The Intel® HE transformer for nGraph™ is a Homomorphic Encryption (HE) backend to the Intel® nGraph Compiler, Intel's graph compiler for Artificial Neural Networks.

Homomorphic encryption is a form of encryption that allows computation on encrypted data, and is an attractive remedy to increasing concerns about data privacy in the field of machine learning. For more information, see our original paper. Our updated paper showcases many of the recent advances in he-transformer.

This project is meant as a proof-of-concept to demonstrate the feasibility of HE on local machines. The goal is to measure performance of various HE schemes for deep learning. This is not intended to be a production-ready product, but rather a research tool.

Currently, we support the CKKS encryption scheme, implemented by the Simple Encrypted Arithmetic Library (SEAL) from Microsoft Research.

To help compute non-polynomial activiations, we additionally integrate with the ABY multi-party computation library. See also the NDSS 2015 paper introducing ABY. For more details about our integration with ABY, please refer to our ARES 2020 paper.

We also integrate with the Intel® nGraph™ Compiler and runtime engine for TensorFlow to allow users to run inference on trained neural networks through Tensorflow.

The examples folder contains a deep learning example which depends on the Intel® nGraph™ Compiler and runtime engine for TensorFlow.

• Operating system: Ubuntu 16.04, Ubuntu 18.04.
• CMake >= 3.12
• Compiler: g++ version >= 6.0, clang >= 5.0 (with ABY g++ version >= 8.4)
• OpenMP is strongly suggested, though not strictly necessary. You may experience slow runtimes without OpenMP
• python3 and pip3
• virtualenv v16.1.0
• bazel v0.25.2

For a full list of dependencies, see the docker containers, which build he-transformer on a reference OS.

• nGraph - v0.28.0-rc.1
• nGraph-bridge - v0.22.0-rc3
• SEAL - v3.4.5
• TensorFlow - v1.14.0
• Boost v1.69
• Google protobuf v3.10.1

    wget https://github.com/bazelbuild/bazel/releases/download/0.25.2/bazel-0.25.2-installer-linux-x86_64.sh
    bash bazel-0.25.2-installer-linux-x86_64.sh --user

Add and source the bin path to your ~/.bashrc file to call bazel

 export PATH=$PATH:~/bin
 source ~/.bashrc

Before building, make sure you deactivate any active virtual environments (i.e. run deactivate)

git clone https://github.com/IntelAI/he-transformer.git
cd he-transformer
export HE_TRANSFORMER=$(pwd)
mkdir build
cd $HE_TRANSFORMER/build
cmake .. -DCMAKE_CXX_COMPILER=clang++-6.0

Note, you may need sudo permissions to install he_seal_backend to the default location. To set a custom installation prefix, add the -DCMAKE_INSTALL_PREFIX=~/my_install_prefix flag to the above cmake command.

See 1a and 1b for additional configuration options. To install, run the below command (note, this may take several hours. To speed up compilation with multiple threads, call make -j install)

make install

To enable an integration with an experimental multi-party computation backend using garbled circuits via ABY, call

cmake .. -DNGRAPH_HE_ABY_ENABLE=ON

See MP2ML for details on the implementation.

We would like to thank the ENCRYPTO group from TU Darmstadt, particularly Hossein Yalame and Daniel Demmler, for helping with the ABY implementation.

Note: this feature is experimental, and may suffer from performance and memory issues. To use this feature, build python bindings for the client, and see 3. python examples.

First install the additional required dependencies:

sudo apt-get install doxygen graphviz

Then add the following CMake flag

cd doc
cmake .. -DNGRAPH_HE_DOC_BUILD_ENABLE=ON

and call

make docs

to create doxygen documentation in $HE_TRANSFORMER/build/doc/doxygen.

To build a client-server model with python bindings (recommended for running neural networks through TensorFlow):

cd $HE_TRANSFORMER/build
source external/venv-tf-py3/bin/activate
make install python_client

This will create python/dist/pyhe_client-*.whl. Install it using

pip install python/dist/pyhe_client-*.whl

To check the installation worked correctly, run

python3 -c "import pyhe_client"

This should run without errors.

cd $HE_TRANSFORMER/build
# To run single HE_SEAL unit-test
./test/unit-test --gtest_filter="HE_SEAL.add_2_3_cipher_plain_real_unpacked_unpacked"
# To run all C++ unit-tests
./test/unit-test

See examples/README.md for examples of running he-transformer for deep learning inference on encrypted data.

Please run maint/apply-code-format.sh before submitting a pull request.

• Fabian Boemer, Yixing Lao, Rosario Cammarota, and Casimir Wierzynski. nGraph-HE: a graph compiler for deep learning on homomorphically encrypted data. In ACM International Conference on Computing Frontiers 2019. https://dl.acm.org/doi/10.1145/3310273.3323047
• Fabian Boemer, Anamaria Costache, Rosario Cammarota, and Casimir Wierzynski. 2019. nGraph-HE2: A High-Throughput Framework for Neural Network Inference on Encrypted Data. In WAHC’19. https://dl.acm.org/doi/pdf/10.1145/3338469.3358944
• Fabian Boemer, Rosario Cammarota, Daniel Demmler, Thomas Schneider, and Hossein Yalame. 2020. MP2ML: A Mixed-Protocol Machine Learning Framework for Private Inference. In ARES’20. https://doi.org/10.1145/3407023.3407045