Reproduce “Differentially Private Mixing of Public Datasets Improves Private Learning” 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
Yufei Chen, Tejumade Afonja, Anvith Thudi, Nicolas Papernot
DP-MixMin learns how to weight public data sources using a small private split, then pretrains on the tailored mixture and privately fine-tunes on the remaining sensitive data. This replaces evaluating candidate mixtures with directly learning their weights. Across tested privacy budgets, it outperforms uniform pretraining and random search on NIH chest X-ray classification and Enron email completion, under stated assumptions and data-allocation differences.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.