Episode 56 - True News, Fake Faces, and Adversarial Algorithms
Google's "VP of Trust and Safety" has put out Google's strategy for weeding out fake news. Do they have the right definition? Will it work? What is a VP of Trust and Safety? We'll discuss whether Google is looking out for their users.
In that vein, we wanted to discuss new ways fake news might be generated. The site thispersondoesnotexist.com created images of people who are not real but look pretty convincing. They use Generative Adversarial Networks. After you listen, you'll know how GANs work to create plausible images.
Live Stream
This is also the first time we’ve livestreamed our discussion on YouTube. This is still experimental (Aaron and I are working out the kinks) but I’ll paste what we recorded below:
Links
Google’s Blog Post on Fighting Disinformation
Link to Google’s PDF Whitepaper
The Google News Initiative
Background on VP of Trust and Safety, Kristie Canegallo, White House Page
On the Generative Adversarial Network for Photos of Human Faces
thispersondoesnotexist.com: The generator
Whichfaceisreal.com: The discriminator
Article on the website inverse.com
Article from interestingengineering.com
For a more technical treatment:
2014 Paper on GANs by Goodfellow et al.
GANs: A Beginners Guide
GANs on Wikipedia
The following face looks good enough, but what’s up with her friends mouth?
Related Episodes
Episode 50 on Bypassing Gatekeepers
Episode 45 on Inside Google
Episode 33 on Ads, Bias, Censorship, Democracy
Episode 28 on the Banning of Alex Jones
Episode 25 on The Great Twitter Purge. They also have their own VP of Trust and Safety.
Episode 9 on Fixing Facebook
Episode 4 on Codebreaking with Multiple Local Maxima