Robust Adversarial Examples

I recently wrote a post on the OpenAI blog about how we created images that reliably fool neural network classifiers when viewed from various scales and perspectives, challenging a recent claim that adversarial examples aren’t an issue for self-driving cars because cars see images from multiple scales and angles.

The video above shows that our adversarial kitten photo, printed on a standard color printer, fools the classifier into thinking it’s a “desktop computer” regardless of how it’s zoomed or rotated.

Read the full story about how we did it on the OpenAI blog.