Machine Vision Calamities

  Source: Devumi bot retweet sales

Source: Devumi bot retweet sales

Let's look at how increasing computing power and algorithm efficiency are leading to some pretty wacky technology in the realm of computer vision. The building blocks are as follows. Neural networks can be trained on large data sets of objects to recognize those objects. They run on video cards (GPUs) and power everything from tagging cat photos to Tesla's self-driving cars. The more GPUs, the more things you can recognize, and the better your data and algorithm efficiency, the more accurate your recognition. 

So here's the example -- Amazon and its magic store, Amazon Go. The company has been testing a check-out free shopping experience for a few years, and the acquisition of Whole Foods has only encouraged speculation about the future of food retail. New information has come out about how the technology works. First, a shopper scans an identifier on their phone when entering the store. From that moment on, the hundreds of video cameras on the ceiling watching all the activity in the store track every single shopper and every single product on video. To do this successfully, not only do you need gazillions of hours of footage (i.e., what Amazon is in fact doing), but a massive cloud infrastructure to process the machine vision demands in real time. Good thing there's AWS!

The same neural network that can recognize images can also hallucinate them. Generative neural networks can manufacture images of a type, where the type is their source data set. And if you put an editor on top of that, like an adversary, you can manufacture pretty accurate renditions of whatever it is you want.

Thus, deepfakes. In their current NSFW form (and this is how the trend is being reported), deep fakes use machine vision to swap out the faces of celebrities onto adult entertainment. But that's just the beginning. Using a free desktop app called FakeApp, a derivative of the many mobile face-swap apps, a user can masterfully replace one speaker's face with that of another. And the effects can be good enough to look better than a multi-million dollar 3D rendering by the best Hollywood studios.

Samantha Cole at Mortherboard, which broke this article, goes on to say -- "An incredibly easy-to-use application for DIY fake videos—of sex and revenge porn, but also political speeches and whatever else you want—that moves and improves at this pace could have society-changing impacts in the ways we consume media. The combination of powerful, open-source neural network research, our rapidly eroding ability to discern truth from fake news, and the way we spread news through social media has set us up for serious consequences."

Yeah, it's not great. Especially when such messages can be validated for peanuts on social networks using cheap bot armies. According to the New York times, the going rate for 25,000 fairly active Twitter bots is $225. Want to know where the profile descriptions and pictures come from that make these bots look like real users? Stolen identities from humans. 

  Source: Top frame shows rendered Carrie Fischer in Star Wars, bottom one uses FakeApp

Source: Top frame shows rendered Carrie Fischer in Star Wars, bottom one uses FakeApp