https://arxiv.org/abs/2103.09151

Attacking End-to-End Autonomous Driving Systems

• Adversarial Driving
  • Overview
  • Quick Start
    • 1. Setup the simulator
    • 2. Setup the server
    • 3. Setup the browser
  • How to add your own attack
    • Add the attack on the UI
    • Implement the attack
  • Training the model
  • Resources

Adversarial Driving: The behaviour of end-to-end autonomous driving model can be manipulated by adding unperceivable perturbations to the input image.

This simulator was built for Udacity's Self-Driving Car Nanodegree, and it's available here.

• Windows: please download term1-simulator-windows.zip, and open beta_simulator.exe.

• Linux: please download term1-simulator-linux.zip, and open beta_simulator.x86_64.

• Mac: please download term1-simulator-mac.zip, and install beta_simulator_mac.app.

You may use anaconda or miniconda.

$ cd model
$ conda env create -f environment.yml
$ conda activate adversarial-driving
$ python drive.py model.h5

This is just a website, your can use any web server, just serve all the content under client/web.

If you use windows, click on client/client.exe. It's a single executable that packages everything.

For Linux and Mac, or other Unix, the server can be built with:

go get -u github.com/gobuffalo/packr/packr
go get github.com/gobuffalo/[email protected]
packr build

The web page will be available at: http://localhost:3333/

That's it!

This demo aims to raise concerns over applications of end-to-end models in safety-critical systems

The system is scalable to include more attacks for further research. For example, if you would like to add an attack that increases each pixel by one. (one-pixel-attack)

First, you can add the attack option here inside client/web/index.html by changing the <input id="" value="one_pixel"> and <label for="">:

<div class="form-check">
  <input class="form-check-input" type="radio" name="flexRadioDefault" id="flexRadioOnePixel" value="one_pixel">
  <label class="form-check-label" for="flexRadioOnePixel">
    One-Pixel-Attack
  </label>
</div>

Then you can find a new option on the website :

(Remember to recompile client.exe using packr build if you use the built-in server):

Second and last, you can implement your attack here in model/adversarial_driving.py.

# One-Pixel Attack
# The attack_type == "one_pixel" here is defined in index.html in the last step.
# <input id="" value="one_pixel">
if self.attack_type == "one_pixel":
	# The attack increases each pixel by one
    noise = np.ones((160, 320, 3))
    return noise

Congratulations, now open the simulator and python server, you should see the outcome of the new attack on the website http://localhost:3333/. The one-pixel attack demonstrates how to add a new attack, but it has negligible effect on the performance of the driving model. Hope you can find a strong attack!

You can use the dataset from Udacity to train your own autonomous driving model. Extract the dataset to model/data, and the folder structure should be like this.

├───model
│   ├───data
│   │   ├───IMG
|   |   └───driving_log.csv
│   └───model.py

And then start training:

python model.py

This will generate a file model-<epoch>.h5 whenever the performance in the epoch is better than the previous best. More details are illustrated here.

• Dataset: https://d17h27t6h515a5.cloudfront.net/topher/2016/December/584f6edd_data/data.zip

• Simulator: https://github.com/udacity/self-driving-car-sim

• Project Structure: https://github.com/udacity/CarND-Behavioral-Cloning-P3

• Nvidia End-to-End model: https://developer.nvidia.com/blog/deep-learning-self-driving-cars/

• Behavior Cloning: https://github.com/naokishibuya/car-behavioral-cloning

• Youtube: https://github.com/llSourcell/How_to_simulate_a_self_driving_car

@misc{wu2021adversarial,
      title={Adversarial Driving: Attacking End-to-End Autonomous Driving Systems}, 
      author={Han Wu, Syed Yunas and Wenjie Ruan},
      year={2021},
      eprint={2103.09151},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}