Counting on Dummies enables privacy amplification with truncated batches
Original title: Counting on Dummies: Privacy Amplification with Observable Counts and Truncated Batches
Mechanism uses dummy points to privatize batch sizes in DP-SGD while preserving privacy under truncation and subsampling.
DP-SGD batch sizes are observable or truncated in practical deployments. The paper introduces a simple dummy-point mechanism that privatizes batch sizes and maintains privacy guarantees for both Poisson sampling and random allocation. It supports efficient accounting and yields better privacy-efficiency tradeoffs than direct truncation accounting.
- Uses random dummy points to privatize batch sizes
- Maintains privacy under truncation
- Analysis covers Poisson sampling and random allocation
- Efficient and relatively tight privacy accounting
- Empirically better tradeoffs than direct truncation
From the original article
1 min read at arxiv.orgIn many practical deployments of Differentially Private Stochastic Gradient Descent (DP-SGD), the server processes data in batches. Randomly sampling these batches allows for better privacy bounds via privacy amplification results. Most sampling schemes with provable privacy amplification, including Poisson sampling and random allocation, result in variable batch sizes.
By Vitaly Feldman · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.