Inversion of Deep Facial Templates using Synthetic Data
This package is part of the signal-processing and machine learning toolbox Bob. It contains the source code to reproduce the following paper:
@inproceedings{ijcb2023faceti,
title={Inversion of Deep Facial Templates using Synthetic Data},
author={Shahreza, Hatef Otroshi and Marcel, S{\'e}bastien},
booktitle={2023 IEEE International Joint Conference on Biometrics (IJCB)},
pages={1--8},
year={2023},
organization={IEEE}
}
Installation
The installation instructions are based on conda and works on Linux systems only. Therefore, please install conda before continuing.
For installation, please download the source code of this paper and unpack it. Then, you can create a conda environment with the following command:
$ git clone https://gitlab.idiap.ch/bob/bob.paper.ijcb2023_face_ti
$ cd bob.paper.ijcb2023_face_ti
# create the environment
$ conda create --name bob.paper.ijcb2023_face_ti --file package-list.txt
# or
# $ conda env create -f environment.yml
# activate the environment
$ conda activate bob.paper.ijcb2023_face_ti
# install paper package
$ pip install ./ --no-build-isolation
We use StyleGAN3 as a pretrained face generator network. Therefore, you need to clone its git repository and download available pretrained model:
$ git clone https://github.com/NVlabs/stylegan3.git
We use stylegan3-r-ffhq-1024x1024.pkl
checkpoint in our experiments.
Downloading the datasets
In our experiments, we use FFHQ dataset for training our face reconstruction network. Also we used MOBIO and LFW datasets for evaluation. All of these datasets are publicly available. To download the datasets please refer to their websites:
Downloading Pretrained models
In our experiments, we used different face recognition models. Among which ArcFace and ElasticFace are integrated in Bob and the code automatically downloads the checkpoints. For other models (such AttentionNet, Swin, etc.) we used FaceX-Zoo repository. Therefore you need to download checkpoints from this repositpry (this table) and put in a folder with the following structure:
├── backbones
│ ├── AttentionNet92
│ │ └── Epoch_17.pt
│ ├── HRNet
│ │ └── Epoch_17.pt
│ └── SwinTransformer_S
│ └── Epoch_17.pt
└── heads
You can use other models from FaceX-Zoo and put in this folder with the aforementioned structure.
Configuring the directories of the datasets
Now that you have downloaded the three databases. You need to set the paths to
those in the configuration files. Bob supports a configuration file
(~/.bobrc
) in your home directory to specify where the
databases are located. Please specify the paths for the database like below:
# Setup FFHQ directory
$ bob config set bob.db.ffhq.directory [YOUR_FFHQ_IMAGE_DIRECTORY]
# Setup MOBIO directories
$ bob config set bob.db.mobio.directory [YOUR_MOBIO_IMAGE_DIRECTORY]
$ bob config set bob.db.mobio.annotation_directory [YOUR_MOBIO_ANNOTATION_DIRECTORY]
# Setup LFW directories
$ bob config set bob.db.lfw.directory [YOUR_LFW_IMAGE_DIRECTORY]
$ bob config set bob.bio.face.lfw.annotation_directory [YOUR_LFW_ANNOTATION_DIRECTORY]
If you use FaceX-Zoo models you need to define the paths to the checkpoints of FaceX-Zoo models too:
# Setup LFW directories
$ bob config set facexzoo.checkpoints.directory [YOUR_FACEXZOO_CHECKPOINTS_DIRECTORY]
Running the Experiments
Step 1: Training face reconstruction model
You can train the face reconstruction model by running train.py
. For example, for an attack against ElasticFace
using ArcFace
in loss function, you can use the following commands:
python train.py --path_stylegan_repo <path_stylegan_repo> --path_stylegan_checkpoint <path_stylegan_checkpoint> \
--FR_system ElasticFace --FR_loss ArcFace
Note: Pre-trained models for attacks against ArcFace and ElasticFace models are available in ./checkpoints
folder of this repository.
Step 2: Evaluation
After the model is trained, you can use it to run evaluation.
For evaluation, you can use evaluation_pipeline
script and evaluate on an evaluation dataset (MOBIO/LFW). For example, for evaluation of a face reconstruction of ArcFace on MOBIO dataset, you can use the following commands:
python evaluation_pipeline.py --path_stylegan_repo <path_stylegan_repo> --path_stylegan_checkpoint <path_stylegan_checkpoint> \
--FR_system ArcFace --checkpoint <path_checkpoint> --dataset MOBIO
After you ran the evaluation pipeline, you can use eval_SAR_TMR.py
to caluclate the vulnaribility in terms of Sucess Attack Rate (SAR).
Contact
For questions or reporting issues to this software package, please contact the first author (hatef.otroshi@idiap.ch) or our development mailing list.