The authors introduce a training-free empirical denoiser that averages over a dataset with a bounded per-sample contribution, clipped so that its sensitivity is exactly 2C/|D|.
Empirical denoiser enables privacy accounting for diffusion sampling
The authors construct a privacy-accountable empirical denoiser with explicit per-step sensitivity, showing that privacy loss concentrates in the critical window where semantic content forms and can be reduced via subsampling and hybrid samplers.
Top University
William Cheng · Mijung Park · Dongjun Kim
University of British Columbia · Stanford University
Research Digest··3 min read
Cheng et al.
§