Reproduce “Pre-trained Perceptual Features Improve Differentially Private Image Generation” 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
Fredrik Harder, Milad Jalali Asadabadi, Danica J. Sutherland, Mijung Park
DP-MEPF uses public-data features to generate private image data with differential privacy. It extracts activations from a publicly trained network, adds noise once to a summary of private images, then trains a generator to match that summary. On unlabeled 32×32 CIFAR-10 at ε=2 and δ=10⁻⁵, it reports FID 38.7 versus 62.4 for DP-GAN, not establishing performance under substantial domain shift.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.