UPDATE: Link to preliminary implementation and writeup here

Criminal Kitty This project inspired by this lovely kitty I breeded, which I admit looks like a criminal…

CryptoKitties

I’ve been spending a lot of time in the dark, neckbeard basements of blockchain. Recently, I stumbled upon CryptoKitties, which their homepage advertises as:

CryptoKitties is a game centered around breedable, collectible, and oh-so-adorable creatures we call CryptoKitties! Each cat is one-of-a-kind and 100% owned by you; it cannot be replicated, taken away, or destroyed.

There’s a lot to unpack here and let’s get straight to it:

Algorithmic Bias

I’ve also been spending a lot of time reading about human rights, ethics, and social good. The physiognomy article we read for class really piqued my interest because as junk science as it can be people are actually using it. Perhaps even more troublesome is that beyond the obviously visceral reactions to phrenology, opaque algorithms do encode subtle biases. For example, there are articles floating about claiming self-driving cars do a poor job at detecting dark-skinned pedestrians.

But alas, it seems people avoid these heavy topics.

Phrenology

The physiognomy reading for class is interesting because despite being so wrong, people can still find ways to draw such meaningful conclusions. Many problems abound in the study. Of note:

At the same time, our guts tell us that there are interesting questions in the realm of Physiognomy. That is, we believe that they do have some measure of prediction – e.g. a beautiful (proportional?) person is generally perceived to have an easier life, making the opposite scenario plausible.

Plan

Anyway, I wanted to explore intersecting these interests, that is junkish/inaccurate data science, bringing light to algorithmic bias by attacking the heart of the internet… cats, and exploring the world of digital collectibles.

  1. Collecting Data – gathering genomes and images of a large diversity of CryptoKitties.
  2. Labeling Data – IMO this is where the project gets interesting because there isn’t an actual database of criminal kitties. Instead, we get the beautiful opportunity to do shoddy data collection, meaning get our own labels. Instead of real data, I propose we use peoples’ gut reactions toward CryptoKitty images. Also let’s explore getting ITP to label the data, and compare that with results from mechanical turk.
  3. Train a Neural Network – a neural network is probably overkill, but something I find fascinating about neural networks are their opaqueness, and hence their deniability of blame when questioned.
  4. Profit? – Not sure. I’d love to build an installation to put some CryptoKitties in jail or a chrome extension that stems the sale of criminal CryptoKitties. Hell, maybe we can build a border wall on the blockchain…

Next Steps

The project is implemented here! https://github.com/epylinkn/operation-criminal-kitty

References