FORGET ME: Federated Unlearning for Face Generation Models

Setup

conda env create -f environment.yaml/environment2.yaml

Dataset

This project utilizes several large-scale public face datasets. You are required to download them manually from their official sources. The core datasets are:VGG-Face2 、 CelebA-HQ 、 CASIA-WebFace After downloading the datasets, you need to perform a custom partitioning to simulate a federated learning environment. The data should be organized into 128 client datasets, with the constraint that each client possesses images of only a single, unique face identity. Before the forgetting operations, all face images must be pre-processed and augmented. The data augmentation should follow the methodology used for processing face images in the Deep3DFaceReconstruction project. This ensures that the input data is normalized and prepared in a consistent manner for the model.

Unlearn Train

pyhton Fed_Unlearn_gan.py
pyhton Fed_Unlearn_dm.py

Acknowledgements

We would like to express our gratitude to the authors of the following projects for their invaluable contributions and for open-sourcing their work:CompVis/stable-diffusion、GOAE 、EG3D、