Kuan-Chieh Wang, Yan Fu, Ke Li, Ashish Khisti, Richard Zemel, Alireza Makhzani

• Most commands are in run_scripts.
• We outline a few example commands here.
  • Commands below end with a suffix <mode>. Setting <mode>=0 will run code locally. <mode>=1 was used with SLURM on a computing cluster.
• The environment variable ROOT1 was set to my home directory.

Check out the StyleGAN repo and place it in the same directory hierarchy as the present repo. This is used to make sure you can load and run the pretrained StyleGAN checkpoints.

For CelebA experiments:

• Data --
  • download the "Align&Cropped Images" from the CelebA website into the directory data/img_align_celeba.
  • make sure in data/img_align_celeba, there are 000001.jpg to 202599.jpg.
  • download identity_CelebA.txt and put it in data/celeb_a.
• Pretrained DCGAN -- download and untar this into the folder pretrained/gans/neurips2021-celeba.
• Pretrained StyleGAN -- download and untar this into the folder pretrained/stylegan/neurips2021-celeba.
• Pretrained Target Classifier -- download and untar this into the folder pretrained/classifiers/neurips2021-celeba.
• Evaluation Classifier --
  • check out the InsightFace repo and place it in the same directory hierarchy as the present repo.
  • follow instructions in that repo, and download the ir_se50 model, which is used as the evaluation classifier.

CelebA

• the script below runs VMI attack on the first 100 IDs and saves the results to results/celeba-id<ID>.

run_scripts/neurips2021-celeba-stylegan-flow.sh

• generate and aggregate the attack samples by running the command below. The results will be saved to results/images_pt/stylegan-attack-with-labels-id0-100.pt.

python generate_vmi_attack_samples.py

• evaluate the generated samples by running:

fprefix=results/images_pt/stylegan-attack-with-labels-id0-100

python evaluate_samples.py \
	--name load_samples_pt \
	--samples_pt_prefix $fprefix \
	--eval_what stats \
	--nclass 100

Code contain snippets from:
https://github.com/adjidieng/PresGANs
https://github.com/pytorch/examples/tree/master/mnist
https://github.com/wyharveychen/CloserLookFewShot