informatiCup2019-LabWITT-CPPN

This repository is part of a submission for the informatiCup 2019 competition hosted by the Gesellschaft für Informatik. The task of the 14th informatiCup is to generate adversarial examples for a given neural network based classification API.

Our solution in this repository is designed for the following setting:

• The attacker does not have access to the dataset used to train the API classifier (and obviously does not have time to collect a similar dataset himself).
• The API classifier only returns the top-5 results (class label and confidence)
• The number of necessary API queries should be as low as possible.

This implementation uses CPPNs (Compositional-Pattern-Producing-Networks) with dynamic net depths, which are optimized using evolutionary strategies. You can find a nice blog post explaining CPPNs here.

The theoretical background, implementation details and results are highlighted in our short paper.

Example Results

Some Adversarial Images (hover to see the label), all >95% confidence:

Visualized optimization process:

The examples/ directory contains some more examples, as well as high resolution versions of the adversarial images.

Usage Guide for Linux-based Platforms

Download

Clone this repository and cd into it:

git clone https://github.com/MaximilianIdahl/informatiCup2019-LabWITT-CPPN.git
cd informatiCup2019-LabWITT-CPPN

Dependencies

This repository includes a conda environment in environment.yml as well as a pip environment in requirements.txt, both including all the necessary dependencies. We used python version 3.7.2 and although our code might be compatible with previous versions, we cannot guarantee downwards compatibility. We highly recommend installing the dependencies in a anconda environment or a python virtualenv.

To create a conda environment from this file, simply run

conda env create -f environment.yml

and activate the conda environment with

source activate infcup-cppn

or install the requirements via pip using

pip install -r requirements.txt

Prerequisites

Put your API key in the API_config.ini file.

[API]
URL = https://phinau.de/trasi
KEY = <your-api-key>

Generating Adversarial Images

• You can generate single adversarial images by running:

  python generate_adversarial.py --input_img <some-image-file> --output_dir <some-directory>

  where <some-image-file> is an image file representing the target class. You can find some example images in the test_images/ directory.

• To generate an RGB image, include the --color flag:

  python generate_adversarial.py --input_img <some-image-file> --output_dir <some-directory> --color

• Full list of argument options:

  Option	Required	Default	Description
  --input_img	True	-	Input image file representing the target class, e.g. a .png file of a stop sign.
  --output_dir	True	-	Output directory to save the results.
  --color	False	False	Use RGB instead of grayscale.
  --target_conf	False	0.95	The targeted confidence value. The optimizer stops after finding an adversarial with a confidence greater than this value.
  --high_res	False	False	Output the adversarial image in high resolution (2000 x 2000) as well as in standard resolution.
  --max_queries	False	1000	The maximum number of API queries.
  --init	False	False	Use the input image to find a good initial CPPN config (e.g. net depth). Still in experimental status.
  --no_gif	False	False	Prevent the output of a .gif visualizing the optimization process
  --random_seed	False	None	Provide a random seed for reproducible results

• You can run the run.sh script to run the program once for each class in the training set of the GTSRB dataset, i.e. the images contained in test_images/. Even though being clearly identifiable by humans, some of those images are misclassified by the API, thus leading to wrong target labels when using this script.

• The unittests can be run with

  python test.py

Overview of the Repository Contents

• examples/ contains all the generated adversarial images, in both 64 x 64 and 2000 x 2000 pixel resolutions, as well as corresponding .gif files visualizing the generation process.
• API_config.ini is a config file for the APIs URL and the used key
• cppn_init.py contains some experimental methods for initializing a CPPN based on image similarity measures. Currently in experimental phase.
• Our CPPN implementation written in PyTorch is contained in cppn_model.py.
• generate_adversarial.py is the main script, which contains all the steps of the adversarial generation process.
• util.py contains all the utility functions, e.g. loading or saving image file and performing API requests.
• environment.yml and requirements.txt contain all the dependencies installable using conda or pip respectively.
• test.py includes some basic unit testing.
• You can find images randomly sampled from the GTSRB dataset in the test_images/ directory. These were used as input images during our experiments.