A Hack Cambridge Ternary project
You, the drawer, create a drawing of a certain word or phrase. Your friend, the guesser, tries to guess what that word is, by looking at your image. However, a Convolutional Neural Network trained on hundreds of thousands of images is also in the running - and if the AI can guess the image, you both lose!
You're thus encouraged to come up with interesting, alternative ways of depicting the given word or phrase. Maybe draw a flamingo with its head under the water, or an upside down umbrella - the choice is yours!
Bad Flamingo isn't just a game. The robustness and security of machine learning algorithms are becoming increasingly critical as AI systems make important decisions in work, life, and play. Crucial to understanding ML security are adversarial training examples, which are training examples that humans easily identify but puzzle machines.
Most adversarial training examples are generated by adding imperceptible noise to the training examples. In Bad Flamingo, however, users must draw fundamentally alternative images to fool the ML classifier. These proposed semantic adversarial examples test relations between captions and images that humans can understand, but ML models cannot. Playing Bad Flamingo will thus generate more difficult datasets for modern ML models and encourage further research in robust, human-like computer vision.
To start the server, run
npm install
npm run devCurrently matchmaking infrastructure hasn't been set up yet. Simply navigate to the following addresses on any browser:
{IP_ADDRESS}:8000/#/{GAME_CODE}/0to play as the drawer{IP_ADDRESS}:8000/#/{GAME_CODE}/1to play as the traitor (see below){IP_ADDRESS}:8000/#/{GAME_CODE}/2to play as the guesser
where GAME_CODE is any alphanumeric game code, and IP_ADDRESS is the
address of the machine hosting the server (e.g. localhost if running on your
own computer).
We noticed that simple ways of fooling the AI include drawing squiggles over an image, or surrounding an image with a border or a box, and others.
One option of fixing this (and making the game slightly more interesting) is to include a third player, the traitor, who is on the side of the AI. The traitor is not given the original label of the image, but is allowed to erase lines in the image to attempt to make it more interpretable for the AI. The guesser sees the original image as usual.
If you don't want to play with the traitor, then (for now) simply click submit on his player - options for enabling/disabling this player are forthcoming.
Original name for this project inspired by Colin Morris's blog post.