Leveraging Frequency Analysis for Deep Fake Image Recognition
2020Conference / Medium
Research Hub
Research Hub C: Sichere Systeme
Research Challenges
RC 9: Intelligent Security Systems
Abstract
Deep neural networks can generate images thatare astonishingly realistic, so much so that it isoften hard for humans to distinguish them fromactual photos. These achievements have beenlargely made possible by Generative Adversar-ial Networks (GANs). Whiledeep fakeimageshave been thoroughly investigated in the imagedomain—a classical approach from the area ofimage forensics—an analysis in thefrequencydomainhas been missing so far. In this paper,we address this shortcoming and our results re-veal that in frequency space, GAN-generated im-ages exhibit severe artifacts that can be easilyidentified. We perform a comprehensive analysis,showing that these artifacts are consistent acrossdifferent neural network architectures, data sets,and resolutions. In a further investigation, wedemonstrate that these artifacts are caused by up-sampling operations found in all current GANarchitectures, indicating a structural and funda-mental problem in the way images are generatedvia GANs. Based on this analysis, we demon-strate how the frequency representation can beused to identify deep fake images in an automatedway, surpassing state-of-the-art methods.