This repository contains data and code corresponding to

D. Stutz, M.Hein, B.Schiele. Disentangling Adversarial Robustness and Generalization. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019.

Please cite as:

@inproceedings{Stutz2019CVPR,
  title = {Disentangling Adversarial Robustness and Generalization},
  author = {Stutz, David and Hein, Matthias and Schiele, Bernt},
  booktitle = {IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  publisher = {IEEE Computer Society},
  year = {2019}
}

Also check the project page for the final publication, code and data.

This repository includes PyTorch implementations of the PGD attack [1], the C+W attack [2], adversarial training [1] as well as adversarial training variants for adversarial deformations and on-manifold adversarial examples.

Additionally, the repository allows to easily reproduce the experiments as reported in the paper.

The main goal of the provided code is to reproduce the results reported in the paper.

The code includes several modules, most of which are tailored to the reported experiments; some, like the implemented attacks, can also be used standalone:

• attacks: implemented adversarial attacks, including [1] and [2]

• common: various utilities, including:

  • latex.py, vis.py, plot.py: utilities for plotting and visualization;
  • cuda.py, torch.py, numpy.py: PyTorch and NumPy utilities;
  • ppca.py: an implementation of probabilistic PCA [3];
  • state.py, scheduler.py, timer.py: utilities for training;

• data/fonts: generation code for the synthetic FONTS dataset;

• experiments: implemented experiments;

• training: command line tools for training, testing, attacking models.

  [1] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. ICLR, 2018. [2] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In SP, 2017. [3] Michael E. Tipping, Chris M. Bishop. Probabilistic Principal Component Analysis. Journal of the Royal Statistical Society, 1999.

The experiments module will be the main focus of the following guide.

Before running experiments, the environment needs to be set up, the datasets need to be downloaded, and the pre-trained VAE-GAN models need to be downloaded

Installation. The experiments were run with on a custom Debian 9 with Tesla V100 GPUs and

• Python 3.5.3
• CUDA 9
• torch 0.4.0
• torchvision 0.2.1
• h5py 2.9.0
• numpy 1.16.3
• scipy 1.2.1
• scikit-learn 0.19.1
• scikit-image 0.14.2
• umap-learn 0.2.3
• xgboost 0.72
• imageio 2.5.0

Data. The data in the correct formats can be downloaded in the following table:

Each datasets provides HDF5 files for training/test images and labels; FONTS additionally provides test/training latent codes.

For FONTS, as example, these HDF5 files look as follows; images are stored as floats in [0,1]:

• database.h5: 1000x10x28x28 (only for FONTS, prototype image for each of the 1000 fonts and 10 classes)
• codes.h5: 1120000x3 (only for FONTS, classes not splitted into train/test)
• theta.h5: 1120000x6 (only for FONTS, latent codes not splitted into train/test)
• images.h5: 1120000x28x28x1 (only for FONTS, images not splitted into train/test)
• train_codes.h5: 960000x3 (for FONTS, the class label is stored in the third column)
• train_theta.h5: 960000x6 (only for FONTS, latent codes)
• train_images.h5: 960000x28x28x1
• test_codes.h5: 160000x3 (for FONTS, the class label is stored in the third column)
• test_theta.h5: 160000x6 (only for FONTS, latent codes)
• test_images.h5: 160000x28x28x1

CelebA needs to be converted manually as the license does not allow redistribution. For this, data/celeba includes two simple scripts, convert_images.py and convert_labels.py; for these scripts, the dataset should be downloaded into BASE_DATA such that the following directories/files exist:

BASE_DATA/CelebA/img_align_celeba
Base_DATA/CelebA/Anno/list_attr_celeba.txt

VAE-GAN Models. To approximate the manifolds of these datasets, we use VAE-GANs; The models for all datasets can be downloaded here:

For each dataset, a class-agnostic VAE-GAN in the form of encoder, decoder and classifier is provided and several class-specifiv VAE-GANs.

Setting up Paths. To automate the experiments, two main directories are required: a data directory and an experiment directory. Both can be set in common/paths.py by setting BASE_DATA and BASE_EXPERIMENTS.

BASE_DATA should include the directories for the datasets, for example fonts, emnist, fashion and/or CelebA (mind capitalization!). If the directories are named differently, the corresponding paths in paths.py need to be changed.

In the end, run python3 setup.py to check the above steps.

