Optimization and Tradeoffs in Secure Floating-Point Computation
Octavian Catrina · 2019
A broad range of privacy preserving collaborative applications require efficient and accurate secure computation with real numbers (e.g., statistical analysis, data mining, and optimizations). Secure fixed-point arithmetic offers efficient solutions for certain tasks, but many applications require the dynamic range and accuracy provided by floating-point arithmetic. We focus in this paper on secure multi-operand multiplication and related tasks. For these tasks, secure floating-point arithmetic can offer better tradeoffs between performance and accuracy than fixed-point arithmetic. We present optimized protocols for evaluating products, powers, and polynomials, that offer important performance gains with respect to generic constructions. These protocols are part of a more comprehensive framework for secure multiparty computation with real numbers, constructed using a small collection of building blocks based on Shamir secret sharing. With these additional protocols, the framework offers a better foundation for secure evaluation of mathematical functions (e.g., by polynomial approximation).