Model Inversion

This is the code for the manuscript "A Sample-Level Evaluation and Generative Framework for Model Inversion Attacks". All experiments are conducted on 4 GPUs of NVIDIA GeForce RTX 3090 server with Linus OS, CUDA 12.2, and PyTorch 1.7.1. Our code utilizes DistributedDataParallel from PyTorch to train the model in parallel manner.

Setup

1. Install StyleGAN Docker environment at https://github.com/NVlabs/stylegan2-ada-pytorch.
2. Download relevant datasets used in the manuscript and preprocess these datasets with the scripts in the ./datasets folder. For preparing the CelebA-HQ dataset, you need to follow the instruction in https://github.com/nperraud/download-celebA-HQ.
3. Download pretrained StyleGAN weights from https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada-pytorch/pretrained/. Create a ./pretrained directory and place these weights in this directory.

Run

Run the scripts in the ./scripts folder. You can change the parameters in the scripts for different settings of MI attacks. All experimental results will be saved to a automatically created folder ./.logger.