Abstract

With the development of advanced machine learning techniques, it is now possible to generate fake images that may appear authentic to the naked eye. Realistic faces generated using Generative Adversarial Networks have been the focus of discussion in the media for exactly this reason. This study examined how well people can distinguish between real and generated images. 30 real and 60 generated were gathered and put into a survey. Subjects were shown a random 30 of these faces in random sequence and asked to specify whether or not they thought the faces were real. Based on a statistical analysis, the participants were not able to reliably distinguish between all real and generated images, but real images were correctly distinguished in 81% of cases, where generated images were correctly distinguished in 61% of cases. Some generated images did receive very high scores, with one generated image being classified as real in 100% of the cases.

COinS