The experiments on all datasets can be found in experiments/. The base experiment is verify_hypotheses.py which can be used to reproduce our results for all four claims:

1. adversarial examples leave the manifold;
2. on-manifold adversarial examples exist;
3. on-manifold robustness is essentially generalization;
4. and regular robustness is not necessarily contradicting generalization.

For the fourth claim, verify_long_resnet_hypotheses.py is useful, as well.

In the paper, we trained 5 networks, using different strategies, on N training examples with N between 100 and 40000. This can be done using

# Starting with the experiment on FONTS:
cd experiments/fonts
python3 verify_hypotheses.py -training_sizes=250,500,1000,1500,2000,2500,5000,7500,10000,15000,20000,30000,40000 -max_models=5

For smaller-scale experiments, less models can be trained, or the scale of N can be reduced. For larger scale experiments, individual Ns can be run on different GPUs. Additionally, the model to start with can be controlled using -start_model. This means -start_model=4 -max_models=5 will only run the fifth model for each N.

The experiment can be evaluated using -mode=evaluate. This will create a directory 0_evaluation, for example BASE_EXPERIMENTS/VerifyHypotheses/Fonts/0_evaluation which contains the plots as reported in the paper. The plots are provided as images and as LaTeX source. Note that for the paper, the LaTeX source was adapted for better readability. The script creates plots for error rate and success rates (off-manifold, on-manifold - true and learned manifold). On Fonts, the on-manifold success rate might look as follows:

The following experiments are implemented for most datasets:

• verify_hypotheses.py: main experiment as reported in the paper, using L_inf adversarial examples and a simple LeNet-like architecture consisting of 3 to 4 convolutional layers with stride two followed by ReLU activations and batch normalization;
• verify_l2_hypotheses.py: main experiment as above but using L_2 adversarial examples;
• verify_[mlp|resnet|vgg2]_hypotheses.py: main experiment using a multi-layer perceptron architecture, ResNet architecture or VGG architecture (without dropout);

Licenses for source code and data corresponding to:

D. Stutz, M.Hein, B.Schiele. Disentangling Adversarial Robustness and Generalization. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019.

Note that the source code and/or data is based on other projects for which separate licenses apply:

• MNIST and EMNIST;
• Fashion-MNIST;
• Google Fonts;

Copyright (c) 2019 David Stutz, Max-Planck-Gesellschaft

Please read carefully the following terms and conditions and any accompanying documentation before you download and/or use this software and associated documentation files (the "Software").

The authors hereby grant you a non-exclusive, non-transferable, free of charge right to copy, modify, merge, publish, distribute, and sublicense the Software for the sole purpose of performing non-commercial scientific research, non-commercial education, or non-commercial artistic projects.

Any other use, in particular any use for commercial purposes, is prohibited. This includes, without limitation, incorporation in a commercial product, use in a commercial service, or production of other artefacts for commercial purposes.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

You understand and agree that the authors are under no obligation to provide either maintenance services, update services, notices of latent defects, or corrections of defects with regard to the Software. The authors nevertheless reserve the right to update, modify, or discontinue the Software at any time.

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. You agree to cite the corresponding papers (see above) in documents and papers that report on research using the Software.

Copyright (c) 2019 David Stutz, Max-Planck-Gesellschaft

Please read carefully the following terms and conditions and any accompanying documentation before you download and/or use the data (the "Dataset").

The authors grant you a non-exclusive, non-transferable, free of charge right: To download the Dataset and use it on computers owned, leased or otherwise controlled by you and/or your organisation; To use the Dataset for the sole purpose of performing non-commercial scientific research, non-commercial education, or non-commercial artistic projects.

Any other use, in particular any use for commercial purposes, is prohibited. This includes, without limitation, incorporation in a commercial product, use in a commercial service, or production of other artefacts for commercial purposes.

Without prior written approval from the authors, the Dataset, in whole or in part, shall not be further distributed, published, copied, or disseminated in any way or form whatsoever, whether for profit or not. This includes further distributing, copying or disseminating to a different facility or organizational unit in the requesting university, organization, or company.

THE DATASET IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE DATASET OR THE USE OR OTHER DEALINGS IN THE DATASET.

You understand and agree that the authors are under no obligation to provide either maintenance services, update services, notices of latent defects, or corrections of defects with regard to the Dataset. The authors nevertheless reserve the right to update, modify, or discontinue the Dataset at any time.

You agree to cite the corresponding papers (see above) in documents and papers that report on research using the Dataset.