Reproduce “Differentially Private Gradient Flow based on the Sliced Wasserstein Distance” 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

Ilana Sebag, Muni Sreenivas Pydi, Jean-Yves Franceschi, Alain Rakotomamonjy, Mike Gartrell, Jamal Atif, Alexandre Allauzen

The authors propose a differentially private particle-based generative flow that moves samples toward a target distribution. It smooths the one-dimensional projections used in the sliced-Wasserstein comparison with Gaussian noise, while diffusion noise in particle updates further amplifies privacy. On MNIST at ε=5, the resampling flow reports mean FID 117 versus 203 for a generator, in these settings and against this baseline.

Summary by alphaXiv.

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