Skip to main content

TF-Shell: Privacy preserving machine learning with Tensorflow and the SHELL encryption library

Project description

tf-shell

The tf-shell library supports privacy preserving machine learning with homomorphic encryption via the SHELL library and tensorflow.

This is not an officially supported Google product.

Getting Started

pip install tf-shell

See ./examples/ for how to use the library.

Background

Homomorphic encryption allows computation on encrypted data. For example, given two ciphertexts a and b representing the numbers 3 and 4, respectively, one can compute a ciphertext c representing the number 7 without decrypting a or b. This is useful for privacy preserving machine learning because it allows training a model on encrypted data.

The SHELL encryption library supports homomorphic encryption with respect to addition and multiplication. This means that one can compute the sum of two ciphertexts or the product of two ciphertexts without decrypting them. SHELL does not support fully homomorphic encryption, meaning computing functions of ciphertexts with arbitrary depth. That said, because machine learning models are of bounded depth, the performance benefits of leveled schemes (without bootstrapping, e.g. SHELL) outweight limitations in circuit depth.

Design

This library has two modules, tf_shell which supports Tensorflow Tensors containing ciphertexts with homomorphic properties, and tf_shell_ml some (very) simple machine learning tools supporting privacy preserving training.

tf-shell is designed for Label-DP SGD where training data is vertically partitioned, e.g. one party holds features while another party holds labels. The party who holds the features would like to train a model without learning the labels. The resultant trained model is differentially private with respect to the labels.

Building

Build From Source

  1. Install Bazelisk and python3 or use the devcontainer.

  2. Run the tests.

    bazelisk test --config test ...
    
  3. Build the code.

    bazelisk build --config release //:wheel
    bazelisk run //:wheel_rename
    
  4. (Optional) Install the wheel, e.g. to try out the ./examples/. You may first need to copy the wheel out of the devcontainer's filesystem.

    cp -f bazel-bin/*.whl ./  # Run in devcontainer if using.
    

    Then install.

    pip install tf_shell-...-manylinux.whl  # Run in target environment.
    

Note the cpython api is not compatible across minor python versions (e.g. 3.10, 3.11) so the wheel must be rebuilt for each python version.

Code Formatters and Counters

bazelisk run //:bazel_formatter
bazelisk run //:python_formatter
bazelisk run //:clang_formatter
cloc ./ --fullpath --not-match-d='/(bazel-.*|.*\.venv)/'

Update Python Dependencies

Update requirements.in and run the following to update the requirements files for each python version.

for ver in 3_9 3_10 3_11; do
  touch requirements_${ver}.txt
  bazelisk run //:requirements_${ver}.update
done

bazelisk clean --expunge

If updating the tensorflow dependency, other dependencies may also need to change, e.g. abseil (see MODULE.bazel). This issue usually manifests as a missing symbols error in the tests when trying to import the tensorflow DSO. In this case, c++filt will help to decode the mangled symbol name and nm --defined-only .../libtensorflow_framework.so | grep ... may help find what the symbol changed to, and which dependency is causing the error.

Contributing

See CONTRIBUTING.md for details.

License

Apache 2.0; see LICENSE for details.

Disclaimer

This project is not an official Google project. It is not supported by Google and Google specifically disclaims all warranties as to its quality, merchantability, or fitness for a particular purpose.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distributions

No source distribution files available for this release.See tutorial on generating distribution archives.

Built Distributions

tf_shell-0.1.1-cp311-cp311-manylinux_2_35_x86_64.whl (16.0 MB view details)

Uploaded CPython 3.11 manylinux: glibc 2.35+ x86-64

tf_shell-0.1.1-cp310-cp310-manylinux_2_35_x86_64.whl (16.0 MB view details)

Uploaded CPython 3.10 manylinux: glibc 2.35+ x86-64

tf_shell-0.1.1-cp39-cp39-manylinux_2_35_x86_64.whl (16.0 MB view details)

Uploaded CPython 3.9 manylinux: glibc 2.35+ x86-64

File details

Details for the file tf_shell-0.1.1-cp311-cp311-manylinux_2_35_x86_64.whl.

File metadata

File hashes

Hashes for tf_shell-0.1.1-cp311-cp311-manylinux_2_35_x86_64.whl
Algorithm Hash digest
SHA256 237daff421005876d1d366ca7f2941ffd4c19c2bd8014f9e926f7dabe9b4b94a
MD5 881275b98d59dab97345f1d8d97a260d
BLAKE2b-256 333dfbeb36e79219b753a72c5d87150a3a5cc0bca7564f49975c2d5b77daf6d3

See more details on using hashes here.

Provenance

File details

Details for the file tf_shell-0.1.1-cp310-cp310-manylinux_2_35_x86_64.whl.

File metadata

File hashes

Hashes for tf_shell-0.1.1-cp310-cp310-manylinux_2_35_x86_64.whl
Algorithm Hash digest
SHA256 27624b1bf965d0f228b253cb6f14cf403543522352c23d9ec668c49115a38de2
MD5 d33e187ae953e0791b305e9283908154
BLAKE2b-256 8e9243ac34e5d737ac21bf2a75500c550e1f1c93c5ba72a8678ea7aec14f8910

See more details on using hashes here.

Provenance

File details

Details for the file tf_shell-0.1.1-cp39-cp39-manylinux_2_35_x86_64.whl.

File metadata

File hashes

Hashes for tf_shell-0.1.1-cp39-cp39-manylinux_2_35_x86_64.whl
Algorithm Hash digest
SHA256 828f665cf1bd8677510faf062278ad49169baa92ab441c1094fbd22689015b0b
MD5 c67d4b62f5091ea5c14464b2b1b21c41
BLAKE2b-256 bcec3f79a623aa0463ab5acd6fc3cdc41a865efd6bfd1e1c09d75b4c96cefef9

See more details on using hashes here.

Provenance

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page