Human Perception Robustness Quiz
The human vision system is considered robust to small perturbations, but we found that nearly imperceptible perturbations to images that fool state-of-the-art neural networks also trick humans into choosing a new label.
Humans & NNs see "horse"
Humans & NNs see "dog"
Instructions
1. For each of the 25 images, select the CIFAR-10 class that best describes it. First, label the version of the image with a small adversarial perturbation added, then label the original clean image.
2. The 'Automobile' class includes sedans & SUVs. 'Truck' includes big trucks. Neither includes pickup trucks. The 'Deer' class includes moose.
3. The clean image's label will be highlighted in green after you label the clean version of each image. When you are done with all 25 images, your latest score will appear below and be recorded to update the averages above. To avoid biasing the results, take the quiz just once.
Read more about the project: ICML 2024 Paper