Natural Adversarial Examples Slash Image Classification Accuracy by 90% | Synced

Researchers from UC Berkeley and the Universities of Washington and Chicago have released a set of natural adversarial examples, which they call “ImageNet-A.”