Skip to content

alan-turing-institute/SHEEP

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

SHEEP is a Homomorphic Encryption Evaluation Platform

Disclaimer

SHEEP is development code. We strongly recommend that you do not deploy this prototype application in a production setting. SHEEP is intended to be an academic research framework to explore the homomorphic encryption (HE) space, and the characteristics of some HE libraries. However, we are not responsible for the content of these external libraries, and we offer no guarantees regarding their use.

Summary

The overall aim of the SHEEP project is to provide a platform for practitioners to evaluate the state-of-the-art of (fully) homomorphic encryption technology in the context of their concrete application. SHEEP enables this evaluation to be done across libraries implementing different HE schemes (with in some cases incomparable security).

Components of the SHEEP platform. These include an interpreter for the SHEEP language, an abstraction layer to incorporate HE libraries via their corresponding wrappers, and a reporting component. The user can either provide a program written in the high-level SHEEP language, or run a set of predefined benchmarks. Different aspects of the results can be visualized through a web interface.

To achieve this goal, the SHEEP platform includes an library agnostic language so that computations can be described generically. One can think of it as an assembly language.

To incorporate a library into the platform, all that needs to be done is write a wrapper for it, implementing the operations in the SHEEP instruction set. These corresponds to common operations in a HE scheme such as encryption, decryption, homomorphic addition, multiplication, comparison, etc.

Besides allowing the user to execute their own programs via either a web interface or programatically, the library incorporates a set of native benchmarks. These are classified in three categories according to their complexity, ranging from low-level microbenchmarks of the basic homomorphic arithmetic operations, to data analysis tasks.

Either on a native benchmark or a user-defined input, the platform evaluates the corresponding program on all the HE libraries that support the instructions of the program. The result is checked for correctness, and metrics such as running time (encryption, decryption, and evaluation) and ciphertext size are reported to a web interface and stored in a database. The web interface also allows convenient plotting of the results for easy visualisation.

The platform hence enables two main use cases:

  1. It allows users to evaluate a concrete custom computation across libraries (and thus encryption schemes) by simply describing it in a high-level language. This allows to set up a easily reproducible baseline when trying to optimize a given primitive based on HE, e.g. Private Information Retrieval, secure inner product, etc. For example, addition of two numbers in SHEEP looks like this:
INPUTS a b res
OUTPUTS res
a b ADD res
  1. The set of predefined benchmarks allows performance snapshots of the state-of-the-art of (fully) homomorphic encryption implementations to be continually produced, as new schemes are invented and their implementations are updated and optimized.

In summary, the general problem addressed by SHEEP is that of getting a satisfying answer, from the practitioner's point of view, to the following questions:

"What is the current state of the implementations of homomorphic encryption?"

"What homomorphic encryption scheme (and implementation) should I use in my application, given some functional and non-functional, e.g. security, latency or bandwidth constraints?"

In its current state, the SHEEP platform does not provide a complete solution to these problems, as we believe this requires a community effort. This is why SHEEP's code is open source.

Installation - docker (recommended)

Assuming you have docker installed, from this directory, do:

docker-compose build

This will build three Docker images: one for the "server", one for a web-based frontend, and one for some Jupyter notebooks containing examples of SHEEP usage.

To run the web frontend, do:

docker-compose run -p 5000:5000 sheep-frontend

If you point your browser to 0.0.0.0:5000 you should get the SHEEP frontend.

To run the notebooks, do:

docker-compose run -p 8888:8888 sheep-notebook

Now if you paste the URL given in the terminal for "sheep-notebook_1" into your browser, but with localhost instead of the hex string immediately after the "http://", you can access the Jupyter notebook examples.

Installation - local

Clone this repository onto your local machine. There are a few packages which the SHEEP platform itself uses, and then there are the HE libraries that we want to evaluate. Instructions for both follow:

Core SHEEP dependencies

SHEEP uses the cpprestsdk library to provide a RESTful API for the sheep server. On OSX, this can be installed with homebrew:

brew install cpprestsdk

For Linux (Ubuntu 16.04), see the relevant section of Dockerfile for a working recipe.

For parallel evaluation of circuits, SHEEP uses Intel's tbb library. This can be installed on OSX using homebrew:

brew install tbb

or for apt-based Linux installations:

apt-get -y install libtbb-dev

There are a few Python packages that are needed for the web front-end and/or the Jupyter notebooks. Install these with:

pip install  flask
pip install  wtforms
pip install  pytest
pip install  sqlalchemy
pip install  python-nvd3
pip install  requests
pip install  jupyter
pip install  matplotlib
pip install  pandas

Example HE libraries

After cloning the SHEEP repo, change to the SHEEP directory then do:

git submodule update --init --recursive

This will checkout the HElib and TFHE submodules in the lib/ directory.

TFHE

To build TFHE, first install the fftw library (download source from http://www.fftw.org/download.html then follow the instructions on that page)

Then build TFHE as follows:

cd lib/tfhe
mkdir build
cd build
cmake ../src -DENABLE_TESTS=on -DENABLE_FFTW=on -DCMAKE_BUILD_TYPE=optim -DENABLE_NAYUKI_PORTABLE=off -DENABLE_SPQLIOS_AVX=off -DENABLE_SPQLIOS_FMA=off -DENABLE_NAYUKI_AVX=off
make
sudo make install

HElib

HElib depends on the GMP and NTL libraries. To install gmp on OSX do:

brew install gmp

(this will install into /usr/local/Cellar/gmp - this location is needed when building ntl.) For Linux, gmp can be installed from source - see the relevant section of Dockerfile.

For ntl, download and follow instructions from here. At the ./configure step, add the argument GMP_PREFIX=/usr/local/Cellar/gmp if you installed gmp using homebrew, or the relevant location if you used another method.

Then build HElib as follows:

cd ../../HElib/src
make
mkdir build
cd build
cmake ..
make all

SEAL

Download and follow the instructions here.

libpaillier

Download the tar.gz file from here, unpack the archive, cd to the directory, then do

./configure
make
make install

Note that libpaillier depends on gmp - follow the instructions for installing this from the "HElib" section if you haven't already.

Building SHEEP

And finally build SHEEP:

cd <base SHEEP dir>
mkdir build
cd build
cmake ../
make all

This builds the examples in build/bin.

Run the tests with:

make test

Circuit language and input files

The inputs to SHEEP are circuit files, which have the extension .sheep and the following format:

INPUTS <circuit_input_0> ...
OUTPUTS <circuit_output_0> ...
<gate_input_0> ...  <GATE> <gate_output>
...