This is the official implementation of the CycleGAN robust to self-adversarial attacks used in paper "Adversarial Self-Defense for Cycle-Consistent GANs" by Dina Bashkirova, Ben Usman and Kate Saenko, NeurIPS'2019. Paper

In this repository you can find the extension of the original implementation of CycleGAN in pytorch. This extension contains two defense techniques against the self-adversarial attacks performed by the unsupervised image-to-image translation methods that hold the cycle-consistency property. More information on self-adversarial attacks can be found in our paper.

You can find the instructions on how to train and test the model with additional guess loss or with Gaussian noise here and here respectively.

In order to reproduce our results, you can find all the configuration files in configs directory.