Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Efficient Skip Connections Realization for Secure Inference on Encrypted Data

Nir Drucker, Itamar Zimerman

TL;DR

This paper shows that by replacing (mid-term) skip connections with Dirac parameterization and shared-source skip connection the authors were able to reduce the skip connections burden for HE-based solutions, achieving x1.3 computing power improvement for the same accuracy.

Abstract

Homomorphic Encryption (HE) is a cryptographic tool that allows performing computation under encryption, which is used by many privacy-preserving machine learning solutions, for example, to perform secure classification. Modern deep learning applications yield good performance for example in image processing tasks benchmarks by including many skip connections. The latter appears to be very costly when attempting to execute model inference under HE. In this paper, we show that by replacing (mid-term) skip connections with (short-term) Dirac parameterization and (long-term) shared-source skip connection we were able to reduce the skip connections burden for HE-based solutions, achieving x1.3 computing power improvement for the same accuracy.

Efficient Skip Connections Realization for Secure Inference on Encrypted Data

TL;DR

This paper shows that by replacing (mid-term) skip connections with Dirac parameterization and shared-source skip connection the authors were able to reduce the skip connections burden for HE-based solutions, achieving x1.3 computing power improvement for the same accuracy.

Abstract

Homomorphic Encryption (HE) is a cryptographic tool that allows performing computation under encryption, which is used by many privacy-preserving machine learning solutions, for example, to perform secure classification. Modern deep learning applications yield good performance for example in image processing tasks benchmarks by including many skip connections. The latter appears to be very costly when attempting to execute model inference under HE. In this paper, we show that by replacing (mid-term) skip connections with (short-term) Dirac parameterization and (long-term) shared-source skip connection we were able to reduce the skip connections burden for HE-based solutions, achieving x1.3 computing power improvement for the same accuracy.
Paper Structure (10 sections, 3 figures, 2 tables)

This paper contains 10 sections, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Our contribution
  3. Organization.
  4. HE
  5. Chain Index and Bootstrapping.
  6. HE Packing.
  7. Skip connections
  8. Handling skip connections in HE
  9. Experiments
  10. Conclusions

Figures (3)

  • Figure 1: An illustration of ResNet50 every layer contains several blocks and there is a skip connection between every block.
  • Figure 2: An illustration of our modified HE-friendly ResNet50. Every layer contains several blocks and there is a shared-source skip connection from the first layer output to the output of the four other layers. Red layers were modified to make the network HE-friendly as in CryptoNets2016.
  • Figure 3: Test accuracy per training epoch of 3 HE-friendly ResNet50 network variants: Reference (blue line), Our modified network (red line), No skip-connection network (green line).