Improving Robustness Without Sacrificing Accuracy with Patch Gaussian\n Augmentation

Abstract

Deploying machine learning systems in the real world requires both high\naccuracy on clean data and robustness to naturally occurring corruptions. While\narchitectural advances have led to improved accuracy, building robust models\nremains challenging. Prior work has argued that there is an inherent trade-off\nbetween robustness and accuracy, which is exemplified by standard data augment\ntechniques such as Cutout, which improves clean accuracy but not robustness,\nand additive Gaussian noise, which improves robustness but hurts accuracy. To\novercome this trade-off, we introduce Patch Gaussian, a simple augmentation\nscheme that adds noise to randomly selected patches in an input image. Models\ntrained with Patch Gaussian achieve state of the art on the CIFAR-10 and\nImageNetCommon Corruptions benchmarks while also improving accuracy on clean\ndata. We find that this augmentation leads to reduced sensitivity to high\nfrequency noise(similar to Gaussian) while retaining the ability to take\nadvantage of relevant high frequency information in the image (similar to\nCutout). Finally, we show that Patch Gaussian can be used in conjunction with\nother regularization methods and data augmentation policies such as\nAutoAugment, and improves performance on the COCO object detection benchmark.\n