Vulnerability of State-of-the-Art Face Recognition Models to Template Inversion Attack

This package is part of the signal-processing and machine learning toolbox Bob. It contains the source code to reproduce the following paper:

@article{tifs2024vulnerability,
  title={Vulnerability of State-of-the-Art Face Recognition Models to Template Inversion Attack},
  author={Shahreza, Hatef Otroshi and Hahn, Vedrana Krivoku{\'c}a and Marcel, S{\'e}bastien},
  journal={IEEE Transactions on Information Forensics and Security},
  year={2024},
  publisher={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:

$ cd bob.paper.tifs2024_face_ti

# create the environment
$ conda create --name bob.paper.tifs2024_face_ti --file package-list.txt
# or conda env create -f environment.yml

$ conda activate bob.paper.tifs2024_face_ti  # activate the environment
$ buildout

Downloading the datasets

In our experiments, we use FFHQ dataset for training our face reconstruction network. Also we used MOBIO, LFW, and AgeDB datasets for evaluation. All of these datasets are publicly available. To download the datasets please refer to their websites:

• FFHQ
• MOBIO
• LFW
• AgeDB

Configuring the directories of the datasets

Now that you have downloaded the four 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]

# Setup AgeDB directories
$ bob config set  bob.db.agedb.directory [YOUR_AGEDB_IMAGE_DIRECTORY]

Running the Experiments

Step 1: Generating Training Dataset

You can use the GenDataset.py in DataGen folder to generate training dataset for embeddings of a terget face recognition model. For example, for ArcFace you can use the following commands:

cd experiments/DataGen
python GenDataset.py --FR_system ArcFace

Note: If you want to use face recognition models from FaceXZoo, you need to run with ./bin/python, where bin folder should be generated after buildout in in the installation step.

Step 2: Training Face reconstruction model

After the data is generated, you can train the face reconstruction model by running train.py in TrainNetwork folder. For example, for ArcFace you can use the following commands:

cd experiments/TrainNetwork
python train.py  --FR_system ArcFace

Step 3: Evaluation

After the model is trained, you can use it to run evaluation. For evaluation, you can use eval folder in the directory of each model and then open the folder of an evaluation dataset (MOBIO/LFW/AgeDB). Then, you can run the evaluation using ./run_pipeline.sh. For example, for evaluation of ArcFace on MOBIO dataset, you can use the following commands:

cd experiments/eval/MOBIO
./run_pipeline.sh

You can change the face recognition in ./run_pipeline.sh script. After you ran the evaluation pipeline, you can use eval_hist_plot.py, eval_SAR_plot.py, eval_SAR_values.py to find the score distributions and calculate the vulnerability in terms of Sucess Attack Rate (SAR).

Contact

For questions or reporting issues to this software package, contact the first author or our development mailing list.