Privacy in Collaborative Deep Learning Systems: A Taxonomy and Archetypes

Abstract

Collaborative deep learning (CDL) systems can help overcome resource constraints in neural network training and use by enabling multiple parties to share resources. Despite their promising benefits, they also raise privacy and confidentiality concerns. To address these concerns, CDL systems integrate privacy-preserving techniques (PPTs), such as differential privacy, homomorphic encryption, secure multi-party computation, and trusted execution environments. Researchers and practitioners face a plethora of CDL system designs that are scattered across research streams. To support consistent description and informed selection of CDL systems, we present a privacy-focused taxonomy and describe four CDL system archetypes and their use of PPTs.