From the source
Apple researchers recast Private Evolution, a differentially private algorithm for synthetic data generation, as generative model-augmented Wasserstein learning, arguing that accounting for the generator yields much better performance bounds than worst-case analyses predict.
- The paper shows that if the generator produces samples in the same low-dimensional space as the distribution, sample complexity depends on intrinsic rather than ambient dimension.
- The authors also report that standard variants of Private Evolution can fail to converge on simple well-clustered instances, and propose a geometry-aware version with provable convergence on such instances.
- Experimentally, they say the new algorithm is competitive with standard baselines and can improve recall.
- The work was done while one author was at Apple, with a co-author from Yale University.
