Photo Source: Pikrepo
Fake people and stories threaten to infiltrate social media and challenge the nature of truth in what we see and read online and on television. Advances in machine learning continue to make the generation of believable fakes easier and more convincing. One source of fake content, such as images, comes from a relatively new type of machine learning called generative adversarial networks or GANs for short.
A GAN works by pitting two machine learning systems against each other—in other words, they learn through battling back and forth about the "truth." GANs have two systems—one trained on the truth about a particular subject such as recognizing human faces in images and another system that tries to fool the first system by making fake images. The first system goes by the name "discriminator" because it judges the real from the imitation, and the other computer system called the "generator" because it pumps out forgeries to fool the discriminator.
What makes GANs special lies in the relationship between the discriminator and the generator—they help each other learn, and, in the process, each one becomes better at their respective jobs. They communicate back and forth after every forgery and judgment. The discriminator will tell the generator if it performed better or worse than the last forgery, allowing the generator to alter itself and improve. Such a powerful arrangement lets both systems to learn and grow. By analogy, imagine if a master forgery artist specializing in Dutch Master paintings such as Rembrandt's could work intimately with a top art historian every day arguing and refining fake masterworks. Together, they could create new masterworks nearly indistinguishable from legitimate ones.
GANs emerged in 2014 from the work of Ian Goodfellow. Goodfellow invented GANs while finishing up his Ph.D. in machine learning at the Université de Montréal in Canada. He first introduced GANs in a paper titled "Generative Adversarial Nets" (NIPS 2014). Since then, GANs have become one of the hottest topics in artificial intelligence affecting many areas of research and commerce from pharmaceutical discovery and robotics training to game development and simulator creation.
A live example of GANs in action resides at a free public website called thispersondoesnotexist.com. The site created by Phil Wang generates unique pictures of people based on the StyleGAN program. The following two examples represent unique images of people that StyleGAN generated on June 1, 2020.
The remarkable detail of the synthetic images demonstrates the power of GANs to generate realistic images of faces that do not match any real people (reverse google image search).
GANs work on more than just images. Another example of GANs in action involves the generation of fake news stories. A site called "World News Articles (AI written)" (thisarticledoesnotexist.com/) regularly generates synthetic news. The site grabs headlines every thirty minutes from Reddit and Hacker News and generates fake stories to go along with the headline. The program creates fake data and quotes from real individuals. For example, a fake article on coronavirus from the site makes up quotes from seemingly official sources such as the World Health Organization and the Chief Medical Officer of Australia. The artificial article titled "Australia's Chief Medical Officer advises against injecting disinfectant to treat coronavirus," includes quotes from professors and public officials. The report states, "The head of the Australian Health Department, Professor Susanne Bartsch, says the global vaccine effort is about to find the source of the virus." (Link to article). Although not convincingly realistic, fake news demonstrates how automated content production will continue to expand with GANs.
Advances in machine learning and artificial intelligence took a significant step forward with the invention of generative adversarial neural networks or GANs by Ian Goodfellow at the Université de Montréal in Canada. Dr. Goodfellow's introduction of GANs opened the door to computers creating novel images such as faces and text with synthetic stories. GANs use two systems, a generator that makes content and a discriminator that judges whether that content is real or fake. The power GANs comes not from the generator trying to fool the discriminator, but that they share the results of each battle to improve their performance. The process has demonstrated success in generating synthetic images of people and fake stories based on headlines. Researchers continue to look for new applications of GANs in many fields, from pharmaceutical research to robot development. As GANs continue to improve at generating synthetic content, researchers will also need to develop tools that can better discriminate the real from the fake.
Dr. Smith’s career in scientific and information research spans the areas of bioinformatics, artificial intelligence, toxicology, and chemistry. He has published a number of peer-reviewed scientific papers. He has worked over the past seventeen years developing advanced analytics, machine learning, and knowledge management tools to enable research and support high level decision making. Tim completed his Ph.D. in Toxicology at Cornell University and a Bachelor of Science in chemistry from the University of Washington.
You can buy his book on Amazon in paperback here and in kindle format here.