BARC talk by Jacob John Imola
Tuesday, February 18, 2025, Jacob Imola, Postdoc here with us at BARC, KU, will give a talk on "Metric Differential Privacy at the User-Level Via the Earth-Mover’s Distance".
Abstract:
Metric Differential Privacy is a fine-grained privacy definition that can offer improved privacy for data generated from realistic priors in a metric space. At the same time, in modern data applications it is more realistic that a user contributes a set of data points, rather than a single point as many prior works in privacy assume. This is known as the user-level setting, and providing privacy for such data is a recent trend in privacy research.
In this work, we explore how to provide a metric DP guarantee in the user-level setting. We propose using the Earth-Mover’s Distance (EMD) between user data as our privacy metric, as it provides similar appealing semantics as the single-item setting. We then show how to bound the EMD sensitivity of the average value of a function over a dataset, providing a general EMD-DP mechanism framework. We provide another general framework for the pointwise release of a dataset histogram, in which we also prove a generalisation of privacy amplification by shuffling which may be of independent interest. As our last tool, we show how to bound the number of data points collected from each user while providing EMD-DP. This facilitates mechanism design by enabling simpler privacy proofs and more black-box analysis.
Bio:
Jacob John Imola is a postdoctoral researcher in computer science with BARC at the University of Copenhagen hosted by Prof. Rasmus Pagh. From 2018 to 2023, Jacob was a PhD student at UCSD where he was advised by Prof. Kamalika Chaudhuri. His primary focus is differential privacy, where he is intrigued by how to overcome information and algorithmic challenges introduced by privacy. Jacob's work tends to be theoretical in nature, but he is also interested in using interdisciplinary approaches, such as multi-party computation and formal methods, to better implement privacy in practice.