CycleGAN learns to cheat by hiding information in generated images

Packtpub: CycleGAN is an algorithm for performing image-to-image translation where the neural network needs to learn the mapping between an input image and an output image with the help of a training set of aligned image pairs. What sets CycleGAN apart from other GAN algorithms is that it does not require paired training data. It translates images from a source domain X to a target domain Y without needing paired examples.

At NeurIPS 2017, a group of Stanford and Google researchers presented a very intriguing study on how a neural network, CycleGAN learns to cheat. The researchers trained CycleGAN to transform aerial images into street maps, and vice versa. They found that the neural network learned to hide information about the original image inside the generated one in the form of a low-amplitude high-frequency signal, which almost appears to be noise. Using this information, the generator can then reproduce the original image and thus satisfy the cyclic consistency requirement....read more>>>...