Shulou(Shulou.com)06/03 Report--

Big Data Digest

Authors: Jiang Baoshang, Wei Zimin

Let's start with a simple quiz.

Here are several sets of images, and don't doubt that one of each set is a composite "fake face."

False on the left.

The right is false.

False on the left.

StyleGAN: Image Generation New Artifacts

The three images above are from a website that has recently exploded on reddit-"which face is real?" ", netizens are very enthusiastic about the results on the website competition, and posted the test results.

You can open the following URL and do more photo recognition tests yourself

http://www.whichfaceisreal.com/index.php

Generative Adversarial Networks (GANs) have changed dramatically since Ian Goodfellow proposed them in 2014.

The early images were also very "hot eyes," far from deceiving. For example, this batch of "fake faces" in 2004.

However, after more than ten years of development, the pictures generated now are almost the same as the real ones, and they cannot be recognized by old drivers.

For example, you can see some of the pictures at the beginning of the article. Seriously, Digest Fungus is almost all about "Meng" when it first starts playing! There were times when the fake face was even more realistic than the real one.

After reading the relevant introduction, sure enough, all the pictures used on the website were generated by StyleGAN.

StyleGAN comes from Nvidia, which can be said to be a "fake" artifact that has recently become popular all over the Internet. Unlike other generators, StyleGAN can change the results of the generated images as needed to draw more realistic images, not only to create fake human portraits, but also to be wildly applied to other machine learning applications, such as cars, rooms, and even anime heads.

Fake house net ": from pictures to text descriptions are automatically generated by computers

The picture above is the "fake Airbnb" website that has been discussed a lot recently on Reddit. It is also generated by StyleGAN. The pictures and text on the website depict nothing at all.

Website Address:

https://thisairbnbdoesnotexist.com/

Every time the fake house generation website refreshes, a fake house will appear. The photos, text descriptions and published portraits on the website are automatically generated by the computer. Because the model used is very simple, the text description is illogical, but at a glance it can still be false.

StyleGAN was originally developed by NVIDIA in a paper entitled A Style-Based Generator Architecture for Generative Adversarial Networks.

Paper download address:

https://arxiv.org/pdf/1812.04948.pdf

StyleGAN is a step-by-step approach to generating artificial images, starting at very low resolution and working its way up to high resolution (1024×1024). By modifying the inputs at each level of the network separately, it can control the visual features represented at that level, from coarse features (pose, facial shape) to fine details (hair color), without affecting the other levels.

This technique not only allows for a better understanding of the generated output, but also produces the highest level of results, i.e., high-resolution images that look more realistic than previously generated images.

StyleGAN is a breakthrough technology that not only generates high-quality and realistic images, but also allows better control and understanding of the generated images, and even makes it easier to generate fake images with high credibility than before. Some of the techniques proposed in StyleGAN, particularly mapping networks and adaptive instance standardization (AdaIN), may be the basis for many future innovations in GAN.

"Counterfeiting" has skills! Tips for identifying fake pictures

Is there no way to identify such a realistic fake picture? Don't panic, no matter how smart AI is, it will still leave some traces when it is fake. Professional researchers, while faking, also left some tips for "cracking down on counterfeit goods."

spots

StyleGAN algorithm, although powerful, but there is a significant feature, is that the generated images tend to have shiny spots, although these spots look like photos aged, the product of chemical reactions, but this is indeed the fatal flaw of these composite images.

These spots can appear anywhere in the image, with hair and background areas most likely to appear.

These speckled pictures, they're all fake, they're all GAN composite.

background issues

Another fatal flaw can be found in the background image of the photo, where neural networks tend to pay little attention to recognizing faces in people's images. In some cases, the background of the photo will appear very messy. Don't think too much. This is not an impressionist painting. It is the neural network that does not handle the background well enough when generating the image.

glasses

Even though StyleGAN has been very powerful, like its predecessors, it also cannot handle glasses perfectly. The most common problem is that the two sides of the glasses are not symmetrical. Take frames for example, usually composite images, the frame style on the left and the frame style behind are not the same, as shown below, one side of the frame sometimes appears curved and jagged.

