# CIFAR-100-C

CIFAR-100-C is a benchmark dataset of corrupted versions of CIFAR-100 images, used to evaluate the robustness of machine learning models against common image corruptions.

CIFAR-100-C is a benchmark dataset designed to evaluate the robustness of machine learning models against common image corruptions. It consists of corrupted versions of the CIFAR-100 test set, created by applying 15 types of distortions at five severity levels. The dataset is widely used in robustness research to measure how well models generalize to unexpected input perturbations, complementing standard accuracy metrics.

The original CIFAR-100 dataset contains 60,000 32x32 color images in 100 classes, with 50,000 training images and 10,000 test images. CIFAR-100-C takes the 10,000 test images and applies each corruption type at each severity, resulting in 75,000 corrupted images (15 corruptions x 5 severities x 10,000 images). The corruptions simulate real-world distortions such as noise, blur, weather effects, and digital artifacts.

## Corruption Types

The 15 corruption types are grouped into four categories: noise (Gaussian, shot, impulse), blur (defocus, frosted, glass, motion, zoom), weather (snow, frost, fog, brightness), and digital (contrast, elastic, pixelate, JPEG). Each corruption is applied algorithmically, with severity levels ranging from 1 (mild) to 5 (severe). For example, Gaussian noise adds random pixel values, while motion blur simulates camera movement. The dataset is generated using a standardized codebase, ensuring reproducibility across studies.

## Usage in Robustness Evaluation

CIFAR-100-C is commonly used to measure the robustness of [neural networks](https://www.wikiprompt.org/wiki/neural-network) and other [machine learning](https://www.wikiprompt.org/wiki/machine-learning) models. Researchers report the average error rate across all corruptions and severities, often called the mean Corruption Error (mCE). Models trained with [data augmentation](https://www.wikiprompt.org/wiki/data-augmentation) techniques, such as random cropping or color jitter, typically perform better on CIFAR-100-C than models trained without. The benchmark has become a standard tool in the robustness community, alongside CIFAR-10-C and ImageNet-C.

## Relationship to Other Benchmarks

CIFAR-100-C is part of a family of corruption benchmarks introduced by Hendrycks and Dietterich in 2019, which also includes CIFAR-10-C and ImageNet-C. These datasets share the same corruption types and severity levels, allowing cross-dataset comparisons. The benchmark is often used in conjunction with adversarial robustness evaluations, though it focuses on natural corruptions rather than malicious perturbations. It is also used to study the effects of [batch normalization](https://www.wikiprompt.org/wiki/batch-normalization) and other architectural choices on robustness.

## Limitations and Extensions

While CIFAR-100-C is widely used, it has limitations. The corruptions are synthetic and may not fully capture real-world distribution shifts. Some researchers have proposed extensions, such as CIFAR-100-P, which includes temporally correlated corruptions for video-like sequences. Others have used CIFAR-100-C to evaluate the robustness of [large language models](https://www.wikiprompt.org/wiki/large-language-model) when applied to vision tasks, though such applications are less common. The dataset remains a valuable resource for benchmarking robustness, especially for small-scale models and quick experiments.

## See Also

- [data-augmentation](https://www.wikiprompt.org/wiki/data-augmentation)
- [residual-network](https://www.wikiprompt.org/wiki/residual-network)
- [deep-learning](https://www.wikiprompt.org/wiki/deep-learning)
- [artificial-intelligence](https://www.wikiprompt.org/wiki/artificial-intelligence)

---
Source: https://www.wikiprompt.org/wiki/cifar-100-c
License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/)
Last updated: 2026-09-12T16:27:25.518241+00:00