Adversarially robust segmentation models learn perceptually-aligned gradients

Abstract

The effects of adversarial training on semantic segmentation networks has not been thoroughly explored. While previous work has shown that adversarially-trained image classifiers can be used to perform image synthesis, we have yet to understand how best to leverage an adversarially-trained segmentation network to do the same. Using a simple optimizer, we demonstrate that adversarially-trained semantic segmentation networks can be used to perform image inpainting and generation. Our experiments demonstrate that adversarially-trained segmentation networks are more robust and indeed exhibit perceptually-aligned gradients which help in producing plausible image inpaintings. We seek to place additional weight behind the hypothesis that adversarially robust models exhibit gradients that are more perceptually-aligned with human vision. Through image synthesis, we argue that perceptually-aligned gradients promote a better understanding of a neural network's learned representations and aid in making neural networks more interpretable.