Reducing Steganography In Cycle-consistency GANs.

Horia Porav, Valentina Musat, Paul Newman

Abstract: In this work we present a simple method of improving the suitability of data generated using cycle-consistency GANs in the context of day-to-night domain adaptation. While CycleGANs produce visually pleasing outputs, they also encode hidden (steganographic) information about the source domain in the generated images, which makes them less suitable as training data generators. We reduce the amount of steganographic information hidden in the generated images by introducing an end-to-end differentiable image de-noiser in between the two generators. The role of the de-noiser is to strip away the high frequency, low amplitude encoded information, making it harder for the generators to hide information that is invisible to the discriminator. We benchmark the suitability of data generated using our simple method in the context of simple domain adaptation for semantic segmentation, comparing with standard Cycle- GAN, MUNIT and DRIT and show that our method yields better performance.