Armory-library is a pure Python library which allows the measurement of ML systems in the face of adversarial attacks. It takes the years of experience gained and techniques discovered under the DARPA GARD program and makes it available to the general ML user.
pip install armory-libraryThis is all that is needed to get a working Armory installation. However, Armory-library
is a library and does not contain any sample code. We provide examples in the
armory-examples repository which is released concurrently with Armory-library.
To install the examples, run:
pip install armory-examplesThe example source code, along with the Armory-library documentation is a good place to learn how to construct your own evaluations using armory-library.
We have provided an sample notebook using Armory to evaluate a food101 classifier in the presence of a Project Gradient Descent (PGD) attack. The notebook can be run for free on Google Colab to get a taste of how Armory works.
The Armory-library documentation is published on Read the Docs or can be viewed directly in the docs directory of this repository.
Armory-library is the successor to the GARD-Armory research program run under DARPA. As that program is nearing its conclusion, that repository will be archived sometime in 2024 and there will be no further development in GARD-Armory by the time you are reading this sentence. The development teams for both GARD-Armory and Armory-library can be reached at armory@twosixtech.com
This material is based upon work supported by the Defense Advanced Research Projects Agency (DARPA) under Contract No. HR001120C0114 and US Army (JATIC) Contract No. W519TC2392035. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the Defense Advanced Research Projects Agency (DARPA) or JATIC.