Reproduce “DP-MERF: Differentially Private Mean Embeddings with Random Features for Practical Privacy-Preserving Data 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
Frederik Harder, Kamil Adamczewski, Mijung Park
DP-MERF trains a generator to match a noisy mean embedding released from sensitive data, rather than repeatedly accessing that data during training. It approximates the embedding with random Fourier features and bounds replacement sensitivity by 2/m. On MNIST, classifiers trained on its synthetic data reached 0.61 accuracy at (0.2, 10⁻⁵)-DP, versus 0.53 for GS-WGAN at ε=10; this comparison uses unequal privacy budgets.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.