Discrimination of Facial Image Generated via GAN (Work-in-Progress)

Abstract

Generated fake facial images remain a serious problem for corporations, governments, developers and individuals, as the voice of anxiety about the side effects of artificial intelligence grows. However, today the AI is still done mainly as a way to keep up with a real facial image rather than researching how to discriminate the generated image. As the world that is no longer able to distinguish between real and fake facial images is coming, the need for radical AI technology to detect generated images arises. In this paper, we introduce an approach that addresses these issues, describing in feasible detail the discriminative models based on various machine learning algorithms. Specifically, we show that the model with the highest accuracy in supervised learning achieved a 92.5% detection rate at 7.5% false positive rate (FPR), out of 400 images. And we have also achieved positive results in unsupervised learning. Our results demonstrate that the fake facial images generated by the GAN can be discriminated by the machine learning algorithms. Since GAN models tend to improve rapidly, we foresee new neural network discrimination models gaining in importance as part of a generated image detection strategy in coming years.