# Gazelle: A Low Latency Framework for Secure Neural Network Inference

Source: [Gazelle: A Low Latency Framework for Secure Neural Network Inference](<https://www.sota2.com/research/paper/68fa63f51e209131cea01daa>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2025-10-25T20:54:10.749Z

Authors: Chiraag Juvekar, Vinod Vaikuntanathan, Anantha Chandrakasan

Published: 2018-01-16T23:58:47.000Z

## 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.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Sparse Matrix-Vector Multiplication on SuiteSparse arc130 (test)](<https://www.sota2.com/research/sota/sparse-matrix-vector-multiplication-on-suitesparse-arc130-test.md>): Sparse Matrix-Vector Multiplication; SuiteSparse arc130 (test); Time (s): 7.52.

- [Sparse Matrix-Vector Multiplication on SuiteSparse stat96 v5 (test)](<https://www.sota2.com/research/sota/sparse-matrix-vector-multiplication-on-suitesparse-stat96-v5-test.md>): Sparse Matrix-Vector Multiplication; SuiteSparse stat96 v5 (test); Time (s): 5465.94.

- [Sparse Matrix-Vector Multiplication on SuiteSparse M80PI\_n1 (test)](<https://www.sota2.com/research/sota/sparse-matrix-vector-multiplication-on-suitesparse-m80pi_n1-test.md>): Sparse Matrix-Vector Multiplication; SuiteSparse M80PI\_n1 (test); Time (s): 294.71.

- [Sparse Matrix-Vector Multiplication on SuiteSparse mycielskian13 (test)](<https://www.sota2.com/research/sota/sparse-matrix-vector-multiplication-on-suitesparse-mycielskian13-test.md>): Sparse Matrix-Vector Multiplication; SuiteSparse mycielskian13 (test); Execution Time (s): 1030.4.

- [Sparse Matrix-Vector Multiplication on SuiteSparse nemeth24 (test)](<https://www.sota2.com/research/sota/sparse-matrix-vector-multiplication-on-suitesparse-nemeth24-test.md>): Sparse Matrix-Vector Multiplication; SuiteSparse nemeth24 (test); Execution Time (s): 1859.81.

- [Sparse Matrix-Vector Multiplication on SuiteSparse as-caida (test)](<https://www.sota2.com/research/sota/sparse-matrix-vector-multiplication-on-suitesparse-as-caida-test.md>): Sparse Matrix-Vector Multiplication; SuiteSparse as-caida (test); Execution Time (s): 18750.38.

- [Memory Evaluation for Sparse Matrix-Vector multiplication on stat96 v5](<https://www.sota2.com/research/sota/memory-evaluation-for-sparse-matrix-vector-multiplication-on-stat96-v5.md>): Sparse Matrix-Vector Multiplication; stat96 v5; Memory (MB): 4916.19.

- [Memory Evaluation for Sparse Matrix-Vector multiplication on nemeth 24](<https://www.sota2.com/research/sota/memory-evaluation-for-sparse-matrix-vector-multiplication-on-nemeth-24.md>): Sparse Matrix-Vector Multiplication; nemeth 24; Memory (MB): 10924.3.

- [Sparse Matrix-Vector multiplication on msc10848](<https://www.sota2.com/research/sota/sparse-matrix-vector-multiplication-on-msc10848.md>): Sparse Matrix-Vector Multiplication; msc10848; Memory (MB): 7051.2.

- [Sparse Matrix-Vector multiplication on as-caida](<https://www.sota2.com/research/sota/sparse-matrix-vector-multiplication-on-as-caida.md>): Sparse Matrix-Vector Multiplication; as-caida; Memory Consumption (MB): 30425.07.

- [Sparse Matrix-Vector multiplication on ship 001 (Memory Metrics)](<https://www.sota2.com/research/sota/sparse-matrix-vector-multiplication-on-ship-001-memory-metrics.md>): Sparse Matrix-Vector Multiplication; ship 001; Memory (MB): 40130.06.

- [Sparse Matrix-Vector multiplication on cca](<https://www.sota2.com/research/sota/sparse-matrix-vector-multiplication-on-cca.md>): Sparse Matrix-Vector Multiplication; cca; Memory (MB): 56485.48.

- [Sparse Matrix-Vector multiplication on p2p-Gnutella 31](<https://www.sota2.com/research/sota/sparse-matrix-vector-multiplication-on-p2p-gnutella-31.md>): Sparse Matrix-Vector Multiplication; p2p-Gnutella 31; Memory (MB): 71923.83.

- [Sparse Matrix-Vector Multiplication on SuiteSparse ship\_001 (test)](<https://www.sota2.com/research/sota/sparse-matrix-vector-multiplication-on-suitesparse-ship_001-test.md>): Sparse Matrix-Vector Multiplication; SuiteSparse ship\_001 (test); Execution Time (s): 21398.58.

- [Sparse Matrix-Vector Multiplication on SuiteSparse cca (test)](<https://www.sota2.com/research/sota/sparse-matrix-vector-multiplication-on-suitesparse-cca-test.md>): Sparse Matrix-Vector Multiplication; SuiteSparse cca (test); Time (s): 17686.71.

- [Sparse Matrix-Vector Multiplication on SuiteSparse p2p-Gnutella31 (test)](<https://www.sota2.com/research/sota/sparse-matrix-vector-multiplication-on-suitesparse-p2p-gnutella31-test.md>): Sparse Matrix-Vector Multiplication; SuiteSparse p2p-Gnutella31 (test); Execution Time (s): 74553.33.

- [Sparse Matrix-Vector Multiplication on SuiteSparse msc10848 (test)](<https://www.sota2.com/research/sota/sparse-matrix-vector-multiplication-on-suitesparse-msc10848-test.md>): Sparse Matrix-Vector Multiplication; SuiteSparse msc10848 (test); Execution Time (s): 56198.48.

- [Image Classification on CIFAR-100 (Communication and Latency Evaluation)](<https://www.sota2.com/research/sota/image-classification-on-cifar-100-communication-and-latency-evaluation.md>): Image Classification; CIFAR-100; Communication (GB): 8.9.

- [Sparse Matrix-Vector multiplication on ARC 130](<https://www.sota2.com/research/sota/sparse-matrix-vector-multiplication-on-arc-130.md>): Sparse Matrix-Vector Multiplication; ARC 130; Memory Consumption (MB): 317.16.

- [Sparse Matrix-Vector multiplication on M80PI n1](<https://www.sota2.com/research/sota/sparse-matrix-vector-multiplication-on-m80pi-n1.md>): Sparse Matrix-Vector Multiplication; M80PI n1; Memory (MB): 4662.45.

- [Sparse Matrix-Vector multiplication on mycielskian13](<https://www.sota2.com/research/sota/sparse-matrix-vector-multiplication-on-mycielskian13.md>): Sparse Matrix-Vector Multiplication; mycielskian13; Memory (MB): 7059.54.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/1801.05507>)

[PDF](<https://arxiv.org/pdf/1801.05507.pdf>)
