dreamingraven / python-fhez

Official mirror of Python-FHEz; Python Fully Homomorphic Encryption (FHE) Library for Encrypted Deep Learning as a Service (EDLaaS).

Python-FHEz

(Formerly Python-ReSeal)

Python-FHEz is a fully-homomorphically-encrypted deep-learning library, made with love, and tears.

Cypherpunks write code.

Docs

Cypherpunks, read the docs:

Version	Provider	Link to Documentation
Master	RTD_
Stable	RTD_
Staging	RTD_
Dev	RTD_
Staging (Full Autodoc)	GitLab Pages

Cite

Either:

@online{reseal,
  author = {George Onoufriou},
  title = {Python-FHEz Source Repository},
  howpublished = {GitLab},
  year = {2021},
  url = {https://gitlab.com/DeepCypher/python-fhez},
}

Or if you do not have @online support:

@misc{reseal,
  author = {George Onoufriou},
  title = {Python-FHEz Source Repository},
  howpublished = {GitLab},
  year = {2021},
  note = {\url{https://gitlab.com/DeepCypher/python-fhez}},
}

python-fhez's Issues

How far from proper installation with `pip`?

Hi @DreamingRaven,

First of all, thanks for your preliminary work enabling FHE in machine learning. I had two questions in mind specifically about the longevity and long-term goals of this project, since the last commit was more than nine months ago (not a hit on you, I know myself how difficult it can be to maintain and document a medium-to-large library on your own).

1. I saw the installation instructions, and the preferred installation method seems to be still by running the code inside the published Docker containers due to the difficulties building the bindings with the SEAL library. I know that the team working on SEAL has no intention of actively maintaining Python bindings, as they prefer to focus their efforts on the C++ compatibility. Is this GitLab repository the one containing your custom SEAL bindings?
2. As a side note, how far do you think you are from having a proper self-contained installation backend with pip? Having working containers is nice because we can play with the package, but being able to install it would be crucial.

I am quite interested in your package, since we are trying to solve a real-world problem in the healthcare domain, where the preservation of privacy is the major concern on the data provider's side. Do you think the package is mature enough to build a complex network? Being very RAM and storage hungry, I am trying to figure out what the requirements of such a network would be.

I know that Intel has also developed its own SEAL integration for HE in machine learning with the nGraph compiler (now moved to the OpenVINO toolkit). Are there any major differences between your approach to FHE neural networks and theirs?

Finally, are you still actively working on this project? What are your short- and long-term goals with it?

Performance of the Network

Hi!
I had to download Fashion MNIST Dataset manually and then place it inside to train the network as there was an unzipping problem. I reached about 62% accuracy on this. Previous results here are also showing same kind of accuracy. Can you kindly tell me that is this the performance you are getting ? and is it fine?
Thank you.