Tighter Privacy Analysis for Truncated Poisson Sampling

2025-08-20

AI Summary

This work addresses the privacy amplification effect of truncated Poisson sampling—where batches exceeding a capacity threshold are truncated—in differentially private learning. Unlike conventional analyses assuming ideal Poisson sampling, this paper provides the first rigorous probabilistic modeling and privacy analysis for the truncated setting, establishing a theoretical framework aligned with practical training scenarios (e.g., gradient clipping combined with batch truncation). Methodologically, it precisely characterizes the sampling distribution shift induced by truncation and derives a tighter upper bound on the Rényi differential privacy (RDP) loss. The resulting bound strictly improves upon existing RDP bounds under identical privacy budgets. Empirically, experiments on standard benchmarks validate both the tightness of the theoretical bound and the improved model utility—demonstrating significant gains in accuracy or convergence speed without compromising privacy guarantees.

Technology Category

• Machine Learning: Privacy
• Search and Optimization: Learning to Search
• Reasoning under Uncertainty: Stochastic Optimization

Application Category

• Security and Privacy: Large-scale security measurements
• User Modeling, Personalization and Recommendation: User privacy protection in personalized systems
• Responsible Web: Data and user privacy-enhancing technologies for the Web

Abstract

We give a new privacy amplification analysis for truncated Poisson sampling, a Poisson sampling variant that truncates a batch if it exceeds a given maximum batch size.

Problem

• Analyzing privacy amplification for truncated Poisson sampling
• Addressing privacy in Poisson sampling with batch truncation
• Improving privacy analysis for bounded batch size sampling

Innovation

• Truncated Poisson sampling technique
• Privacy amplification analysis method
• Maximum batch size truncation mechanism