Preventing Adversarial Use of Datasets through Fair Core-Set Construction

TL;DR

This paper introduces a method for constructing a core-set of datasets that maintains performance on desired tasks while reducing the risk of adversarial use, enhancing privacy and security.

Contribution

The paper presents novel core-set construction techniques for both linear models and neural networks to improve dataset privacy against adversarial tasks.

Findings

Core-sets enable strong primary task performance.

Core-sets reduce effectiveness of unwanted tasks.

Methods are effective on various data types.

Abstract

We propose improving the privacy properties of a dataset by publishing only a strategically chosen "core-set" of the data containing a subset of the instances. The core-set allows strong performance on primary tasks, but forces poor performance on unwanted tasks. We give methods for both linear models and neural networks and demonstrate their efficacy on data.