HideFace

Please see http://bit.ly/rccreager_slides for the slides corresponding to this package!

HideFace is a testing framework for face detectors and adversarial attacks. It was built with the intention of participating in Google's Adversarial Examples Challenge. This framework is useful for adding privacy filters to images (to hide them from face detection) or for testing the robustness of face detection algorithms against attacks.

Generally, we can divide image attacks into two classes:

• White-box: attacks that use information about the algorithm being attacked (for example, model weights)
• Black-box: attacks that assume no knowledge of the algorithm being attacked

This package is designed for black-box attacks against facial detection algorithms because no information about the model itself is used.

Attacks are implemented in the hideface/attacks.py file -- a few examples are provided. Detectors are provided as a dictionary in the run_attacks.py script -- for now, only the DLib HoG detector is used.

Setup and Installation

Setup with Docker

The easiest way to run this package is with Docker.

1. Install Docker

2. Pull the Docker image from Docker Hub:

    docker pull rccreager/hideface:1.0
   

3. Run the Docker image:

    docker run -it rccreager/hideface:1.0 
   

Now you can go to the next section, ``Usage''.

Build yourself on Ubuntu

This package uses DLib for its face detection -- this is a C++ package and it takes a long time to build, so be patient! The script build/apt_get.sh includes sudo commands to install all the necessary systemwide tools. After that, you can use build/requirements.txt to install the needed python packages to your pip3. I would recommend setting up a virtual environment first (the second and third lines):

git clone https://github.com/rccreager/hideface.git
cd hideface
source build/apt_get.sh
python3 -m venv build/hideface_env
source build/hideface_env/bin/activate
pip3 install -r build/requirements.txt

Usage

For a test and example of how this package works, run example_attack.py, which will test 25 images (1 attack test each) from a subset of the WIDER-FACE dataset:

python3 example_attack.py

You can create some histograms to visualize the attack performance across all images tested in your example like so:

python3 example_histo_plots.py