Bluff is an interactive system for deciphering adversarial attacks on Deep Neural Networks (DNNs).

• Video: https://youtu.be/b6PHyYnassc
• Live demo: https://poloclub.github.io/bluff/
• Paper: https://arxiv.org/pdf/2009.02608.pdf
• Code and Data: https://github.com/poloclub/bluff

Bluff: Interactively Deciphering Adversarial Attacks on Deep Neural Networks.
Nilaksh Das*, Haekyu Park*, Zijie J. Wang, Fred Hohman, Robert Firstman, Emily Rogers, Duen Horng (Polo) Chau.
IEEE Visualization Conference (VIS), 2020.
* Both contributed equally

For a live demo, visit: https://poloclub.github.io/bluff/

Code (both frontend and backend) and data are available at https://github.com/poloclub/bluff.

For the presentation video, visit: https://youtu.be/b6PHyYnassc

• Download or clone this repository:

  git clone https://github.com/poloclub/bluff.git

• Within bluff repo, run:

  python -m http.server <PORT>
  

  For example,

  python -m http.server 8080

  To run this command, python 3 is needed.

• Open any web browser and go to http://localhost:<PORT>. For example, http://localhost:8080 if you used port 8080.

• You can find the frontend code in bluff/code/frontend.

This project is written in python 3.7.6. You can install the dependencies by running:

pip install -r requirements.txt

• We used tfrecord ImageNet Datasete.

The backend code is in bluff/code/frontend.

• Generate Adversarial Images
  • Please run gen_adversarial_images.py.
• Compute Neuron Importance
  • Please run preprocess_scores.py
• Compute Propagation of Influence
  • Please run edges.py

@inproceedings{das2020bluff,
  title={Bluff: Interactively Deciphering Adversarial Attacks on Deep Neural Networks},
  author={Wang, Zijie J. and Turko, Robert and Shaikh, Omar and Park, Haekyu and Das, Nilaksh and Hohman, Fred and Kahng, Minsuk and Chau, Duen Horng (Polo)},
  booktitle={IEEE Visualization Conference (VIS)},
  publisher={IEEE},
  year={2020},
  url={https://github.com/poloclub/bluff}
}

MIT License. See LICENSE.md.

For questions contact Nilaksh Das or Haekyu Park.