Towards Differential Relational Privacy and its use in Question Answering

TL;DR

This paper introduces Relational Memorization and Relational Privacy to understand and control how models memorize relations between entities, aiming to improve privacy in question answering systems without sacrificing learning effectiveness.

Contribution

It formalizes Relational Privacy and Differential Relational Privacy, providing new tools to quantify and bound relational memorization in models, especially for question answering tasks.

Findings

Bounding relational memorization does not hinder learning of general concepts.

Relational Privacy can be formally defined and measured in models.

Experiments demonstrate the applicability of these concepts in large-scale question answering models.

Abstract

Memorization of the relation between entities in a dataset can lead to privacy issues when using a trained model for question answering. We introduce Relational Memorization (RM) to understand, quantify and control this phenomenon. While bounding general memorization can have detrimental effects on the performance of a trained model, bounding RM does not prevent effective learning. The difference is most pronounced when the data distribution is long-tailed, with many queries having only few training examples: Impeding general memorization prevents effective learning, while impeding only relational memorization still allows learning general properties of the underlying concepts. We formalize the notion of Relational Privacy (RP) and, inspired by Differential Privacy (DP), we provide a possible definition of Differential Relational Privacy (DrP). These notions can be used to describe and compute bounds on the amount of RM in a trained model. We illustrate Relational Privacy concepts in experiments with large-scale models for Question Answering.