Paper 2020/1013

A Study on Privacy-Preserving GRU Inference Framework

Shou-Ching Hsiao, Zi-Yuan Liu, and Raylin Tso

Abstract

Gated Recurrent Unit (GRU) has broad application fields, such as sentiment analysis, speech recognition, malware analysis, and other sequential data processing. For low-cost deployment and efficient machine learning services, a growing number of model owners choose to deploy the trained GRU models through Machine-learning-as-a-service (MLaaS). However, privacy has become a significant concern for both model owners and prediction clients, including model weights privacy, input data privacy, and output results privacy. The privacy leakage may be caused by either external intrusion or insider attacks. To address the above issues, this research designs a framework for privacy-preserving GRU models, which aims for privacy scenarios such as predicting on textual data, network packets, heart rate data, and so on. In consideration of accuracy and efficiency, this research uses additive secret sharing to design the basic operations and gating mechanisms of GRU. The protocols can meet the security requirements of privacy and correctness under the Universal Composability framework with the semi-honest adversary. Additionally, the framework and protocols are realized with a proof-of-concept implementation. The experiment results are presented with respect to time consumption and inference accuracy.

Metadata
Available format(s)
PDF
Category
Cryptographic protocols
Publication info
Preprint. MINOR revision.
Keywords
Privacy-preservingGated Recurrent Unit ModelSecret SharingUniversal Composability Framework
Contact author(s)
zyliu @ cs nccu edu tw
History
2020-08-22: received
Short URL
https://ia.cr/2020/1013
License
Creative Commons Attribution
CC BY

BibTeX

@misc{cryptoeprint:2020/1013,
      author = {Shou-Ching Hsiao and Zi-Yuan Liu and Raylin Tso},
      title = {A Study on Privacy-Preserving GRU Inference Framework},
      howpublished = {Cryptology ePrint Archive, Paper 2020/1013},
      year = {2020},
      note = {\url{https://eprint.iacr.org/2020/1013}},
      url = {https://eprint.iacr.org/2020/1013}
}
Note: In order to protect the privacy of readers, eprint.iacr.org does not use cookies or embedded third party content.