Counting on Dummies: Privacy Amplification with Observable Counts and Truncated Batches

TL;DR

This paper introduces a simple mechanism using dummy data points to privatize batch sizes in DP-SGD, enabling privacy amplification under truncation and observable batch counts. It provides tight privacy accounting for both Poisson sampling and random allocation, reducing the expected number of dummy points needed compared to direct noise-based approaches. Empirical results show improved privacy-efficiency tradeoffs over prior methods.

Why It Matters

In real-world DP-SGD deployments, batch size variability and truncation due to system constraints can undermine privacy guarantees. This work enables practical, privacy-preserving training in distributed settings where batch sizes may be exposed, without sacrificing efficiency or accuracy.

Key Findings

• Dummy points effectively privatize batch sizes even when batch sizes are observable or truncated.
• The proposed mechanism achieves tighter privacy accounting than direct noise-based approaches for truncated batches.
• Empirical results demonstrate superior privacy-efficiency tradeoffs compared to recent methods (Ganesh, 2025).