Paper 2023/1890 Original: eprint.iacr.org/2023/1890 ↗

Lightning Fast Secure Comparison for 3PC PPML

Abstract

Privacy-preserving machine learning (PPML) techniques have gained significant popularity in the past years. Those protocols have been widely adopted in many real-world security-sensitive machine-learning scenarios. Secure comparison is one of the most important non-linear operations in PPML. In this work, we focus on maliciously secure comparison in the 3-party MPC over ring $\mathbb{Z}_{2^\ell}$ setting. In particular, we propose a novel constant round sign-bit extraction protocol in the preprocessing model. The communication of its semi-honest version is only 12.5% of the state-of-the-art (SOTA) constant-round semi-honest comparison protocol by Zhou et al.(Bicoptor, S&P 2023); communication complexity of its malicious version are approximately 25% of the SOTA by Patra and Suresh (BLAZE, NDSS 2020), for $\ell=64$. Finally, the resulting ReLU protocol outperforms the SOTA secure ReLU evaluation solution (Bicoptor, S&P 2023) by $6\times$ in the semi-honest setting and $20\times$ in the malicious setting, respectively.