m-kahla / label-only-model-inversion-attacks-via-boundary-repulsion Goto Github PK

Hi,

Thank you for sharing your amazing work.

I am currently working on your code and curious about the details on the evaluation protocol for BREP-MI, which might not be found in the paper.

May I trouble you about these info:

1. What is the attacked IDs? It is 0..299 or 300 random IDs from 1000 private IDs. The default in your code is 300 random IDs?
2. How many reconstructed Images per ID? The default in your code is 1 per ID?

Thank you

We want to use your attack as a baseline, and also need GAN model for FFHQ -> CelebA, but can't find in your Google drive.
In addition, do you use GAN structure as same as KEDMI? (inversion specified GAN)
How about using other GAN structure pretrained on FFHQ dataset (such as StyleGAN 2)

Hi,
Could you please share the pretrained models and the IDs list (public and private images) for Facescrub and Pubfig83 experiments?
Thank you very much.
Ngoc

I'm sorry, I have been wondering why we should use another model to test the accuracy of the target model. Isn't the distribution different for different models? Is such an accuracy rate instructive?

I use the command from readme.
Why can't the reconstructed images correspond to the images in the CelebA dataset?
Maybe I'm using a wrong dataset?
What is the reconstructed original image dataset?
Can you answer it for me.