Adversarial Environment

Overview

Attacking car racing in OpenAI Gym using adversarial attacks on environment. The fully trained agent and its associated environment wrappers, networks are taken from pytorch_car_caring.

Installation

Clone the repo and cd into directory.

$ git clone https://github.com/DesignInformaticsLab/adversarial-environment.git

$ cd adversarial-environment

Requirements

Here are some of the dependencies that are required. For complete dependencies check requirements.txt file

• pytorch 1.4.0
• gym 0.15.4
• Box2D 2.3.2
• visdom 0.1.8
• numpy 1.18.1
• matplotlib 3.1.3
• torchvision 0.5.0
• tensorboard 1.14.0

Note: If you are facing errors related to Box2D while running, try installing Box2D-kengz v2.3.3. Also, some versions of libraries are upgraded to support tensorboard.

Running

General attack (Level 0)

To train the attack, run python adv_attack_train.py --attack_type=general. To test the attack and render the environment, run python adv_attack_test.py --render --attack_type=general

Patch attack (Level 1)

To train the attack, run python adv_attack_train.py --attack_type=patch --patch_type=circle. To test the attack and render the environment, run python adv_attack_test.py --render --attack_type=patch --patch_type=circle. Currently supports only two types of patches box and circle.

Disclaimer

This work is highly based on the following repo:

1. xtma/pytorch_car_caring