SecureBiNN: 3-Party Secure Computation for Binarized Neural Network Inference

  • Wenxing Zhu
  • , Mengqi Wei
  • , Xiangxue Li*
  • , Qiang Li
  • *Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

9 Scopus citations

Abstract

The paper proposes SecureBiNN, a novel three-party secure computation framework for evaluating privacy-preserving binarized neural network (BiNN) in semi-honest adversary setting. In SecureBiNN, three participants hold input data and model parameters in secret sharing form, and execute secure computations to obtain secret shares of prediction result without disclosing their input data, model parameters and the prediction result. SecureBiNN performs linear operations in a computation-efficient and communication-free way. For non-linear operations, we provide novel secure methods for evaluating activation function, maxpooling layers, and batch normalization layers in BiNN. Communication overhead is significantly minimized comparing to previous work like XONN and Falcon. We implement SecureBiNN with tensorflow and the experiments show that using the Fitnet structure, SecureBiNN achieves on CIFAR-10 dataset an accuracy of 81.5%, with communication cost of 16.609MB and runtime of 0.527s/3.447s in the LAN/WAN settings. More evaluations on real-world datasets are also performed and other concrete comparisons with state-of-the-art are presented as well.

Original languageEnglish
Title of host publicationComputer Security – ESORICS 2022 - 27th European Symposium on Research in Computer Security, Proceedings
EditorsVijayalakshmi Atluri, Roberto Di Pietro, Christian D. Jensen, Weizhi Meng
PublisherSpringer Science and Business Media Deutschland GmbH
Pages275-294
Number of pages20
ISBN (Print)9783031171420
DOIs
StatePublished - 2022
Event27th European Symposium on Research in Computer Security, ESORICS 2022 - Hybrid, Copenhagen, Denmark
Duration: 26 Sep 202230 Sep 2022

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume13556 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference27th European Symposium on Research in Computer Security, ESORICS 2022
Country/TerritoryDenmark
CityHybrid, Copenhagen
Period26/09/2230/09/22

Keywords

  • Binarized neural network
  • Privacy-preserving machine learning
  • Secure multi-party computation

Fingerprint

Dive into the research topics of 'SecureBiNN: 3-Party Secure Computation for Binarized Neural Network Inference'. Together they form a unique fingerprint.

Cite this