Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Low-Latency Privacy-Preserving Deep Learning Design via Secure MPC

Ke Lin, Yasir Glani, Ping Luo

TL;DR

The paper tackles the latency bottleneck in privacy-preserving deep learning via secure MPC by introducing a low-latency secret-sharing design that employs multivariate multiplication and communication coalescing to minimize rounds of interaction. It extends Beaver-triple based MPC to $n$-ary operations, develops univariate polynomial and nonlinear function techniques, and demonstrates practical latency gains on varied network settings and model architectures. The key contributions include the theory and implementation of $n$-ary Beaver multiplication, a framework for nonlinear approximations within secret sharing, and a comprehensive experimental evaluation showing $10\sim20\%$ improvements in communication latency with maintained accuracy. These results advance the practicality of MPC-based DL in real-world, bandwidth-constrained environments and offer guidance for integrating these ideas with other privacy-preserving paradigms. The work is relevant for researchers and practitioners aiming to deploy privacy-preserving inference in healthcare, finance, and large-scale vision-language systems.

Abstract

Secure multi-party computation (MPC) facilitates privacy-preserving computation between multiple parties without leaking private information. While most secure deep learning techniques utilize MPC operations to achieve feasible privacy-preserving machine learning on downstream tasks, the overhead of the computation and communication still hampers their practical application. This work proposes a low-latency secret-sharing-based MPC design that reduces unnecessary communication rounds during the execution of MPC protocols. We also present a method for improving the computation of commonly used nonlinear functions in deep learning by integrating multivariate multiplication and coalescing different packets into one to maximize network utilization. Our experimental results indicate that our method is effective in a variety of settings, with a speedup in communication latency of $10\sim20\%$.

Low-Latency Privacy-Preserving Deep Learning Design via Secure MPC

TL;DR

The paper tackles the latency bottleneck in privacy-preserving deep learning via secure MPC by introducing a low-latency secret-sharing design that employs multivariate multiplication and communication coalescing to minimize rounds of interaction. It extends Beaver-triple based MPC to -ary operations, develops univariate polynomial and nonlinear function techniques, and demonstrates practical latency gains on varied network settings and model architectures. The key contributions include the theory and implementation of -ary Beaver multiplication, a framework for nonlinear approximations within secret sharing, and a comprehensive experimental evaluation showing improvements in communication latency with maintained accuracy. These results advance the practicality of MPC-based DL in real-world, bandwidth-constrained environments and offer guidance for integrating these ideas with other privacy-preserving paradigms. The work is relevant for researchers and practitioners aiming to deploy privacy-preserving inference in healthcare, finance, and large-scale vision-language systems.

Abstract

Secure multi-party computation (MPC) facilitates privacy-preserving computation between multiple parties without leaking private information. While most secure deep learning techniques utilize MPC operations to achieve feasible privacy-preserving machine learning on downstream tasks, the overhead of the computation and communication still hampers their practical application. This work proposes a low-latency secret-sharing-based MPC design that reduces unnecessary communication rounds during the execution of MPC protocols. We also present a method for improving the computation of commonly used nonlinear functions in deep learning by integrating multivariate multiplication and coalescing different packets into one to maximize network utilization. Our experimental results indicate that our method is effective in a variety of settings, with a speedup in communication latency of .
Paper Structure (27 sections, 1 theorem, 7 equations, 3 figures, 2 tables, 3 algorithms)

This paper contains 27 sections, 1 theorem, 7 equations, 3 figures, 2 tables, 3 algorithms.

Table of Contents

  1. Introduction
  2. Background
  3. Arithmetic Secret Sharing Based Scheme
  4. Arithmetic Secret Sharing Based MPC
  5. Addition.
  6. Multiplication.
  7. Linear functions.
  8. Nonlinear functions.
  9. Notations
  10. Related Work
  11. Methodology
  12. Multivariate Multiplication
  13. Univariate Polynomials
  14. Nonlinear Approximations
  15. Exponentiation.
  16. ...and 12 more sections

Key Result

Theorem 1

Let $\{x_i'\}$ and $\{x_i"\}$ be random values. The distribution of the view of each party is identical when $x_i=x_i'$ or $x_i=x_i"$.

Figures (3)

  • Figure 1: Communication percentage of different models.
  • Figure 2: Transimission data and latency of naïve and our proposed methods when a different number of parties are involved. The experiment is conducted using the LeNet model on CIFAR-10 using a medium latency network.
  • Figure 3: Latency of naïve and our proposed methods concerning different network bandwidth. The experiment is conducted using the LeNet model on CIFAR-10 using a medium latency network.

Theorems & Definitions (1)

  • Theorem 1