Adversarial Mask: Real-World Universal Adversarial Attack on Face Recognition Models

This is a PyTorch implementation of Adversarial Mask: Real-World Universal Adversarial Attack on Face Recognition Models by Alon Zolfi, Shai Avidan, Yuval Elovici, Asaf Shabtai. Mask projection code is partially inspired from FaceXZoo.

Face Recognition Models

Please put the downloaded weights in a local directory called "weights" under each model directory (or change location in the config file).

ArcFace and CosFace

Code is taken from here. Download weights from here.

MagFace

Code is taken from here. Download weights from here.

Landmark Detection Models

MobileFaceNet

Code is taken from here. Download weights from here. (Weights file is already included in this repository under landmark_detection/pytorch_face_landmark/weights).

Face Alignment

Code is taken from here. Weights are downloaded automatically on the first run.

Note: this model is more accurate, however, it is a lot larger than MobileFaceNet and requires a large memory GPU to be able to backpropagate when training the adversarial mask.

3D Face Reconstruction Model

Code is taken from here. Download weights from here. (Weights file is already included in this repository under prnet).

Datasets

CASIA-WebFace

The cleaned up version of the dataset can be found here, suggested by this github issue.

Installation

Install the required packages in req.txt.

Usage

Configuration

Configurations can be changed in the config file.

Train

Run the patch/train.py file.

Test

Run the patch/test.py file. Specify the location of the adversarial mask image in main function.

Citation

@inproceedings{zolfi2023adversarial,
  title={Adversarial Mask: Real-World Universal Adversarial Attack on Face Recognition Models},
  author={Zolfi, Alon and Avidan, Shai and Elovici, Yuval and Shabtai, Asaf},
  booktitle={Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2022, Grenoble, France, September 19--23, 2022, Proceedings, Part III},
  pages={304--320},
  year={2023},
  organization={Springer}
}