Reproduce “Oblivious Sampling Algorithms for Private Data Analysis” 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
Sajin Sasy, Olga Ohrimenko
Researchers introduced data-oblivious sampling algorithms for Trusted Execution Environments (TEEs) to prevent side-channel leakage of sample identities, enabling differentially private analysis to fully leverage privacy amplification. These methods achieve model accuracies comparable to existing approaches while significantly reducing privacy loss, particularly with Poisson sampling.
Key insights
- Memory access patterns in TEEs represent a critical side-channel vulnerability that can compromise privacy amplification during differentially private sampling.
- Data-oblivious algorithmic design is essential to enable the full benefits of differentially private sampling within hardware-secured TEEs by preventing leakage of sample identities.
- The choice between shuffling, Poisson sampling, and sampling without replacement for mini-batch generation does not substantially impact neural network model accuracy in both non-private and differentially private training contexts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.