Efficient 3PC for Binary Circuits with Application to Maliciously-Secure DNN Inference

Authors: 

Yun Li, Tsinghua University, Ant Group; Yufei Duan, Tsinghua University; Zhicong Huang, Alibaba Group; Cheng Hong, Ant Group; Chao Zhang and Yifan Song, Tsinghua University

Abstract: 

In this work, we focus on maliciously secure 3PC for binary circuits with honest majority. While the state-of-the-art (Boyle et al. CCS 2019) has already achieved the same amortized communication as the best-known semi-honest protocol (Araki et al. CCS 2016), they suffer from a large computation overhead: when comparing with the best-known implementation result (Furukawa et al. Eurocrypt 2017) which requires 9× communication cost of Araki et al., the protocol by Boyle et al. is around 4.5× slower than that of Furukawa et al.

In this paper, we design a maliciously secure 3PC protocol that matches the same communication as Araki et al. with comparable concrete efficiency as Furukawa et al. To obtain our result, we manage to apply the distributed zero-knowledge proofs (Boneh et al. Crypto 2019) for verifying computations over Z2 by using prime fields and explore the algebraic structure of prime fields to make the computation of our protocol friendly for native CPU computation.

Experiment results show that our protocol is around 3.5× faster for AES circuits than Boyle et al. We also applied our protocol to the binary part (e.g. comparison and truncation) of secure deep neural network inference, and results show that we could reduce the time cost of achieving malicious security in the binary part by more than 67%.

Besides our main contribution, we also find a hidden security issue in many of the current probabilistic truncation protocols, which may be of independent interest.

Open Access Media

USENIX is committed to Open Access to the research presented at our events. Papers and proceedings are freely available to everyone once the event begins. Any video, audio, and/or slides that are posted after the event are also free and open to everyone. Support USENIX and our commitment to Open Access.

BibTeX
@inproceedings {291086,
author = {Yun Li and Yufei Duan and Zhicong Huang and Cheng Hong and Chao Zhang and Yifan Song},
title = {Efficient {3PC} for Binary Circuits with Application to {Maliciously-Secure} {DNN} Inference},
booktitle = {32nd USENIX Security Symposium (USENIX Security 23)},
year = {2023},
isbn = {978-1-939133-37-3},
address = {Anaheim, CA},
pages = {5377--5394},
url = {https://www.usenix.org/conference/usenixsecurity23/presentation/li-yun},
publisher = {USENIX Association},
month = aug
}

Presentation Video