Domain Obfuscation for Efficient Secure Aggregation of Sparse Feature Vectors

Abstract

To collaboratively compute the sum of user private feature vectors, secure aggregation protocols are used to preserve the users’ privacy. The existing protocols are expensive in terms of computational/communication costs, making them impractical where the feature vectors are of size in thousands or millions. Also, they do not provide a possibility to define their own privacy and computation budget. In this work, we propose the Domain Obfuscated Secure Aggregation (DO-SAFE) protocol that allows users to define their privacy budget as per their preference. Our theoretical analysis demonstrates that the proposed protocol can significantly reduce communication and computation costs.