Defensive Collaborative Learning: Protecting Objective Privacy in Data Sharing

Cynthia Huang, Pascal Poupart

Abstract: In collaborative machine learning, protecting proprietary information and safeguarding competitive advantages are crucial for participating organizations. This necessitates the development of algorithms that target a general notion of privacy defined by the data owner: objective privacy. In this paper, we formalize the idea of objective privacy as the protection of private value propositions characterized by predictive functions. We propose Defensive Collaborative Learning (DCL), where participants share data collaboratively while safeguarding their objective privacy. Formulating a min-max optimization problem that trades off utility and privacy protection, we propose algorithms that leverage mutual information backpropagation in both decentralized and centralized settings. Empirical studies show that the proposed algorithms protect objective privacy while enabling data sharing.