The ImageNet-A dataset consists of real-world, unmodified, naturally occurring examples that are misclassified by ResNet models. The dataset contains 7,500 adversarially filtered images, which can easily cause significant performance degradation of machine learning models.

This dataset is the same as the ImageNet test set, but it is more challenging for existing models. The 200 categories of ImageNet-A cover most of the large categories covered by the ImageNet-1K dataset.

This dataset includes two datasets: ImageNet-A and ImageNet-O.