search

sky4689524 / DefenseGAN-Pytorch

12 stars 4 forks source link

Pre-trained WGAN generator and discriminator #1

Open matteoghirardelli opened 4 years ago

matteoghirardelli commented 4 years ago

Hi, I'm trying to run the code of this repository to implement the same experiments of the original DEFENSE-GAN paper, but with the CIFAR-10 dataset. I was wondering if you have managed to verify the robustness of DEFENSE-GAN on the CIFAR-10 dataset, and if you have any pre-trained models that I could use in order to compare your results with mine.

Also, is this version of the repository the most recent?

sky4689524 commented 4 years ago

Hi, yes I verified the robustness of DEFENSE-GAN on the CIFAR-10 dataset. Unfortunately, I don't have any pre-train models anymore.

yes. It is the recent version.

matteoghirardelli commented 4 years ago

How much accuracy did you obtain when reconstructing the original images using the functions in util_defense_GAN.py? I obtained a very low accuracy (around 0.11) for large values of L and R, even by adapting your code to the simpler assessment with the F-MNIST dataset. How many epochs did you use to train the GAN on the CIFAR-10 dataset?Maybe my GAN wasn't trained enough, but it seems to generate reasonable deepfakes, at least with the F-MNIST dataset

sky4689524 commented 4 years ago

To be honest, I don't remember how much epochs did I train for CIFAR-10 dataset. But I guess you do not need to train a lot of epochs because CIFAR-10 dataset is small size.