Reproduce “Differential Privacy with Random Projections and Sign Random Projections” 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
Ping Li, Xiaoyun Li
Researchers from LinkedIn Ads developed a suite of differentially private (DP) algorithms leveraging random projections (RP) and sign random projections (SignRP), which demonstrably improve data utility compared to direct noise addition on high-dimensional data. Their methods, including a novel 'smooth flipping probability' for SignRP, achieve superior privacy-utility trade-offs for tasks like similarity search and SVM classification, with some achieving over 85% accuracy where baselines reached 60% at relevant privacy budgets.
Key insights
- Optimal noise calibration and choice of projection matrix (e.g., Rademacher) significantly enhance utility in full-precision DP-RP by reducing inherent sensitivity and noise requirements.
- The inherent robustness of SignRP's 'aggregate-and-sign' operation can be leveraged with a 'smooth flipping probability' to achieve pure ε-DP with considerably improved utility.
- Individual Differential Privacy (iDP) combined with SignRP offers exceptional utility at very small ε values, suggesting its suitability for non-interactive data release scenarios where point-wise privacy is sufficient.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.