Exposing GAN-Generated Profile Photos From Compact Embeddings
Generative adversarial networks (GANs) have been used to create remarkably realistic images of people. More recently, diffusion-based techniques have taken image synthesis to the next level. From only a text prompt, these techniques can synthesize any image seemingly limited only by our imagination. Along with the many clever and creative use cases, synthetically-generated faces are being used to create more convincing fake social-media profiles. We describe two related techniques that learn low-dimensional (128-D) embeddings of GAN-generated faces. We show that these embeddings capture common facial structures found in these synthetically-generated faces that are uncommon in real profile photos. These low-dimensional models, trained on a relatively small data set, achieve higher classification performance than larger and more complex state-of-the-art classifiers.