Google's DeepMind Can Now Create Fake Photos That Look Totally Realistic

Wednesday, 03 October 2018 - 12:41PM
Weird Science
Wednesday, 03 October 2018 - 12:41PM
Google's DeepMind Can Now Create Fake Photos That Look Totally Realistic
< >
Image: Outer Places from Pixabay
A few weeks ago, we covered Congress' letter to the Director of National Intelligence, which asked the agency to look into 'deepfakes', images and videos altered by 'deep learning' AI software. What makes deepfakes so dangerous is that the edits (especially face swaps) can look real enough to fool most people, even under close scrutiny. Now, Google's DeepMind has taken things a step further—instead of editing someone into a compromising photo or video, they can create photo-realistic images almost from whole cloth. Their new research paper, titled "Large Scale GAN training for high fidelity natural image synthesis," was posted on Arxiv.org.

Like other 'deep-learning' neural nets, DeepMind's network was trained by showing it huge volumes of real photos, first from ImageNet.org, then from JFT-300M (which hosts 300 million pictures). The network was dubbed 'BigGAN' because the system uses Generators and Discriminators (GAN), where the former attempts to generate realistic-looking new images and the latter tries to distinguish between which images are real and which are fake. According to the pre-published study, the DeepMind team managed to significantly increase the record for realistic-looking computer-generated images, scoring 166.3 on the Inception Score and 9.6 on the Frechet Inception Distance Score. These numbers are double or triple the previous best scores (52.52 and 18.65, respectively).

Most of the photos synthesized by BigGAN are frighteningly accurate—a brown dog facing the camera, a shot of a tropical island landscape, a juicy burger, all of which look 100% real, even when you see the photos that were used to make the composite. The new photo just looks like any other, down to the lighting and the gritty details. It's astonishing, and demonstrates that AI identification of deepfakes may become a growth industry very, very quickly.

You can see some of the sample sheets from the study here.

In the meantime, people will continue to edit Nic Cage's face into blockbuster movies

Science
Artificial Intelligence
Weird Science