CIFAR-10-C

Overview

CIFAR-10-C is a corrupted version of the CIFAR-10 dataset designed for benchmarking neural network robustness to common corruptions and perturbations. It consists of the CIFAR-10 test set with 19 different types of corruptions applied at 5 severity levels each.

Test: 950,000 images (19 corruptions × 5 severity levels × 10,000 images)
Train: N/A (test-only dataset for robustness evaluation)

The dataset includes corruptions such as noise, blur, weather effects, and digital artifacts to evaluate how well models trained on clean CIFAR-10 data generalize to corrupted inputs.

Data Structure

When accessing an example using ds[i], you will receive a dictionary with the following keys:

Key	Type	Description
`image`	`PIL.Image.Image`	32×32×3 RGB corrupted image
`label`	int	Class label (0-9)
`corruption_name`	str	Type of corruption applied (e.g., “gaussian_noise”, “fog”)
`corruption_level`	int	Severity level of corruption (1-5, where 5 is most severe)

Corruption Types

The dataset includes 19 corruption types across different categories:

Noise: - gaussian_noise - shot_noise - impulse_noise - speckle_noise

Blur: - defocus_blur - glass_blur - motion_blur - zoom_blur - gaussian_blur

Weather: - snow - frost - fog

Digital: - brightness - contrast - elastic_transform - pixelate - jpeg_compression - saturate - spatter

Usage Example

Basic Usage

from stable_datasets.images.cifar10_c import CIFAR10C

# Load the test set (only split available)
ds = CIFAR10C(split="test")

sample = ds[0]
print(sample.keys())  # {"image", "label", "corruption_name", "corruption_level"}
print(f"Corruption: {sample['corruption_name']} at level {sample['corruption_level']}")

# Optional: make it PyTorch-friendly
ds_torch = ds.with_format("torch")

Typical Workflow

from stable_datasets.images.cifar10 import CIFAR10
from stable_datasets.images.cifar10_c import CIFAR10C

# Train on clean CIFAR-10
train_data = CIFAR10(split="train")
# ... train your model ...

# Evaluate robustness on CIFAR-10-C
test_data = CIFAR10C(split="test")
# ... evaluate on corrupted images ...

References

Paper: Benchmarking Neural Network Robustness to Common Corruptions and Perturbations (ICLR 2019)
Zenodo: https://zenodo.org/records/2535967
License: CC BY 4.0

Citation

@article{hendrycks2019robustness,
  title={Benchmarking Neural Network Robustness to Common Corruptions and Perturbations},
  author={Dan Hendrycks and Thomas Dietterich},
  journal={Proceedings of the International Conference on Learning Representations},
  year={2019}
}