Other asymmetric problems

In addition to glasses, facial hair asymmetry can sometimes occur. The earrings worn on the left ear and right ear are also different, and the collar will also have different shapes on the left and right sides.

Symmetry is now often a challenge for face generation algorithms, and we can exploit this weakness to kill them.

hair

In general, composite images of people, hair is often not very realistic, sometimes the hair will be broken on the face or elsewhere, as shown in the first image below, sometimes the hair of the person will be too straight, and will appear striped. As Kyle McDonald says, it's like someone messed up a pile of acrylic resin with a palette knife. In some cases, strange halos or halos may appear around the hair, as shown in the middle image:

background fluorescence

Another interesting drawback is that fluorescent colors sometimes appear from the background to the hair or face.

teeth

Teeth are not easy to render. Teeth are usually odd or asymmetrical. In some cases, composite images of people will show three front teeth.

If you're still scared of GAN-generated images, Digest Fungi once wrote a more detailed article titled GAN-generated fake faces are too realistic! Don't be afraid, ten tips to teach you to identify AI-generated fake images to help you identify fake photos, stamp here for more tips.

Now that you know how hard it is for neural networks to generate anything, you can find the fatal flaws in each composite image like a game of finding fault and develop confidence in recognizing real images.

"Counterfeiting" Skill Assessment Begins

After reading these skills, let's test your ability to crack down on fake goods. Let's test a few groups of pictures to see if your ability to "look at pictures" has improved.

Fake on the left. Pay attention to the hair aperture.

The one on the right is fake. Notice the messy background.

The one on the right is fake. Watch for the flare.

Where did the website of "counterfeiting" come from?

Finally, I would like to introduce to you how this interesting "counterfeiting" website was born.

This project comes from a popular course at the University of Washington called "Resisting Bullshit--Calling Bullshit." The lecturers were Carl Bergstrom from the Biology Department and Jevin West from the School of Informatics.

Part of the class's popularity stems from its unruly name, and as for why it's called "bull shit," the course's introduction is designed to counter the various types of bullshit information that currently exist, including language, data, graphs, and other forms of presentation that blatantly disregard facts and logic.

And, of course, beneath that vulgar title is a very serious curriculum.

The two teachers posted the full course content and readings on the syllabus page of the course website. Interested friends can read carefully. Below I will give you a brief introduction and analysis.

Curriculum:

http://callingbullshit.org/syllabus.html

In the introduction, the course uses Princeton professor Harry Frankfurt's article "On Bullshit." He actually published a book called On Bullshit. Nanfang Shuo translated it into Chinese and published it in Taiwan under the title Farting! A shortcut to fame and fortune was published in mainland China with a very conservative translation of "On Bullshit."

Week 2 introduces some common ways to distinguish bullshit. Week 3 is about the ecosystems that breed shit, such as how social media facilitates the spread of shit.

Over the next few weeks, the class went into statistical and logical detail on a number of types of bullshit, including confounding correlations and causations, medians and averages, and the prosecutor's fallacy. The course also includes a separate week on common misdirection in data visualization.

The big data part of week 7 focuses on the phenomenon of "garbage in and garbage out" under the bright appearance of big data and algorithms, as well as the abuse of machine learning and misleading parameters.

In the weeks that followed, he delved into science, introducing concepts such as "publication bias,""predatory publication," and the ethics of criticism within and between disciplines.

Week 11 is about fake news. The topics include the economic drivers of fake news, echo chamber effects, and how to conduct fact-checking, all of which are frequently discussed in the News Lab. If the course were offered in a journalism school, there would be enough of it to expand into an entire course. But because the class focuses on science, journalism is compressed into one lesson.

The last week was all about refuting bullshit. Different strategies are needed for different audiences. The content of this aspect is basically the study of persuasion effect in communication science.

This course started in the spring of 2017 and is now over. The teacher has also put all the videos of the course on YouTube. Interested students open the following URL to watch oh?

https://www.youtube.com/playlist? list=PLPnZfvKID1Sje5jWxt-4CSZD7bUI4gSPS

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.