Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PrivQuant: Communication-Efficient Private Inference with Quantized Network/Protocol Co-Optimization

Tianshi Xu, Shuzhang Zhong, Wenxuan Zeng, Runsheng Wang, Meng Li

TL;DR

This paper proposes PrivQuant, a framework that jointly optimizes the 2PC-based quantized inference protocols and the network quantization algorithm, enabling communication-efficient private inference and develops a communicationaware mixed precision quantization algorithm to improve the inference efficiency while maintaining high accuracy.

Abstract

Private deep neural network (DNN) inference based on secure two-party computation (2PC) enables secure privacy protection for both the server and the client. However, existing secure 2PC frameworks suffer from a high inference latency due to enormous communication. As the communication of both linear and non-linear DNN layers reduces with the bit widths of weight and activation, in this paper, we propose PrivQuant, a framework that jointly optimizes the 2PC-based quantized inference protocols and the network quantization algorithm, enabling communication-efficient private inference. PrivQuant proposes DNN architecture-aware optimizations for the 2PC protocols for communication-intensive quantized operators and conducts graph-level operator fusion for communication reduction. Moreover, PrivQuant also develops a communication-aware mixed precision quantization algorithm to improve inference efficiency while maintaining high accuracy. The network/protocol co-optimization enables PrivQuant to outperform prior-art 2PC frameworks. With extensive experiments, we demonstrate PrivQuant reduces communication by $11\times, 2.5\times \mathrm{and}~ 2.8\times$, which results in $8.7\times, 1.8\times ~ \mathrm{and}~ 2.4\times$ latency reduction compared with SiRNN, COINN, and CoPriv, respectively.

PrivQuant: Communication-Efficient Private Inference with Quantized Network/Protocol Co-Optimization

TL;DR

This paper proposes PrivQuant, a framework that jointly optimizes the 2PC-based quantized inference protocols and the network quantization algorithm, enabling communication-efficient private inference and develops a communicationaware mixed precision quantization algorithm to improve the inference efficiency while maintaining high accuracy.

Abstract

Private deep neural network (DNN) inference based on secure two-party computation (2PC) enables secure privacy protection for both the server and the client. However, existing secure 2PC frameworks suffer from a high inference latency due to enormous communication. As the communication of both linear and non-linear DNN layers reduces with the bit widths of weight and activation, in this paper, we propose PrivQuant, a framework that jointly optimizes the 2PC-based quantized inference protocols and the network quantization algorithm, enabling communication-efficient private inference. PrivQuant proposes DNN architecture-aware optimizations for the 2PC protocols for communication-intensive quantized operators and conducts graph-level operator fusion for communication reduction. Moreover, PrivQuant also develops a communication-aware mixed precision quantization algorithm to improve inference efficiency while maintaining high accuracy. The network/protocol co-optimization enables PrivQuant to outperform prior-art 2PC frameworks. With extensive experiments, we demonstrate PrivQuant reduces communication by , which results in latency reduction compared with SiRNN, COINN, and CoPriv, respectively.

Paper Structure

This paper contains 32 sections, 3 theorems, 8 equations, 10 figures, 12 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Network Quantization for 2PC Inference
  4. Notations
  5. Secret Share (SS)
  6. Oblivious Transfer (OT)
  7. Underlying Protocols
  8. Related works
  9. Threat Model and Security of PrivQuant
  10. Motivation
  11. Efficient Mixed Bit-width 2PC Protocols
  12. Operator Level Protocol Optimization
  13. DNN Architecture-Aware Convolution Protocol
  14. Simplified Residual Protocol
  15. Graph Level Protocol Optimization
  16. ...and 17 more sections

Key Result

Proposition 4.1

For a given $\langle x \rangle^{(l_1)}$, $\Pi_{\mathrm{Trunc}}^{l_1,l_2}(\langle x \rangle^{(l_1)})$ can be decomposed into $\Pi_{\mathrm{TR}}^{l_1,l_2}$ followed by $\Pi_{\mathrm{Ext}}^{l_1-l_2,l_1}$ as The decomposition reduce the communication from $\mathrm{O}(\lambda (l_1 + 3))$ to $\mathrm{O}(\lambda (l_1 + 2))$.

Figures (10)

  • Figure 1: Profile the ResNet50 building block with representative 2PC protocols, i.e., CrypTFlow2 (first column) and SiRNN (other columns): the scaling and breakdown of (a) total communication and (b) online communication with different bit-widths of weight and activation.
  • Figure 2: Detailed protocols for one convolution with residual connection in (a) CrypTFlow2 and (b) SiRNN. The bit extension, truncation, and re-quantization are required in SiRNN to align the bit-widths and scales of quantized operands.
  • Figure 3: Overview of PrivQuant.
  • Figure 4: An illustration of OT-based matrix multiplication protocol which extends $X$ and chooses the client to be the sender. We omit $\langle\cdot \rangle$ for simplicity.
  • Figure 5: (a) The baseline protocol for the residual addition in SiRNN; and (b) our proposed simplified protocol. The $l\_,s\_$ means the bit-width and scale of the activations.
  • ...and 5 more figures

Theorems & Definitions (3)

  • Proposition 4.1
  • Proposition 4.2
  • Proposition 4.3