[CVPR 2024] Generative Unlearning for Any Identity

Official PyTorch implementation for CVPR 2024 paper:

Generative Unlearning for Any Identity
Juwon Seo*, Sung-Hoon Lee*, Tae-Young Lee*, Seungjun Moon, and Gyeong-Moon Park^$\dagger$

Environment

• Python 3.9.x
• PyTorch 2.1.0
• Torchvision 0.16.0
• NVIDIA GeForce RTX 3090 / A5000 / A6000
• CUDA 11.8

Getting Started

Environment

git clone [email protected]/KHU-AGI/GUIDE.git
cd GUIDE
conda env create -f environment.yaml
conda activate guide

Files to Run Code

Filename	Description
ffhqrebalanced512-128.pkl	Pretrained weights of EG3D network.
w_avg_ffhqrebalanced512-128.pt	Average latent code computed from pretrained EG3D network.
model_ir_se50.pth	Pretrained weights of ArcFace network to compute identity loss.
encoder_FFHQ.pt	Pretrained weights of GOAEncoder for 3D GAN inversion.
CurricularFace_Backbone.pth	Pretrained weights of CurricularFace network to compute identity similarity.
ffhq_real_feat.npy	Statistics of Inception-v3 feature to easily compute ΔFIDreal.

To run code, please download all of files via Google Drive.

Erase an identity from pretrained 3D GAN

We provide sample data (./data/CelebAHQ/512) to run our code.
The commands for Random and In-Domain scenarios will be updated in near future.

# Baseline
python unlearn.py --exp baseline \
                  --inversion goae \
                  --inversion_image_path ./data/CelebAHQ/512 \
                  --target average \
                  --local \
                  --seed 0

# GUIDE
python unlearn.py --exp guide \
                  --inversion goae \
                  --inversion_image_path ./data/CelebAHQ/512 \
                  --target extra \
                  --target_d 30.0 \
                  --local \
                  --adj \
                  --glob \
                  --seed 0

It takes about 15 minutes in a single 3090 GPU.

Erase in the wild identities

If you want to erasure in the wild identities, please preprocess the images via Deep3DFaceRecon.
Otherwise, both the inversion encoder and pretrained generatior can't recognize them correctly.

Evaluation

Evaluation on identity erasure:

python evaluate_id.py --exp [baseline | guide]

Evaluation on pretrained distribution preservation:

python evaluate_fid.py --exp [baseline | guide]

By running the above commands, we could obtain the results of:

	ID	IDothers	FIDpre	ΔFIDreal
baseline	0.0818	0.4122	9.487	4.891
GUIDE	0.0183	0.2797	7.497	3.331

Acknowledgement

Our code is based on EG3D, GOAE, Deep3DFaceRecon, PerceptualSimilarity, pytorch-fid, InsightFace, and CurricularFace.

BibTex

@inproceedings{seo2024generative,
    author    = {Seo, Juwon and Lee, Sung-Hoon and Lee, Tae-Young and Moon, Seungjun and Park, Gyeong-Moon},
    title     = {Generative Unlearning for Any Identity},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2024},
    pages     = {9151-9161}
}