Adversarial Attacks on Deep Neural Networks for Time Series Classification

This is the companion repository for our paper also available on ArXiv titled "Adversarial Attacks on Deep Neural Networks for Time Series Classification". This paper has been accepted at the IEEE International Joint Conference on Neural Networks (IJCNN) 2019.

Approach

Data

The data used in this project comes from the UCR archive, which contains the 85 univariate time series datasets.

Pre-trained models

You can download the pre-trained ResNet models for each dataset in the archive here. These are the models used to generate the adversarial time series examples. They are published for reproducibility, nevertheless the code can be applied to any model in the h5py format.

Prerequisites

All python packages needed are listed in pip-requirements.txt file and can be installed simply using the pip command.

• numpy
• pandas
• sklearn
• scipy
• matplotlib
• tensorflow-gpu
• keras
• h5py

Note that for convience we copied some of the code from the Cleverhans API and modified it to adapt it for time series data instead of images.

Code

To perform the fgsm (or bim) attack on the datasets in the UCR archive you can run the following command:

python3 main.py attack fgsm 

Once the perturbed time series are generated, you can launch this command to plot and visualize the difference:

python3 main.py draw fgsm 

If you want to visualize the noise as well you can run this command:

python3 main.py noise fgsm 

Finally to plot the Multi-Dimensional Scaling (MDS) and visualize the difference between an original and perturbed dataset, you can issue this command:

python3 main.py mds fgsm 

Results

The following animation shows how the accuracy and the time series shape variation with respect to the amount of perturbation allowed.

The folllowing table shows the accuracy over the 85 datasets with and without adversarial perturbation, using both attacks FGSM and BIM for two models ResNet (white-box mode) and FCN (black-box mode). The raw csv results can be found here. For example column 'resnet_ori' shows the original accuracy of ResNet over the 85 datasets, while column 'resnet_fgsm_adv' shows the accuracy after performing the FGSM attack.

Reference

If you re-use this work, please cite:

@InProceedings{IsmailFawaz2019adversarial,
  Title                    = {Adversarial Attacks on Deep Neural Networks for Time Series Classification},
  Author                   = {Ismail Fawaz, Hassan and Forestier, Germain and Weber, Jonathan and Idoumghar, Lhassane and Muller, Pierre-Alain},
  booktitle                = {IEEE International Joint Conference on Neural Networks},
  Year                     = {2019}
}

Acknowledgement

We would like to thank NVIDIA Corporation for the Quadro P6000 grant and the Mésocentre of Strasbourg for providing access to the cluster.