Scientists explain how to distinguish a deepfake from a real photo

A group of American scientists has discovered a feature that allows distinguishing deepfakes from photos of real people by examining the eyes.

Researchers found that most generative adversarial networks (GANs) used to create fakes do not draw round pupils. Such artifacts are widespread among high-quality deepfakes, they say.

“The incorrect shape of the pupil is a good sign that a person can identify a face generated by a GAN. Even if there are no boundary marks around the pupils, we will easily notice their irregular shapes, which can vary significantly on the same deepfake image,” the paper states.

Researchers note that in practice people can enlarge the image and check the pupil shape to detect a forgery.

Back in June, Facebook developed an algorithm for detecting deepfakes, which tracks the source of the fabrication.

In April, experts stated that the number of online forgeries doubles every six months.

In March, scientists developed a tool that based on reflections of light in the cornea can recognise a deepfake.

Follow ForkLog news on Telegram: ForkLog AI — all the news from the world of AI!