This repository contains trained model weights, training and evaluation code for the paper Increasing the robustness of DNNs against image corruptions by playing the Game of Noise 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 the trained model weights by either running

bash ./Models/download_models.sh

or by downloading them from the Releases on GitHub.

The evaluation results are summarized in the Table below.

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.

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

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

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

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

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

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 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.