Results 1191 - 1200 of 24657

This talk studies privacy-preserving learning when heterogeneity is part of the problem's structure. I will begin with recent work on private personalized learning, which develops a taxonomy of privacy notions for multitask learning and meta-learning, and proves separations between them. These results show that modeling privacy can fundamentally change what we can and cannot do in learning. I will then briefly discuss other forms of heterogeneity that fit within this perspective, particularly settings where privacy requirements differ across the data. This includes feature-specific privacy and heterogeneous local differential privacy, where privacy guarantees vary across coordinates or users. Taken together, these works suggest a broader message: meaningful privacy guarantees must adapt to the heterogeneous structure of modern learning problems. 

We consider the privacy amplification properties of a sampling scheme in which a user’s data is used in k steps chosen randomly and uniformly from a sequence (or set) of t steps. This sampling scheme has been recently applied in the context of differentially private optimization (Chua et al., 2024a; Choquette-Choo et al., 2025) and communication-efficient high-dimensional private aggregation (Asi et al., 2025), where it was shown to have utility advantages over the standard Poisson sampling. Theoretical analyses of this sampling scheme (Feldman & Shenfeld, 2025; Dong et al.,2025) lead to bounds that are close to those of Poisson sampling, yet still have two significant shortcomings. In this work, we demonstrate that the privacy loss distribution (PLD) of random allocation applied to any differentially private algorithm can be computed efficiently. When applied to the Gaussian mechanism, our results demonstrate that random allocation can be used in place of Poisson subsampling with no degradation in resulting privacy guarantees. In addition, our work develops tools for privacy accounting based on approximate stochastic domination of the privacy loss random variable. In particular, we demonstrate how to correctly perform subsampling directly on PLD bounds, enabling accurate privacy accounting for more general algorithms.

Privacy amplification by sampling reduces the noise necessary to achieve a target privacy guarantee when training a model with differential private stochastic gradient descent (DP-SGD), by compounding the randomness in forming batches with the randomness in noise addition in DP-SGD. Historically, the literature on DP-SGD has assumed batches are formed using Poisson sampling, but in practice shuffling-like methods were used, and thus the gains from privacy amplification achieved in the literature could not be achieved in practice. Recently, infrastructure has caught up and privacy amplification by sampling is now feasible in practice, but with a set of restrictions introducing new technical challenges. This talk will cover recent work on these challenges, including (i) handling a need for fixed-size batches due to JIT compilation, (ii) privacy amplification for correlated noise mechanisms, and (iii) privacy guarantees that are robust to side channel information.