Submission to Kaggle NIPS'17 competition on adversarial examples (non-targeted adversarial attack track)
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
datasets Add code Oct 6, 2017
fcn_models Add code Oct 6, 2017
multi_fcn_training
.gitignore Initial commit Oct 6, 2017
LICENSE Add license Oct 6, 2017
README.md Update README.md Oct 24, 2018
attack_non_targeted_multi_fcn.py Add code Oct 6, 2017
examples.png
metadata.json
run_attack.sh Pass None to resnet50_file to skip loading ResNet50 Layers Oct 6, 2017

README.md

Kaggle NIPS'17 Competition

This repository contains the submission of team 'iwiwi' for the non-targeted adversarial attack track of Kaggle NIPS'17 competition on adversarial examples (https://www.kaggle.com/c/nips-2017-non-targeted-adversarial-attack).

Overview

Our approach is to produce adversarial examples by using fully-convolutional neural networks. The basic framework is the same as that of the Adversarial Transformation Networks paper (https://arxiv.org/pdf/1703.09387.pdf), but we used a much larger FCN model and stronger computation power, together with several new ideas such as multi-target training, multi-task training, and gradient hints. For details, we are preparing a technical report that describes our approach.

How to Run

docker pull iwiwi/nips17-adversarial
nvidia-docker run \
  -v ${INPUT_IMAGES}:/input_images \
  -v ${OUTPUT_IMAGES}:/output_images \
  -v ${SUBMISSION_DIRECTORY}:/code \
  -w /code \
  iwiwi/nips17-adversarial \
  ./run_attack.sh \
  /input_images \
  /output_images \
  ${MAX_PERTURBATION}

Examples

The following is the examples of our attack with MAX_PERTURBATION=16 (left: original image, middle: perturbated image, right: perturbation).

Example images

References

  • Shumeet Baluja, Ian Fischer. Adversarial Transformation Networks: Learning to Generate Adversarial Examples. CoRR, abs/1703.09387, 2017.
  • Alexey Kurakin, Ian J. Goodfellow, Samy Bengio, Yinpeng Dong, Fangzhou Liao, Ming Liang, Tianyu Pang, Jun Zhu, Xiaolin Hu, Cihang Xie, Jianyu Wang, Zhishuai Zhang, Zhou Ren, Alan L. Yuille, Sangxia Huang, Yao Zhao, Yuzhe Zhao, Zhonglin Han, Junjiajia Long, Yerkebulan Berdibekov, Takuya Akiba, Seiya Tokui, Motoki Abe. Adversarial Attacks and Defences Competition. CoRR, abs/1804.00097, 2018.

License

MIT License