Dummy items added to each batch to protect privacy during training
A privacy trick adds dummy items to each batch to hide true batch sizes.
- Publication
- arXiv
- Stage
- Preprint
- What we read
- Summary of the paper
- Authors
- Vitaly Feldman, Audra McMillan, Kunal Talwar
- Universities and research institutions
- Apple
Institution metadata: Original paper ↗
What they did and found
In large-scale machine learning with differential privacy, batch sizes can reveal data. The authors suggest adding a random number of dummy items to each batch before processing. This masking keeps privacy intact even if batches are later shortened.
Why it matters
This provides a practical way to run privacy-friendly training in real systems where batch sizes are limited while keeping privacy guarantees.
The idea adds fake items to each batch so outsiders can’t tell how many real examples exist. Training uses the larger batch, with privacy math limiting what counts may reveal.
In practice, extra items cost more compute and memory. Teams should test with their data and privacy goals to balance costs and protection.
What remains uncertain
The observed privacy gains depend on batch construction and data flow; effectiveness may vary by system design.
Original sources · 1
- Counting on Dummies: Privacy Amplification with Observable Counts and Truncated Batches ↗arXiv · 2026-10-30
Check the original paper for its authors, methods, version and access terms.
What is your take?
Ask a question, add useful context or share a different perspective. Keep the conversation respectful and grounded.