PyTorch framework for building, testing and benchmarking Binarised Neural Networks (BNNs), which can be converted into Fully Homomorphic Encryption (FHE) versions to be executed on CONCRETE.

Warning: Much of the code pertaining to FHE is not functional, due to issues with bootstrapping in CONCRETE. Future versions of CONCRETE may address this problem, but would require significant modifications to the code.

First install Poetry. Dependencies for the doren_bnn package can now be installed in a virtual environment with the following command:

poetry install

To view Jupyter notebooks with nbdime, configure the diff/merge drivers for this repository with the following command:

nbdime config-git --enable

To enable the use of Git pre-commit hooks, install the pre-commit package with the following command:

pre-commit install

To set up the Concrete library, please follow the instructions at [https://docs.zama.ai/concrete/lib/user/installation.html]. Take note that the library is only supported on macOS and Linux.

Note: Currently, it seems that compilation only works on Nightly Rust and not stable.

To run the executable in the library, run the following commands from the root directory:

cd doren_bnn_concrete
RUSTCFLAGS="target-cpu=native" cargo run --release

To build and install the Python bindings, run the following commands from the root directory:

cd doren_bnn_concrete
RUSTCFLAGS="target-cpu=native" maturin develop --release

You can now import the package doren_bnn_concrete from the root directory.