Adversarial Robustness Test Bed
Project description
ARMORY is a test bed for running scalable evaluations of adversarial defenses. Configuration files are used to launch local or cloud instances of the ARMORY docker container. Models, datasets, and evaluation scripts can be pulled from external repositories or from the baselines within this project.
Installation
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 and evaluation
outputs. Defaults can be changed by editing ~/.armory/config.json
Usage
ARMORY works by running an evaluation configuration file within the armory docker
ecosystem. To do this, simply run armory run <path_to_evaluation.json>
.
The current working directory and armory installation directory will be mounted
inside the container and the armory.eval.Evaluator
class will proceed to run the
evaluation script that is written in the evaluation['eval_file']
field of the
config.
For more detailed information on the evaluation config file please see the documentation in our example repo.
Note: Since ARMORY launches Docker containers, the python package must be ran on system host.
As an example:
pip install armory-testbed
armory configure
git clone https://github.com/twosixlabs/armory-example.git
cd armory-example
armory run examples/fgm_attack.json
Interactive Debugging of Evaluations
Debugging evaluations can be performed interactively by passing --interactive
and
following the instructions to attach to the container in order to use pdb or other
interactive tools. There is also support for --jupyter
which will open a port on
the container and allow notebooks to be ran inside the armory environment.
Custom Attacks and Defenses
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 adversarial-robustness-toolbox
Docker
Armory is intended to be a lightweight python package which standardizes all evaluations inside a docker container. Docker images will be pulled as needed when evaluations are ran.
However if there are issues downloading the images (e.g. proxy) they can be built within this repo:
version=$(python -c "import armory; print(armory.__version__)")
docker build --build-arg armory_version=${version} --target armory-tf1 -t twosixarmory/tf1:${version} .
docker build --build-arg armory_version=${version} --target armory-tf2 -t twosixarmory/tf2:${version} .
docker build --build-arg armory_version=${version} --target armory-pytorch -t twosixarmory/pytorch:${version} .
Docker Mounts
By default when launching an ARMORY instance the current working directory will be mounted as your default directory.This enables users to run modules from ARMORY baselines, as well as modules from the user project.
Docker Setup
Depending on the task, docker memory for an ARMORY container must be at least 8 GB to run properly (preferably 16+ GB). On Mac and Windows Desktop versions, this defaults to 2 GB. See the docs to change this:
Docker Cleanup
Running armory download-all-data
will download new Docker images, but will not clean up old images.
To download new images and clean up old images:
armory clean
If containers are currently running that use the old images, this will fail. In that case, either stop them with first or run:
armory clean --force
To display the set of current images:
docker images
To manually delete images, see the docs for docker rmi.
In order to see the set of containers that are running:
docker ps
ARMORY will attempt to gracefully shut down all containers it launches; however, certain errors may prevent shutdown and leave running containers. To shut down these containers, please see the docs for docker stop and docker kill.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distributions
Built Distribution
Hashes for armory_testbed-0.5.0-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 91daca7db305b0f2edcc1dba2c027e097a2e324109ee866d8a364efcc05ea3be |
|
MD5 | 2971aead19642448388c74ce7feb1fbc |
|
BLAKE2b-256 | 03baf0ea6b7e9d278dd7273dfdb90bccc4dcf7acb2addffb10aa7a1d9d9d55a9 |