Pattern Recognition (or so called Artificial Intelligence) can be tricked. An overview.
Do you aim to become a luddite? Here is your guide to hacking pattern recognition and disturbing the technocratic wet dreams of engineers, managers, businesses and government agencies.
Most of the current processes attributed to Artificial Intelligence are actually pattern recognition, and artists and scientists have begun to work with adversarial patterns either to test the existing techniques or to initiate a discussion of the consequences of the so called Artificial Intelligence. They create disturbances and misreading for trained neural networks that get calculated against incoming data.
Do neural networks dream of sheep?
Janelle Shane looks into how neural networks just mis-categorize information. In her article Do neural nets dream of electric sheep? she discusses some mis-categorizations of Microsofts’ Azure Computer Vision API, used for creating automatic image captions. Shane points out, that the underlying training data seems to be fuzzy, since in many landscape pictures sheep got detected, where are actually none. »Starting with no knowledge at all of what it was seeing, the neural network had to make up rules about which images should be labeled ›sheep‹. And it looks like it hasn’t realized that ›sheep› means the actual animal, not just a sort of treeless grassiness.«
The author then looks into, how this particular pattern recognition API can be further tricked, pointing out that the neural network looks only for sheep where it actually expects it, for instance in a landscape setting. »Put the sheep on leashes, and they’re labeled as dogs. Put them in cars, and they’re dogs or cats. If they’re in the water, they could end up being labeled as birds or even polar bears. … Bring sheep indoors, and they’re labeled as cats. Pick up a sheep (or a goat) in your arms, and they’re labeled as dogs«, Shane mocks the neural network. I’ll call it the abuse scope method. It applies, whenever you can determine or reverse-engineer (aka guess) the scope and domain to which a neural network is directed, and insert information that is beyond the scope. The abuse scope method could be used for photo collages that trick a neural network, while maintaining relevant information to humans.
Shane went further and asked twitter followers for images depicting sheep. Richard Leeming came up with a photo taken in the English country side. Orange dyed sheep shall deter rustlers from stealing the animals.
This photo is fucking with the neural networks’ expectations and leads to a categorization as »a group of flowers in a field« (Shane 2018). Continue reading