Hanrui Wang, Shuo Wang, Zhe Jin, Yandan Wang, Cunjian Chen, Massimo Tistarelli
The majority of adversarial attack techniques perform well against deep face recognition when the full knowledge of the system is revealed (white-box). However, such techniques act unsuccessfully in the gray-box setting where the face templates are unknown to the attackers. In this work, we propose a similarity-based gray-box adversarial attack (SGADV) technique.
This is a single-task attack. We have a new Multi-task version, which targets more challenging scenarios.
eagerpy (0.30.0)
The versions in ()
have been tested.
Source image name must satisfy 00000_0.jpg
. 00000
and _0
indicates the image id and user id/class/label, respectively. The image id must be unique and auto-increment from 00000
. .jpg
can be any image file format.
20 source samples have been prepared for the demo.
InsightFace: iresnet100 pretrained using the CASIA dataset; automatically downloaded
FaceNet: InceptionResnetV1 pretrained using the VggFace2 dataset; automatically downloaded
Run attack:
python SGADV.py
Objective function: foolbox/attacks/gradient_descent_base.py
New developed tools: foolbox/utils.py
Filter objects of CelebA: tools/fetch_celebAhq.py
Feature embeddings and save to .mat: tools/feature_embedding.py
Dataset | EER (%) | ASR - White box(%) | ASR - Gray box(%) |
---|---|---|---|
FaceNet | 1.2 | 100 | 98.74 |
InsightFace | 6.23 | 100 | 93.23 |
If using this project in your research, please cite our paper.
@inproceedings{wang2021similarity,
title={Similarity-based Gray-box Adversarial Attack Against Deep Face Recognition},
author={Wang, Hanrui and Wang, Shuo and Jin, Zhe and Wang, Yandan and Chen, Cunjian and Tistarelli, Massimo},
booktitle={2021 16th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2021)},
pages={1--8},
year={2021},
}
The code in the folder foolbox is derived from the project foolbox.
Images in the folder data are only examples from LFW and CelebA dataset.
If you have any questions about our work, please do not hesitate to contact us by email.
Hanrui Wang: hanrui_wang@nii.ac.jp