Paper 2025/919

Rep3 Reloaded: On the Cost of Function-Dependent Preprocessing in Semi-Honest 3PC with Honest Majority

Abstract

Rep3 denotes the implementation of semi-honest three-party computation with an honest majority in MP-SPDZ (CCS'20). It uses replicated secret sharing with one message per multiplication and party as proposed by Araki et al. (CCS'16). This approach is rivaled by Astra (CCSW'19) and Trio (PETS'25), which use function-dependent preprocessing. The latter is more involved than, e.g., Beaver triples which can be used as a commodity. In this work, we present a full implementation of Astra and Trio in MP-SPDZ, and we evaluate the costs of the different approaches. We show the equivalence of the schemes, which implies that a protocol in any of the schemes can be translated to one in another with the same overall communication cost. We also present an improvement to two important building blocks for privacy-preserving computation, namely secure comparison and probabilistic truncation used in fixed-point arithmetic. To evaluate our implementation, we have benchmarked machine learning training and inference in all three schemes, improving on Keller and Sun (ICML'22) by over 30%. Our implementation also highlights the large storage requirements of function-dependent preprocessing as it runs the two phases separately. To the best of our knowledge, this is the first implementation to do so.