Performance Analysis of Secure Floating-Point Sums and Dot Products

Abstract

Privacy-preserving collaborative applications enable groups of parties to run joint computations with private inputs, using cryptographic protocols that protect data privacy through-out the computation. The developers of these applications need a collection of protocols that provide efficient secure computation with all basic data types and secure protocol composition. We focus in this paper on protocols for multioperand floating-point addition and dot products, to provide a more comprehensive analysis of their complexity and performance, as well as improved solutions. These protocols are part of a framework that supports all basic data types. The analysis shows that adding protocols optimized for multiple inputs can substantially improve the performance of secure floating-point arithmetic.