README.md

Readme -- Image-Scaling Attacks and Defenses

About

This document describes the steps to set up the python environment.

System Requirements

• Python 3.6

• Conda / (Virtualenv)

• We've tested everything on different UNIX systems. If you set up a virtual machine, we recommend using the latest Ubuntu OS.

Set up environment

In this section, we will describe everything to set up a Python project with all the necessary requirements/dependencies. If you are already familiar with creating and using Python projects, this section will not be so interesting. That said, let's examine the different steps that are necessary:

1. Install Anaconda

• This step will create an own Python environment so that we do not mess up with the system.
• Go to the Anaconda website and download/install the latest Python 3 version. At the time of writing, it was the Python 3.7 version and Anaconda 2019.10.
• If you want to use virtualenv, because you are already familiar with that, you can also try it. With virtualenv, we had some problems to get the requested version from the requirements.txt. But in principle, virtualenv should also work if you can get the correct python packages.

2. Setup Anaconda

• Open a terminal.
• Create a new environment, we need Python 3.6
• Use: conda create --name scaling-attack python=3.6
  • The name is arbitrary, but makes sense for our project.
• Activate the environment via source activate scaling-attack (if this does not work, try conda activate scaling-attack)

3. Project Requirements

• Go to the project directory scaleatt. You will find a requirements.txt there.
• Use pip install -r requirements.txt to install all requirements.
• If you get the following error: libSM.so.6 is missing.
  • Use opencv-python-headless package instead of opencv-python, without libSM6 dependency.
  • Either replace opencv-python in the requirements.txt by opencv-python-headless==4.1.0.25
  • or if all other packages are already installed, install opencv via pip install opencv-python-headless==4.1.0.25

4. Cython compilation

• We've implemented some defenses with Cython.
• Installation
  • Make sure you are in the scaleatt directory. Just run python setup.py build_ext --inplace
  • This will compile the defenses, you might get a warning that the numpy version is deprecated (which can happen due to cython).
• Problems
  • If you have problems with the compilation step, each defense has a pure python version for each cython code, which you can easily use by setting usecythonifavailable=False at the respective code locations where we define a defense object!
  • You will need to remove the cython import in the respective defenses.
• Another note: Our timing benchmark experiments were done with a pure C++ version in the benchmark directory without any python/cython overhead. Their implementation is more optimized. So please do not use the python/cython version of our defenses to repeat the benchmark experiments.

5. Setup Python project

• Now we need to start the python project. You can do that with any IDE you want.
• The following description is based on Pycharm.
  • After installing and opening Pycharm, create a new project
    • with the directory scaleatt as Location. It will be our content root.
    • and our conda environment scaling-attack as python interpreter.
      • Please use existing interpreter and activate the scaling-attack environment that we've created in step 2.

Troubleshooting

1. Depending on your system, it might be necessary to install the package python3-tk.
2. If you get the error libSM.so.6 is missing, see step 3 to solve it.
3. Python paths:

• If you do not use an IDE or Pycharm, you might need to set the PYTHONPATH yourself once, such as: export PYTHONPATH=/home/<path_to_extracted_dir>/scaleatt/:$PYTHONPATH.
• Alternatively, go to the scaleatt directory in a terminal, and start any python script by using the following command: PYTHONPATH=/home/<path_to_extracted_dir>/scaleatt python path/to/python-file/file.py