Reproduce “Leveraging Frequency Analysis for Deep Fake Image Recognition” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Joel Frank, Thorsten Eisenhofer, Lea Schönherr, Asja Fischer, Dorothea Kolossa, Thorsten Holz
Researchers at Ruhr-University Bochum systematically analyzed the frequency domain of images generated by various GAN architectures, revealing ubiquitous grid-like artifacts originating from upsampling operations. They leveraged these spectral anomalies to develop a deep fake detection method that achieved 100% linear separability for high-resolution fakes and significantly improved source identification while exhibiting greater robustness and training efficiency.
Key insights
- All tested GAN architectures consistently produce severe, distinct grid-like artifacts in their DCT spectra, which are absent in natural images.
- These spectral artifacts are directly attributable to the upsampling operations in GAN generators, indicating a structural limitation in current generative models.
- Deep fakes are perfectly linearly separable from real images in the frequency domain, offering a highly discriminative feature space for detection.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.