SuperPerturbator

You don't have to wait for later
To fool a classificator
Ask your local hacker trader
For the SuperPerturbator ¹

This repo contains our work on the InformatiCup, solving the challenge of adversarial example creation in a black-box scenario.

Deprecation Notice

This repository is not maintained and many dependencies are deprecated now (esp. tensorflow). The code requires a fundamental refactoring to adapt to these changes.

Setup

The software is distributed as a git repository and a ready-to-use docker container. To get started, perform the commands corresponding to your use case.

Fetching the Sources

To be able to fetch from our repository, you need to provide us your public SSH key to grant you access.

# clone the git repo
git clone git@gitlab.com:LPirch/informaticup.git
cd informaticup

Alternatively, you can import the repo from a tar ball:

# import from tar
cd /home/user/location/of/targz-file
tar -xzf superperturbator-repo.tar.gz

Building from Source

# change to the local repository directory
cd /home/user/workspace/informaticup
# build the docker image
sudo docker build -t superperturbator .

Import a pre-built Docker Image

To import a docker image which has been distributed as tar archive, execute the following command:

sudo docker load --input superperturbator-image.tar

Prepare a Data Directory (optional)

We recommend saving the dataset and internal program state in a separate directory on the host. The advantage of this is that restarting the container won't require download all the data again and also, the program state is not lost on container termination.
This decouples the program state from the container lifecycle.

# make data dir (adjust the path as you like)
mkdir /home/user/data

Please make sure that the data directory is not located in the directory of the repository if it has been used to create the docker image. Due to internal restrictions on docker contexts, this directory will be ignored otherwise.

Run the Container

To start the container, you must have built or imported the docker image. Check the output of

sudo docker images

whether it contains the superperturbator image.

Next, you can start it by typing

sudo docker run -it --rm \
	-p 80:80/tcp  \
	-v /home/user/workspace/informaticup/data:/informaticup/data \
	superperturbator

The -p flag binds a TCP port to the host such that it can be reached in the user's web browser.
The -v flag is optional and mounts the data directory into the container.
Optionally, you can replace the -it by -d which runs the docker container as a daemon.

Run a nvidia cuda Container

Our software supports the usage of nvidia cude which greatly speeds up computational tasks like training or attacking a model.
However, this technique is not officially supported since it is heavily dependent on the machine executing it.
If you want to use this feature nevertheless, use the following commands to use a cuda docker image.

sudo docker build -t superperturbator-cuda -f cuda_dockerfile
sudo docker run -it --rm \
	--runtime=nvidia \
	-p 80:80/tcp  \
	-v /home/user/workspace/informaticup/data:/informaticup/data \
	superperturbator-cuda

Opening the Program

After starting the container, check the command line output and wait until it says:

Starting development server at http://0.0.0.0:80/
Quit the server with CONTROL-C.

Now open your web browser on "localhost" or "127.0.0.1" and you should see the welcome page. We suggest using the chrome/chromium browser since the user experience has been optimized for it.

Important Notes

the first startup will take some time since we need to fetch the training and test datasets
always provide fully qualified paths; it is not guaranteed to work otherwise
for cuda-docker, you need to be able to execute nvidia-docker. if you want to do that, please read the instructions according to your distro on the official page
cuda-docker is not officially supported; use it at own risk
the files and directories created in the data directory will be owned by the user executing the docker daemon (usually root); do not modify the file permissions because docker won't be able to use them otherwise

1: Adapted from the superperforator advertisement in the movie "The shoe of Manitou" ("Der Schuh des Manitu")

LPirch/informaticup2019