Reproduce “Privacy Accounting for Empirical Denoisers” 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

William Cheng, Dongjun Kim, Mijung Park

The authors present a differentially private diffusion sampler that reuses one data subset through a denoising block. It clips contributions to a training-free empirical denoiser, then uses a public neural denoiser for refinement. On CIFAR-10, a 55% private block reports FID 45.7 with three public steps and 30.4 with nine, at epsilon no greater than 0.3. The guarantee covers this sampler, not privately trained neural denoisers.

Summary by alphaXiv.

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