This repository, now known as GARD-Armory is only to be used by performers involved in the DARPA GARD research program. The adversarial evaluation capabiites that GARD-Armory provides for the laboratory work in GARD has been reworked into a more flexible, easily imported, readily composible armory-library.

Thus, anyone interested in Armory who is not associated with the GARD project should look to https://github.com/twosixlabs/armory-library for the Armory that remains under active development. One can install the most recent release from that repository with

pip install armory-library

Armory is a testbed for running scalable evaluations of adversarial defenses. Configuration files are used to launch local or cloud instances of the Armory docker containers. Models, datasets, and evaluation scripts can be pulled from external repositories or from the baselines within this project.

Our evaluations are created so that attacks and defenses may be interchanged. To do this we standardize all attacks and defenses as subclasses of their respective implementations in the Adversarial Robustness Toolbox (ART) hosted by the LF AI & Data Foundation (LFAI).

TLDR: Try Armory or follow the instructions below to install locally.

pip install armory-testbed

Upon installing armory, a directory will be created at ~/.armory. This user specific folder is the default directory for downloaded datasets, model weights, and evaluation outputs.

To change these default directories simply run armory configure after installation.

If installing from the git repo in editable mode, ensure that your pip version is 22 .

There are three ways to interact with Armory's container system.

• armory run <path/to/config.json> This will run a configuration file end to end. Stdout and stderror logs will be displayed to the user, and the container will be removed gracefully upon completion. Results from the evaluation can be found in your output directory.

• armory run <path/to/config.json> --interactive This will launch the framework-specific container specified in the configuration file, copy the configuration file into the container, and provide the commands to attach to the container in a separate terminal and run the configuration file end to end while attached to the container. A notable use case for this would be to debug using pdb. Similar to non-interactive mode, results from the evaluation can be found in the output directory. To later close the interactive container simply run CTRL C from the terminal where this command was ran.

• armory launch <armory|pytorch-deepspeech> This will launch a framework specific container, with appropriate mounted volumes, for the user to attach to for debugging purposes. A command to attach to the container will be returned from this call, and it can be ran in a separate terminal. To later close the interactive container simply run CTRL C from the terminal where this command was ran.

• armory launch <armory|pytorch-deepspeech> --jupyter. Similar to the interactive launch, this will spin up a container for a specific framework, but will instead return the web address of a jupyter lab server where debugging can be performed. To close the jupyter server simply run CTRL C from the terminal where this command was ran.

• armory exec <armory|pytorch-deepspeech> -- <cmd> This will run a specific command within a framework specific container. A notable use case for this would be to run test cases using pytest. After completion of the command the container will be removed.

Note: Since Armory launches Docker containers, the python package must be run on system host (i.e. not inside of a docker container).

pip install armory-testbed
armory configure

git clone https://github.com/twosixlabs/armory-example.git
cd armory-example
armory run official_scenario_configs/cifar10_baseline.json

All containers have a pre-installed armory package so that baseline models, datasets, and scenarios can be used.

Additionally, volumes (such as your current working directory) will be mounted from your system host so that you can modify code to be run, and retrieve outputs. For more information on these mounts, please see our Docker documentation

Armory provides several baseline threat-model scenarios for various data modalities. When running an armory configuration file, the robustness of a defense will be evaluated against that given scenario. For more information please see our Scenario Documentation.

Please see the frequently asked questions documentation for more information on:

• Dataset format and preprocessing
• Access to underlying models from wrapped classifiers.

Armory is an open source project and as such we welcome contributions! Please refer to our contribution docs for how to get started.

This material is based upon work supported by the Defense Advanced Research Projects Agency (DARPA) under Contract No. HR001120C0114. 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).