Benchmarking Neural Network Robustness to Common Corruptions and Perturbations

This repository contains the datasets and some code for the paper Benchmarking Neural Network Robustness to Common Corruptions and Perturbations (ICLR 2019) by Dan Hendrycks and Thomas Dietterich.

Requires Python 3+ and PyTorch 0.3+. For evaluation, please download the data from the links below.

ImageNet-C

Download ImageNet-C here. (Mirror.)

Download Tiny ImageNet-C here. (Mirror.)

Tiny ImageNet-C has 200 classes with images of size 64x64, while ImageNet-C has all 1000 classes where each image is the standard size. For even quicker experimentation, there is CIFAR-10-C and CIFAR-100-C. Evaluation using the JPEGs above is strongly prefered to computing the corruptions in memory, so that evaluation is deterministic and consistent.

ImageNet-C Leaderboard

ImageNet-C Robustness with a ResNet-50 Backbone

Method	Reference	mCE	Clean Error
Assemble-ResNet50	Lee et al.	56.5%	17.90%
AugMix	Hendrycks and Mu et al. (ICLR 2020)	65.3%	22.47%
Stylized ImageNet Data Augmentation	Geirhos et al. (ICLR 2019)	69.3%	25.41%
Patch Uniform	Lopes et al.	74.3%	24.5%
ResNet-50 Baseline		76.7%	23.85%

Other backbones can obtain better results. For example, a vanilla ResNeXt-101 has an mCE of 62.2%. Note Lopes et al. have a ResNet-50 backbone with an mCE of 80.6, so their improvement is larger than what is immediately suggested by the table.

Submit a pull request if you beat the state-of-the-art with a ResNet-50 backbone on ImageNet-C.

AlexNet ImageNet-C Error

Use these values to normalize raw corruption error to calculate mCE:

Corruption	Average	Severity 1	Severity 2	Severity 3	Severity 4	Severity 5
Gaussian Noise	0.886428	0.69528	0.82542	0.93554	0.98138	0.99452
Shot Noise	0.894468	0.71224	0.85108	0.93574	0.98182	0.99146
Impulse Noise	0.922640	0.78374	0.89808	0.94870	0.98720	0.99548
Defocus Blur	0.819880	0.65624	0.73202	0.85036	0.91364	0.94714
Glass Blur	0.826268	0.64308	0.75054	0.88806	0.91622	0.93344
Motion Blur	0.785948	0.58430	0.70048	0.82108	0.89750	0.92638
Zoom Blur	0.798360	0.70008	0.76992	0.80784	0.84198	0.87198
Snow	0.866816	0.71726	0.88392	0.86468	0.91870	0.94952
Frost	0.826572	0.61390	0.79734	0.88790	0.89942	0.93430
Fog	0.819324	0.67474	0.76050	0.84378	0.87260	0.94500
Brightness	0.564592	0.45140	0.48502	0.54048	0.62166	0.72440
Contrast	0.853204	0.64548	0.76150	0.88874	0.97760	0.99270
Elastic	0.646056	0.52596	0.70116	0.55686	0.64076	0.80554
Pixelate	0.717840	0.52218	0.54620	0.73728	0.87092	0.91262
JPEG Compression	0.606500	0.51002	0.54718	0.57294	0.65458	0.74778
Speckle Noise	0.845388	0.66192	0.74440	0.90246	0.94548	0.97268
Gaussian Blur	0.787108	0.54732	0.70444	0.82574	0.89864	0.95940
Spatter	0.717512	0.47196	0.62194	0.75052	0.84132	0.90182
Saturate	0.658248	0.59342	0.65514	0.51174	0.70834	0.82260

ImageNet-P

_{^{ImageNet-P sequences are MP4s not GIFs. The spatter perturbation sequence is a validation sequence.}}

Download Tiny ImageNet-P here. (Mirror.)

Download ImageNet-P here. (Mirror.)

ImageNet-P Leaderboard

ImageNet-P Perturbation Robustness with a ResNet-50 Backbone

Method	Reference	mFR	mT5D
AugMix	Hendrycks and Mu et al. (ICLR 2020)	37.4%
Low Pass Filter Pooling (bin-5)	Zhang (ICML 2019)	51.2%	71.9%
ResNet-50 Baseline		58.0%	78.4%

Submit a pull request if you beat the state-of-the-art on ImageNet-P.

Citation

If you find this useful in your research, please consider citing:

@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}
}

Part of the code was contributed by Tom Brown.

Icons-50 (From an Older Draft)

Download Icons-50 here or here.