# MAP2V

Validating Privacy-Preserving Face Recognition under a Minimum Assumption To be presented in the IEEE/CVF Conference on Computer Vision and Pattern Recognition 2024

Abstract

The widespread use of cloud-based face recognition technology raises privacy concerns, as unauthorized access to face images can expose personal information or be exploited for fraudulent purposes. In response, privacy-preserving face recognition (PPFR) schemes have emerged to hide visual information and impede recovery. However, their privacy validation methods, built on unrealistic assumptions, raise doubts about proclaimed face privacy preservation capabilities. In this paper, we introduce a general privacy-preserving validation version under the novel concept of minimum assumption coined Map$^2$V. This is the first exploration of formulating a privacy validation method utilizing deep image priors and zeroth-order gradient estimation, with the potential to serve as a generalizable tool for PPFR evaluation. Building upon Map$^2$V, we comprehensively validate the anti-privacy inference capability of PPFRs through a combination of human and machine vision. The results of experiments and analysis demonstrate the effectiveness and generalizability of the proposed Map$^2$V, showcasing its superiority over native privacy validation methods from PPFR works of literature. Additionally, this work exposes privacy vulnerabilities in evaluated state-of-the-art PPFR schemes, laying the foundation for the subsequent effective proposal of countermeasures.

Requirements

• PyTorch 1.9.1
• Torchvision 0.10.1
• CUDA 10.1/10.2

Setup

Download pretrained privacy-preserving face recognition model and weights: DuetFace DCTDP PartialFace

[all_modelweights] (https://pan.baidu.com/s/1OmW93OEEXHTxDrH-gk1rw?pwd=2fb4) Fetch Code: 2fb4 or google drive: https://drive.google.com/drive/folders/1SWxALbbZ8gB0Kt8-ADK-_8vg1GVxFwpu?usp=sharing

After downloading, place all model weights in the directory encoder/pretrained and change the paths in encoder/encoder_conf.yaml.

If you need to validate your own PPFR model, you can update the model and weights to your content.

Download LFW and CelebA datasets:

LFW

CelebA Fetch Code: fd8o

After downloading, change the paths in dataset/dataset_conf.yaml.

Download the latents from the constructed Prior Space:

latents Fetch Code: onrr

After downloading, place the initial latent spaces in latents/ and change the paths in reconstruct.py when the initialization type selects the prior space(FFHQ).

Download the weight of generator（here is stylegan） to generate face images:

generator Fetch Code: yjng

After downloading, place the weights in the directory generator/.

Usage

After the setup is done, simply run validate.py and change the path args.save_dir to save results.

Experimental Results

Privacy scores (%) against different validation systems on LFW and CelebA dataset under 1k1c settings. Examples of reconstructed faces for three SOTAs under 1k1c settings.

Citation

If you find this code useful in your research, please consider citing us.