Aono Tensor Demo

[iamtrask]

Helium : Aono Tensor Demo

Background: One of the most valuable contributions of the PySyft library to the Cryptography and AI communities is our Tensor level abstractions of Homomorphic Encryption algorithms. This allows users with no knowledge of Encryption to perform Encrypted Linear Algebra using a wide variety of encryption techniques. In this (and similar) tickets, we want to extend this ability to new Homomorphic Encryption schemes, such that users with no knowledge of Encryption can empirically try out various backend encryption techniques for their front end use cases with minimal effort. As such, while the backend API will differ from Tensor to Tensor, we want the front-end Linear Algebra interface to be as similar as possible, and to give intuitive explanations if the API differs such that calling functions in an Encrypted Tensor that doesn't support those operations explains WHY that specific type of Encryption cannot support the operation (i.e., Paillier can't multiply two encrypted numbers together... perhaps you should try another backend Encryption scheme).

Part 1: User Stories

• User Story A: As a Linear Algebra user, but NOT a Cryptography expert, I want the ability to perform linear algebra on Homomorphically Encrypted data using the Aono Homomorphic Encryption scheme. If operations are unsupported as a result of the limitations of Aono, I want Error messages to explain so and recommend alternate options. Furthermore, if operations must be approximated using operations such as Taylor Series (or other Polynomial Approximations), I want the option to turn these approximations on or off, with the default functionality to be ON. If I perform an operation for which approximation is necessary (non-linear), and have the "allow_approximations" feature turned off, it should return an error.

• User Story B: As a Linear Algebra user, but NOT a Cryptography expert who has implemented an algorithm in one of PySyft's other Encrypted Tensor libraries, I want the ability to simply change out the tensor type and be able to run all the same operations in this tensor library, with good error messages explaining why functions may not be supported (when relevant).

Part 2: Mockup

This Notebook is an excellent mockup of the functionality that we want to see with this new encrypted tensor class. It should be able to do all the operations in this notebook, with the added functionality of multiplication between two encrypted tensors (since that is explicitly supported by Aono).

Part 3: Acceptance Criteria

• a basic AonoTensor class mirroring the interface of syft.he.paillier.basic.PaillierTensor class, extending the TensorBase class.
• for encryption, it should use the OpenMined/PyAono library.
• the basic operations also listed in PaillierTensor should be included.
• unit tests should exist for all operations
• a notebook should exist demonstrating the same functionality as this one https://github.com/OpenMined/PySyft/blob/master/notebooks/Syft%20-%20Paillier%20Homomorphic%20Encryption%20Example.ipynb

[ivuckovic]

I'd like to take this issue.

[vasu-dev]

I would also like to collaborate on this issue .

[bharathgs]

@ivuckovic @vasu-dev any updates on this?

[ivuckovic]

@bharathgs unfortunately not yet. I'm getting some errors when importing Aono into jupyter notebook so I'm working on resolving them.

[ivuckovic]

Just a quick note... it seems that I had problems with Pari library because I tried to install it from the source from http://pari.math.u-bordeaux.fr/ on a Ubuntu and that didn't work. There is a lot easier way, using Ubuntu package libpari-dev. Now I'll focus on closing this issue.