Watch out! Motion is Blurring the Vision of Your Deep Neural Networks

This repository is the official implementation of Watch out! Motion is Blurring the Vision of Your Deep Neural Networks.

Requirements

We use python 3.6 and pytorch 0.4.1 on Ubuntu 16.04. with NVIDIA Tesla P40 and M40.

To install requirements:

pip install -r requirements.txt

and the PoolNet for saliency object detection is required: https://github.com/backseason/PoolNet.git Please put the poolnet code to the './tools/saliency' folder.

The foolbox2.3.0 is also required. After installing, please replace the installed '/foolbox2.3.0/attacks/iterative_projected_gradient.py' with our new version './foolbox2.3.0/attacks/iterative_projected_gradient.py'.

The DeblurGAN and DeblurGANv2 are also required for the discussion of effects to Deblurring methods: https://github.com/KupynOrest/DeblurGAN.git https://github.com/VITA-Group/DeblurGANv2.git

An implementation of the spatial transformer is also needed: https://github.com/kamenbliznashki/spatial_transformer

Subject Models

In the paper, we attack four models, i.e., Inception v3 (Inc-v3), Inception v4 (Inc-v4), Inception ResNet v2 (IncRes-v2), and Xception. We use the models pre-trained on imagenet and provided by the pytorchcv: https://github.com/osmr/imgclsmob.git

Dataset

We evaluate our method on NeurIPS 2017 adversarial competition dataset : https://www.kaggle.com/c/nips-2017-non-targeted-adversarial-attack

Evaluation

Please test and evaluate the method via the script in './experiments/'

Results

We present the partial evaluation results in the following and more results could be found in the supplementary material.

Bibtex

@inproceedings{Guo2020NeurIPS,
  title={Watch out! Motion is Blurring the Vision of Your Deep Neural Networks},
  author={Qing Guo and Felix Juefei-Xu and Xiaofei Xie and Lei Ma and Jian Wang and Bing Yu and Wei Feng and Yang Liu},
  booktitle={Advances in Neural Information Processing Systems 34},
  year={2020},
}