Reproduce “Exploring Weaknesses of Generative Image Watermarks against Latent Frequency Masking” 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

Kirill Aistov, Khaled Abud, Irina Serzhenko, Egor Kovalev, Aleksey Yakushev, Aleksandr Akimenkov, Dmitry Obydenkov, Yury Markin and 3 more

LFM attacks generative-image watermarks by replacing selected Fourier coefficients in a latent diffusion autoencoder representation and reconstructing the image. Gaussian replacement avoids diffusion generation. On DiffusionDB, averaged across six watermark methods, it leaves TPR@1%FPR of 0.36 and PSNR of 21.6 dB, compared with UnMarker’s 0.35 and 12.8 dB. It can weaken tested watermarks, but Gaussian Shading resists partial masking.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.