Can You Tell the Real People From the GAN Images?
ModelingGANsposted by Luke Coughlin April 19, 2019 Luke Coughlin
What are GANs?
GAN stands for Generative Adversarial Network, but the term is quickly becoming synonymous with the images created by these networks. This form of machine learning uses databases of images to create new, disconcertingly realistic fabrications. These images are then run through a second “adversarial” network which attempts to discern whether the images are real or not, leading to ever-improving “fake” images.
Improvements in the quality of this technology have led to concerns over “deep fakes,” extremely convincing fake identities that use GAN images to fool real people into giving sensitive information, or worse. This potential for misuse is real, but there are many positive potential applications of the technology, particularly in fields like design and architecture, where realistic images can give designers a feel for what a new product, style, or even building might look like without going through an expensive production process.
[Related article: Efficient, Simplistic Training Pipelines for GANs in the Cloud with Paperspace]
What do GAN images look like?
See for yourself! Below are ten faces, five of which are GAN generated and five of which are real people. See if you can tell the real ones from the computer-generated.
ANSWERS: 1. Fake! 2. Fake! 3. Czech novelist Milan Kundera 4. Fake! 5. Reigning Queen of Denmark Margrethe II 6. Fake! 7. Singer/Songwriter Elvis Costello 8. Singer/Songwriter Joan Baez 9. Fake! 10. Irish Prime Minister Leo Varadkar
The possibilities presented by GAN technology are both exciting and disquieting. Will hackers and cyber-criminals use these networks to further their schemes? Many videos already exist of celebrities being made to say things that they never said, and certain stories about scammers or stalkers using generated images to perpetuate false identities are circulating around the cybersphere. At the same time, creatives in various fields will be able, as GANs improve, to make use of highly realistic fabricated images in drafting designs and projects. As with all technological advancements, it is likely that GAN networks will be used for both good and bad ends.