A simple way to make neural networks robust against diverse image corruptions

This repository contains trained model weights, training and evaluation code for the paper A simple way to make neural networks robust against diverse image corruptions by Evgenia Rusak*, Lukas Schott*, Roland Zimmermann*, Julian Bitterwolf, Oliver Bringmann, Matthias Bethge & Wieland Brendel.

We show that a very simple approach -- data augmentation with Gaussian noise -- suffices to surpass state-of-art methods to increase robustness towards common corruptions. Going one step further, we learn the per-pixel distribution to sample noise from adversarially with a simple generative neural network which we call the Noise Generator. Training the Noise Generator and the classifier jointly further increases robustness.

Download models

Download the trained model weights by either running

bash ./Models/download_models.sh

or by downloading them from the Releases on GitHub.

Results

The evaluation results are summarized in the Table below.

Model	ImageNet val top1 accuracy [%] ↗	ImageNet-C top1 accuracy [%] ↗	ImageNet-C mCE ↘
Vanilla	76.1	39.2	76.7
Speckle	75.8	46.4	68.3
Gauss_sigma_0.5	75.9	49.4	64.5
Gauss_mult	76.1	49.2	65.0
ANT	76.1	51.1	62.5
ANT+SIN	74.9	52.2	61.2
ANT3x3	76.1	50.4	63.4
ANT3x3+SIN	74.1	52.6	60.7

Evaluate models

To validate our models, run python3 main.py. It is necessary to download the model weights first!

Arguments:

• --datadir-clean the top-level directory of the ImageNet dataset (mandatory)
• --imagenetc-path the top-level directory of the ImageNet-C dataset (mandatory)
• --model_name name of the model that should be evaluated (optional, default: clean, possible choices: clean, ANT-SIN, ANT, Gauss_mult, Gauss_sigma_0.5, Speckle)
• --workers number of data loading workers (optional, default: 30)
• --test-batch-size number of images in the batch (optional, default: 256)
• --device device for evaluation (optional, default: 'cuda')

The results are saved as txt files.

Example Usage

Evaluate our model trained with Speckle noise data augmentation

It is necessary to specify the paths for the ImageNet validation dataset and the ImageNet-C dataset. Then you can run:

python3 main.py -e True --model_name Speckle --datadir-clean /path-to-imagenet-val --imagenetc-path /path-to-imagenet-c

Evaluate our best performing model trained with Adversarial Noise Training and on Stylized ImageNet ANT3x3+SIN

python3 main.py -e True --model_name ANT3x3+SIN --ng_type 3x3 --datadir-clean /path-to-imagenet --imagenetc-path /path-to-imagenet-c

Train a robust model with ANT

python3 main.py  --mode ANT --datadir-clean /path-to-imagenet --imagenetc-path /path-to-imagenet-c

Train a robust model with ANT3x3

python3 main.py  --mode ANT --ng_type 3x3 --datadir-clean /path-to-imagenet --imagenetc-path /path-to-imagenet-c

Train a robust model with ANT+SIN

python3 main.py  --mode ANT+SIN --datadir-clean /path-to-imagenet --imagenetc-path /path-to-imagenet-c --datadir-sin /path-to-stylized-imagenet

Train a robust model with Gaussian noise with one specific sigma

python3 main.py  --mode Gauss_single --std_gauss 0.5 --datadir-clean /path-to-imagenet --imagenetc-path /path-to-imagenet-c

Examples of adversarial noise

Examples of adversarial noise found by the noise generator during adversarial noise training are displayed below.

A: Examples of learned probability densities over the grayscale version of the noise during ANT1x1 where each density corresponds to one local minimum; B: Example images with sampled uncorrelated adversarial noise; C: Example patches of locally correlated noise with a size of 28x28 pixels learned during ANT3x3; D: Example images with sampled correlated adversarial noise.