"In the future everyone will be anonymous for 15 minutes", British artist Banksy mocked the common urge to be famous. It seems this future is coming really fast: the latest researches showed how AI can recognize human faces even if meant to be incognito.
There are 3 basic methods of masking techniques: mosaicing (also known as pixelation), blurring and Privacy Preserving Photo Sharing or shortly P3 (encrypting part of the data). Scientists at University of Texas at Austin and Cornell Tech trained neural networks to identify relevant objects from images, then obfuscated them and continued training in finding resemblance based on the knowledge of previous images. In the end they tested neural networks and combined the tests with random guess. It turned out that in some cases success rates were around 80-90%. No surprise, P3 was ranked as the best method, but it works with JPG images only. Also the more intensely an image was pixelated, the less successful AI was in reading it. In general blurring and pixelating turned out to be not good enough to cover up a certain personality in front of a neural network.
Luckily AI has been only tested on matching the masked images with the orginal ones. It is still unable to recreate human face based on the distorted image on its own. Confronted with a video of people with blurred faces, such as CCTV footage from a train station, AI can't indentify every individual. However finding a certain person in the same crowd basing on a photo would not be a problem to neural network.
Scientists did the tests to raise awareness of privacy protection technologies problems, they don't have answers yet. Until we will find an alternative, the most common solution from the analogue world might help. Cutting out the face from a photo leaves no chance to anyone, even the neural networks, to recognize a person. Soon all videos might be filled with people having black boxes instead of heads.
Source: Wired. Image: Getty Images