A Unified Analysis of Label Inference Attacks

A Google TechTalk, presented by Marika Swanberg, 2024-03-06 ABSTRACT: We investigate the privacy properties of two proposed methods for attribution reporting through the lens of label inference measures. A label inference measure quantifies the relationship between an adversary’s prior and posterior knowledge about the private labels after a private data release. In this talk, I will discuss new ways that we model our adversary, and present two measures that capture different notions of label inference success. I will also present some empirical and theoretical findings that help guide the discussion around risks and accuracy tradeoffs of potential attribution reporting methods. This is based on research with Andres Muñoz Medina, Travis Dick, Robert Busa-Fekete, Claudio Gentile, and Adam Smith during my Google PhD research internship and subsequent Student Researcher position. Speaker: Marika Swanberg (Boston University)