Reproduce “Revisiting Differentially Private Hyper-parameter Tuning” 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
Zihang Xiang, Tianhao Wang, Chenglong Wang, Di Wang
Researchers from King Abdullah University of Science and Technology and the University of Virginia developed tighter privacy bounds for differentially private hyper-parameter tuning by employing the f-DP framework. Their method enables exploring 5 to 7 times more hyper-parameter configurations under the same privacy budget, directly enhancing the utility of private machine learning models.
Key insights
- Generic private selection privacy bounds are tight in worst-case scenarios for arbitrary differentially private base algorithms, indicating limitations for black-box improvements.
- Empirical privacy audits showed that in practical settings, hyper-parameter tuning adds little to no additional privacy cost beyond a single base run, but under strong adversarial conditions, additional leakage does occur.
- The f-DP framework provides a finer-grained characterization of privacy for DP-SGD, enabling significantly tighter theoretical bounds for hyper-parameter tuning compared to previous RDP-based approaches.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.