You Can Buy a Random Facial Recognition Photo on China’s Black Market for Just $.07
Welcome to General Intelligence, OneZero’s weekly dive into the A.I. news and research that matters.
A picture may be worth a thousand words, but apparently a picture of a face is only worth seven cents.
Black-market sellers in China are now offering packages of up to 20,000 face images and personal information data that can be used to impersonate others for as cheap as $0.07 each, or 0.5 yuan, according to a report from China’s state-owned news outlet Xinhua.
And it’s not just still images in these packages. For around $5, sellers will also provide deepfake-like software that animates a still image to bypass some security features on popular dating apps in China like Tantan. These apps require a person to nod or blink in order to verify that the phone isn’t being held up to a still image, according to a report on similar practices in the South China Morning Post.
These images are marketed on Chinese shopping websites like Taobao and Xianyu. When a buyer expresses interest, sellers then move the conversation to messaging apps like WeChat or QQ, where they share links to cloud storage drives with the images, data, and software to animate faces.
Much of China’s telecommunications infrastructure is being linked to facial recognition, meaning everyone from internet providers, social media apps, and banks are requiring face scans to verify a person’s identity as they use their devices. Being able to trick those systems would be basically assuming another person’s identity online.
It’s unlikely these impersonations would work on smartphones with 3D face scanners like the iPhone, which matches the topography of a person’s face rather than just an image. However, many mid-range and budget smartphones don’t have these kinds of sensors.
Most apps in the U.S. rely on the smartphone’s authentication, whether that be a fingerprint scanner or facial recognition like FaceID. The apps themselves do not attempt to perform facial recognition, making it less likely that these techniques would work as well in the States.
Everyone from internet providers, social media apps, and banks are requiring face scans to verify a person’s identity.
But these packages, including face photos, ID and phone numbers, and banking information, are a glimpse at what identity theft looks like in a world where your face is your password.
And now, here’s some of the most interesting A.I. research of the week:
OpenAI’s Text-Generating Algorithm Learns To Make Images
OpenAI’s GPT-2 algorithm is an enormous neural network that was able to analyze hundreds of millions of texts and spit out its own bits of writing. Seeded with a sentence, it could write a pretty haunting short story. Now, OpenAI is turning to the same algorithm to generate images, in an shockingly successful experiment.
NEC Ups Its Facial Recognition Chops
Facial recognition datasets are typically labeled images of people’s faces, meaning that if you have four photos of me, each is tagged with my name or some other identifier. But making these datasets by tagging every photo is time consuming. NEC has proposed a way to build large-scale facial recognition datasets without needing to label each identity or face. It uses a second algorithm to guess which faces are the same in a batch of millions of unlabeled images, and then adds those images with “pseudo-labels” into the original dataset.
Nvidia’s Play For More Realistic Video Games
Textures make video games look good. New research by Nvidia allows an algorithm to create realistic textures with life-like variation, a stepping-stone for A.I. able to create better virtual worlds in video games.