Generative models can reconstruct face images from encoded representations (templates) bearing remarkable likeness to the original face raising security and privacy concerns. We present FaceCloak, a neural network framework that protects face templates by generating smart, renewable binary cloaks. Our method proactively thwarts inversion attacks by cloaking face templates with unique disruptors synthesized from a single face template on the fly while provably retaining biometric utility and unlinkability. Our cloaked templates can suppress sensitive attributes while generalizing to novel feature extraction schemes and outperforms leading baselines in terms of biometric matching and resiliency to reconstruction attacks. FaceCloak-based matching is extremely fast (inference time cost=0.28ms) and light-weight (0.57MB). Refer to our paper: FaceCloak
Run scripts/FaceCloak.py with the following arguments. Please check the script for details.
folder_path = '/path/to/synthetictemplates'
gallery_path = str(sys.argv[1]) # filename1 from cfpfp_pair.txt or lfw_pair.txt
probe_path = str(sys.argv[2]) # filename2 from cfpfp_pair.txt or lfw_pair.txt
results_path = str(sys.argv[3]) # path to save results
num_epochs = int(sys.argv[4]) # default = 100
num_featureaug = int(sys.argv[5]) # default = 50 (so total 101 (=2 * 50 disruptors +1 unprotected template))
bit_size = int(sys.argv[6]) # default = 64
margin = float(sys.argv[7]) # default = 9.0
dataset_name = str(sys.argv[8) # default = 'lfw' Options: ['lfw', 'cfpfp']
We run the experiments on CFPFP and LFW datasets. Ensure the images are in your working directory and please follow the instructions included as comments in the script. In order to be able to run FaceCloak.py, please make sure to extract the face templates first using either one of three temaplte extractors below:
- InsightFace ArcFace -->
scripts/ExtractTemplate.py - FaceNet512 -->
scripts/ExtractTemplate_FaceNet512.py - ArcFaceOPL -->
scripts/ExtractTemplate_ArcFaceOPL.py(Refer to ArcFace-OPL below for details)
We use StyleGAN3 with truncation_factor=0.7 for generating synthetic faces to be used as disruptor templates. Check out the gen_images.py script from the official repository. (Link)
This is only for generalizability test. We use VGGFace dataset provided by InsightFace code repo for training ArcFace with Orthogonal Projection Loss. We provide pretrained model in this link. Kindly follow the insructions from the official InsightFace library to download insightface and put model.pt under models_zoo folder before extracting face templates using scripts/ExtractTemplate_ArcFaceOPL.py. Link
Note ExtractTemplate_ArcFaceOPL.py is adapted from inference.py Set arguments --network r50 --weight fullpath-to-model_zoo/model.pt --img -path-to-image
Run scripts/FaceCloak_BiometricMatching.py for evaluating biometric performance, scripts/FaceCloak_Irreversibility.py for irreversibility evaluation and scripts/FaceCloak_Unlinkability.py for unlinkability evaluation. Note FaceCloak is trained everytime differently, so, repeat the experiments to compute and report the average performance which may slightly differ from the paper.
Run scripts/GenderPrediction.py to evaluate the gender prediction results. We have provided the ground truth geneder labels extracted from LFW.py using DeepFace.analyze in data/LFW_Gender.csv.
Run scripts/Arc2Face_Disruptor.py to visualize the face images corresponding to the disruptors created using scripts/FaceCloak.py. This is borrowed from Arc2Face github repo. Please follow their installation instructions.
Acknowledgment
If you find this work to be useful, please cite
@article{banerjee2025facecloaklearningprotectface,
title={FaceCloak: Learning to Protect Face Templates},
author={Sudipta Banerjee and Anubhav Jain and Chinmay Hegde and Nasir Memon},
year={2025},
journal={IEEE International Conference on Automatic Face and Gesture Recognition (FG)},
note={Accepted}
}