spring-epfl / trickster

Library and experiments for attacking machine learning in discrete domains

Home Page:https://arxiv.org/abs/1810.10939

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Trickster

trickster

Travis Docs

Library and experiments for attacking machine learning in discrete domains using graph search.

See the documentation on Readthedocs, or jump directly to the guide.

Setup

Library

Install the trickster library as a Python package:

pip install -e git+git://github.com/spring-epfl/trickster#egg=trickster

Note that trickster requires Python 3.6.

Experiments

Python packages

Install the required Python packages:

pip install -r requirements.txt

System packages

On Ubuntu, you need these system packages:

apt install parallel unzip

Datasets

To download the datasets, run this:

make data

The datasets include:

Citing

This is an accompanying code to the paper "Evading classifiers in discrete domains with provable optimality guarantees" by B. Kulynych, J. Hayes, N. Samarin, and C. Troncoso, 2018. Cite as follows:

@article{KulynychHST18,
  author    = {Bogdan Kulynych and
               Jamie Hayes and
               Nikita Samarin and
               Carmela Troncoso},
  title     = {Evading classifiers in discrete domains with provable optimality guarantees},
  journal   = {CoRR},
  volume    = {abs/1810.10939},
  year      = {2018},
  url       = {http://arxiv.org/abs/1810.10939},
  archivePrefix = {arXiv},
  eprint    = {1810.10939},
}

Acknowledgements

This work is funded by the NEXTLEAP project within the European Union’s Horizon 2020 Framework Programme for Research and Innovation (H2020-ICT-2015, ICT-10-2015) under grant agreement 688722.

About

Library and experiments for attacking machine learning in discrete domains

https://arxiv.org/abs/1810.10939

License:MIT License


Languages

Language:Jupyter Notebook 92.2%Language:Python 7.5%Language:Shell 0.3%Language:Makefile 0.0%