Skip to content

GAZELLE: A Low Latency Framework for Secure Neural Network Inference

Authors: Chiraag Juvekar; Vinod Vaikuntanathan; Anantha Chandrakasan
Year: 2018
Venue: USENIX Security 2018

type: system collection: catalogue status: auto

Links: arXiv:1801.05507

Summary

Auto-extracted summary (unreviewed)

Gazelle is a two-party secure CNN inference framework that combines a fast packed additively homomorphic encryption (PAHE) library (SIMD add, SIMD multiply, ciphertext permutation) with garbled circuits, using homomorphic linear-algebra kernels for matrix-vector products and convolutions plus encryption-switching protocols to move between HE and garbled-circuit encodings. On benchmark networks for MNIST and CIFAR-10 it achieves 20x lower online runtime than MiniONN, 30x lower than Chameleon, and about three orders of magnitude lower than the pure-HE CryptoNets.

Abstract

The growing popularity of cloud-based machine learning raises a natural question about the privacy guarantees that can be provided in such a setting. Our work tackles this problem in the context where a client wishes to classify private images using a convolutional neural network (CNN) trained by a server. Our goal is to build efficient protocols whereby the client can acquire the classification result without revealing their input to the server, while guaranteeing the privacy of the server's neural network. To this end, we design Gazelle, a scalable and low-latency system for secure neural network inference, using an intricate combination of homomorphic encryption and traditional two-party computation techniques (such as garbled circuits). Gazelle makes three contributions. First, we design the Gazelle homomorphic encryption library which provides fast algorithms for basic homomorphic operations such as SIMD (single instruction multiple data) addition, SIMD multiplication and ciphertext permutation. Second, we implement the Gazelle homomorphic linear algebra kernels which map neural network layers to optimized homomorphic matrix-vector multiplication and convolution routines. Third, we design optimized encryption switching protocols which seamlessly convert between homomorphic and garbled circuit encodings to enable implementation of complete neural network inference. We evaluate our protocols on benchmark neural networks trained on the MNIST and CIFAR-10 datasets and show that Gazelle outperforms the best existing systems such as MiniONN (ACM CCS 2017) by 20 times and Chameleon (Crypto Eprint 2017/1164) by 30 times in online runtime. Similarly when compared with fully homomorphic approaches like CryptoNets (ICML 2016) we demonstrate three orders of magnitude faster online run-time.
Source: arxiv

Extracted details

Field Value Source Confidence
contribution Gazelle homomorphic encryption library with fast SIMD addition, multiplication and permutation llm 0.80
contribution homomorphic linear algebra kernels for matrix-vector multiplication and convolution llm 0.80
contribution encryption switching protocols between homomorphic and garbled-circuit encodings llm 0.80
contribution parameter selection procedure for 128-bit security llm 0.80
contribution end-to-end evaluation on MNIST and CIFAR-10 networks llm 0.80
schemes packed additively homomorphic encryption (lattice-based) llm 0.80
schemes garbled circuits llm 0.80
libraries Gazelle HE library (own) llm 0.80
models CNN llm 0.80
datasets MNIST llm 0.80
datasets CIFAR-10 llm 0.80
tasks image classification llm 0.80
tasks secure neural network inference llm 0.80
security level 128-bit llm 0.80
threat model two-party client-server; client input and server model both private llm 0.80
interaction interactive (HE+MPC) llm 0.80
code url https://github.com/bristolcrypto/SPDZ-2 (URL found in the full text) regex 0.50
code url https://github.com/cryptobiu/scapi (URL found in the full text) regex 0.50
code url https://github.com/shaih/HElib (URL found in the full text) regex 0.50
code url https://git.njit.edu/ llm 0.80
llm relevance medium llm 0.80
key result online runtime speed-up vs MiniONN: 20x (MNIST / CIFAR-10 benchmark networks) llm 0.80
key result online runtime speed-up vs Chameleon: 30x (MNIST / CIFAR-10 benchmark networks) llm 0.80
key result online runtime speed-up vs CryptoNets: about 1000x (three orders of magnitude) (pure-HE baseline) llm 0.80

System card: Gazelle

Group 5. Interactive MPC and HE+MPC hybrids
Family HE (BFV) + garbled circuits
Arithmetic Exact integers
Interaction Interactive 2PC
Model scale CNNs
Status Foundational
Relation to the SoK Uses HE only for linear layers and garbled circuits for non-linear ones.
Note Template for most hybrid protocols. Requires many communication rounds.
Systems page Gazelle

Keyword mentions

Counts of vocabulary terms in the full text (first page of occurrence in brackets).

  • scheme: Garbled circuits 27 [p.1], Paillier 11 [p.2], MPC 6 [p.2], Secret sharing 5 [p.3], Oblivious transfer 3 [p.2], BFV 2 [p.12], Partially homomorphic encryption 2 [p.2]
  • library: PALISADE 4 [p.16], HElib 2 [p.16], MP-SPDZ 2 [p.2], ABY / ABY3 1 [p.16]
  • model: CNN 20 [p.1]
  • dataset: MNIST 10 [p.1], CIFAR-10 6 [p.1]
  • hardware: Intel Xeon 1 [p.13]
  • technique: Convolution 71 [p.1], Rotations 64 [p.5], SIMD packing 27 [p.1], NTT 25 [p.6], ReLU 17 [p.2], Two-party computation 16 [p.1], Square activation 5 [p.7], Bootstrapping 4 [p.7], Relinearization 3 [p.10], Matrix multiplication 3 [p.14]
  • assumption: Lattice-based 12 [p.2], 128-bit security 2 [p.12]
  • task: Medical / healthcare 7 [p.1]
  • topic: Homomorphic encryption 27 [p.1], Benchmark 24 [p.1], Open source 4 [p.16]

Files

Path (under papers/) Collection Pages Size SHA-256 Note
references/catalogue/5-061_gazelle-a-low-latency-framework-for-secure-neural-network-in.pdf catalogue 17 725.7 KB 63b41ab2c30e

Database key: gazelle-a-low-latency-framework-for-secure-neural-network-inference · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set gazelle-a-low-latency-framework-for-secure-neural-network-inference COLUMN VALUE