MinusFace

Model of the CVPR 2024 paper, Privacy-Preserving Face Recognition Using Trainable Feature Subtraction.

---

Overview

MinusFace is a privacy-preserving face recognition method that transforms an original face image into a visually uninformative protective representation while preserving identity information for face recognition.

The key idea is to learn a high-dimensional residual representation through feature subtraction between the original face and its model-produced regeneration. The residual features are further randomized through channel shuffling and decoded into the final protective representation. The generative and recognition models are collaboratively trained so that the resulting representations remain recognizable while reducing recoverable visual information.

For implementation and training details, please refer to the paper and the official TFace repository.

Citation

If you use this model in your research, please cite:

@InProceedings{mi2024minusface,
    author    = {Mi, Yuxi and Zhong, Zhizhou and Huang, Yuge and Ji, Jiazhen and Xu, Jianqing and Wang, Jun and Wang, Shaoming and Ding, Shouhong and Zhou, Shuigeng},
    title     = {Privacy-Preserving Face Recognition Using Trainable Feature Subtraction},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2024},
    pages     = {297-307}
}

Contact

yxmi20 (at) fudan.edu.cn