shawn-shan/fawkes

Fawkes is a privacy protection tool built by researchers at the University of Chicago. Its purpose is to make small, invisible changes to photos of your face so that facial recognition systems cannot correctly identify you. The changes are so subtle that a human looking at the photo would not notice anything different, but machine learning models that try to learn your face from those photos end up with a distorted picture of who you are. The project was published as an academic paper at USENIX Security 2020, one of the top computer security conferences. The researchers call these invisible edits "cloaks." When a cloaked photo gets fed into a face recognition training pipeline, the system memorizes a version of your face that does not match what you actually look like. Later, when that system tries to recognize your real, uncloaked face, it fails. You run Fawkes from the command line, pointing it at a directory of images. There are three protection modes: low, mid, and high. Higher modes add more distortion and offer stronger protection but are more visible. Processing takes about 60 seconds per image on a regular laptop CPU, and considerably faster on a machine with a GPU. The output is a new set of images in your chosen format, such as PNG or JPG, with the cloaking applied. Installation is straightforward: it is available on PyPI, the standard Python package index, so you can install it with one command. A desktop application for Mac and Windows is also available on the project website for people who prefer not to use the command line. The code is intended for personal privacy protection or academic research. If you are a non-technical person, the practical use case is this: before uploading photos of yourself to social media or any public site, you run them through Fawkes first. The photos look normal to anyone viewing them, but any service that tries to train a facial recognition model on your images will end up unable to recognize you in real life.