and outputs results in data/new_adv_imgs and data/adversarial. Successful generation will be printed through each iteration.

NOTE: If you want to use hinge loss, you must align (detect, crop, and resize) a bucket of your own faces to sizes 160x160 or 96x96. You can use MTCNN to do so. We will integrate support for this shortly.

Amplification

./scripts/amplify.sh

Amplifies the perturbations by an amplification factor (constant at the top of src/amplify.py).

Uploading

./scripts s3_upload.sh

Uploads specified images to AWS S3 bucket for API evaluation.

Embeddings

./scripts/store_embeddings.sh
./scripts/store_adv_embeddings.sh

Creates embeddings using a given model and outputs to a .npz file in data/embeddings/attributes

Evaluation

./scripts/validate_on_lfw.sh
./scripts/norm_distance.sh
./scripts/api_eval.sh
./scripts/tcav_all.sh
./lipschitz/customrecurjac/scripts/newvgg<1-12>.sh
python src/read_lipschitz_constants.py

Evaluates natural and adversarial accuracy metrics. Also evaluates L2 norms of perturbations. Evaluates adversarial examples on APIs like Face++. Runs TCAV evaluation of skin tones on face recognition model. Runs Lipschitz constant estimations. Finally parses Lipschitz constant estimation results and outputs to csv.

Plotting

./scripts/plot.sh
./scripts/plot_accuracy.sh
./scripts/parse_norms.sh
python src/plot_tcav.py
./scripts/plot_lipschitz.sh

Plots TSNE of embeddings. Plots adversarial and natural accuracy metrics. Plots distribution of L2 perturbation norms. Plots TCAV distributions for identities and skin tones. Plots distribution of estimated Lipschitz constants.

Link to the papers

https://arxiv.org/abs/2108.02707

Citation

@misc{rosenberg2021fairness,
      title={Fairness Properties of Face Recognition and Obfuscation Systems}, 
      author={Harrison Rosenberg and Brian Tang and Kassem Fawaz and Somesh Jha},
      year={2021},
      eprint={2108.02707},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Contact

Questions? Contact [email protected] or [email protected] with subject: Fairness Faces