Image-Scaling Attacks & Defenses

This repository belongs to our publication:

Erwin Quiring, David Klein, Daniel Arp, Martin Johns and Konrad Rieck. Adversarial Preprocessing: Understanding and Preventing Image-Scaling Attacks in Machine Learning. Proc. of USENIX Security Symposium, 2020.

Background

For an introduction together with current works on this topic, please visit our website.

In short, image-scaling attacks enable an adversary to manipulate images, such that they change their appearance/content after downscaling. In particular, the attack generates an image A by slightly perturbing the source image S, such that its scaled version D matches a target image T. This process is illustrated in the figure above.

Getting Started

This repository contains the main code for the attacks and defenses. It has a simple API and can be easily used for own projects. The whole project consists of python code (and some cython additions).

Installation

In short, you just need the following steps (assuming you have Anaconda).

Get the repository:

git clone https://github.com/EQuiw/2019-scalingattack
cd 2019-scalingattack/scaleatt

Create a python environment (to keep your system clean):

conda create --name scaling-attack python=3.6
conda activate scaling-attack

Install python packages and compile cython extensions:

pip install -r requirements.txt
python setup.py build_ext --inplace

Check the README in the scaleatt directory for a detailed introduction how to set up the project (in case of problems).

That's it. For instance, to run the tutorial, you can use (assuming you're still in directory scaleatt and use BASH for $(pwd)):

PYTHONPATH=$(pwd) python tutorial/defense1/step1_non_adaptive_attack.py

Tutorial

Jupyter Notebook

For a quick introduction, I recommend you to look at this jupyter notebook.

Main Tutorial

Check the directory scaleatt/tutorial/ for a detailed tutorial how to run the attacks and defenses.

The directory has the same structure as our evaluation. Each subdirectory corresponds to the subsection from our paper:

• The directory defense1 corresponds to experiments from Section 5.2 and 5.3
• The directory defense2 corresponds to experiments from Section 5.4 and 5.5
  • Each subdirectory contains some python scripts that describe the API and the respective steps.

My recommendation: Open each file (in the order of the steps), and then use a python console to run the code step by step interactively.