Cheetah: Lean and Fast Secure Two-Party Deep Neural Network Inference

This repo contains a proof-of-concept implementation for our Cheetah paper.
The codes are still under heavy developments, and should not be used in any security sensitive product.

Repo Directory Description

  • include/ Contains implementation of Cheetah’s linear protocols.
  • SCI/ A fork of CryptFlow2’s SCI library and contains implementation of Cheetah’s non-linear protocols.
  • networks/ Auto-generated cpp programs that evaluate some neural networks.
  • pretrained/ Pretrained neural networks and inputs.
  • patch/ Patches applied to the dependent libraries.
  • credits/ Licenses of the dependencies.
  • scripts/ Helper scripts used to build the programs in this repo.


  • openssl
  • c++ compiler (>= 8.0 for the better performance on AVX512)
  • cmake >= 3.13
  • git
  • make
  • OpenMP (optional, only needed by CryptFlow2 for multi-threading)

Building Dependencies

  • Run bash scripts/ which will build the following dependencies

    • emp-tool We follow the implementation in SCI that using emp-tool for network io and pseudo random generator.
    • emp-ot We use Ferret in emp-ot as our VOLE-style OT.
    • Eigen We use Eigen for tensor operations.
    • SEAL We use SEAL’s implementation for the BFV homomorphic encryption scheme.
    • zstd We use zstd for compressing the ciphertext in SEAL which can be replaced by any other compression library.
    • hexl We need hexl’s AVX512 acceleration for achieving the reported numbers in our paper.
  • The generated objects are placed in the build/deps/ folder.

  • Build has passed on the following setting

    • MacOS 11.6 with clang 13.0.0, Intel Core i5, cmake 3.22.1
    • Red Hat 7.2.0 with gcc 7.2.1, Intel(R) Xeon(R), cmake 3.12.0
    • Ubuntu 18.04 with gcc 7.5.0 Intel(R) Xeon(R), cmake 3.13
    • Ubuntu 20.04 with gcc 9.4.0 Intel(R) Xeon(R), cmake 3.16.3

Building Cheetah and SCI-HE Demo

  • Run bash scripts/ which will build 6 executables in the build/bin folder
    • resnet50-cheetah
    • sqnet-cheetah
    • densenet121-cheetah
    • resnet50-SCI_HE
    • sqnet-SCI_HE
    • densenet121-SCI_HE

Local Demo

  1. On one terminal run bash scripts/ cheetah sqnet. The program will load the pretrained model in the folder pretrained/ which might takes some time when the pretrained model is huge.

  2. On other terminal run bash scripts/ cheetah sqnet. The program will load the prepared input image in the folder pretrained.

    • replace cheetah by SCI_HE to execute the CryptFlow2’s counterpart.
    • replace sqnet by resnet50 to run on the ResNet50 model.

You can change the SERVER_IP and SERVER_PORT defined in the scripts/ to run the demo remotely.
Also, you can use our throttle script to mimic a remote network condition within one Linux machine, see below.

Mimic an WAN setting within LAN on Linux

  • To use the throttle script under scripts/ to limit the network speed and ping latency (require sudo)
  • For example, run sudo scripts/ wan on a Linux OS which will limit the local-loop interface to about 400Mbps bandwidth and 40ms ping latency.
    You can check the ping latency by just ping The bandwidth can be check using extra iperf command.


View Github