Object Sensing and Cognition for Adversarial Robustness

Overview

This repository contains the source code for Intel Lab's and Georgia Tech's submissions to the DARPA GARD program. We currently distribute our submission that contain our defenses, and we are working to open source our training code too.

Getting Started

To recreate our submission:

make submission

This assumes some dependencies are installed on your system. We support installing these system-level dependencies for Ubuntu 18.04:

make ubuntu_deps

To run our submission, you will need to use ARMORY. We do provide some infrastructure to run our submission using our Makefile. Because ARMORY relies upon Python, we manage our Python dependencies are managed using Poetry. The following command will install Poetry and use it to create a virtual environment and install all necessary dependencies:

make python_deps

If you have not previously installed Poetry, you will want to add it to your path:

source $HOME/.poetry/env

Because Armory primarily relies upon Docker images to run evaluations, we created our own Docker image container our dependencies.

make docker_image

After install all Python dependencies, you need to configure Armory using:

poetry run armory configure

Once configured, you can run our UCF101 defense submission via:

make run_ucf101_submission

Contributing

External contributions are not being accepted and pull requests will not be accepted.

Acknowledgements

This material is based upon work supported by the Defense Advanced Research Projects Agency (DARPA) under Contract No. HR001119S0026.