Paper 2023/1890

Aegis: A Lightning Fast Privacy-preserving Machine Learning Platform against Malicious Adversaries

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, e.g., medical care and finance. In this work, we introduce $\mathsf{Aegis}$~-- a high-performance PPML platform built on top of a maliciously secure 3-PC framework over ring $\mathbb{Z}_{2^\ell}$. In particular, we propose a novel 2-round secure comparison (a.k.a., sign bit extraction) protocol in the preprocessing model. The communication of its semi-honest version is only 25% of the state-of-the-art (SOTA) constant-round semi-honest comparison protocol by Zhou et al.(S&P 2023); both communication and round complexity of its malicious version are approximately 50% of the SOTA (BLAZE) by Patra and Suresh (NDSS 2020), for $\ell=64$. Moreover, the communication of our maliciously secure inner product protocol is merely $3\ell$ bits, reducing 50% from the SOTA (Swift) by Koti et al. (USENIX 2021). Finally, the resulting ReLU and MaxPool PPML protocols outperform the SOTA by $4\times$ in the semi-honest setting and $10\times$ in the malicious setting, respectively.