Nettle is an integration layer between the Flower federated learning framework and Carbyne Stack. Nettle allows for large-scale privacy-preserving federated learning with MPC-based secure aggregation. It currently protects against reconstruction attacks. In the future, it will also prevent unauthorized model extraction.

See CSEP-0055: Privacy-Preserving Federated Learning for details. Please note that Nettle currently implements only a subset of the system proposed in CSEP-0055. In particular, the model plane protection mechanisms based on Confidential Computing are still under active development.

DISCLAIMER: Nettle is work-in-progress alpha software. The software is not ready for production use. It has neither been developed nor tested for a specific use case.

A nettle is a chiefly coarse herb armed with stinging hairs. Carbyne Stack Nettle is a fortified version of Flower that can resist certain kinds of attacks.

To set up a virtual environment and install the dependencies, invoke the following commands from the root of the Nettle repository:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install pip-tools
pip-sync

To install dependencies and to build the Docker images required for running Nettle in MP-SPDZ-based emulation mode, invoke

make -j

NOTE: Remember that you have to activate the virtual environment in each shell launched subsequently.

To run a Nettle Federated Learning experiment using CIFAR-10 with secure aggregation performed locally using a Docker-hosted 2-party MP-SPDZ backend, you first have to start a Nettle Orchestrator using

source .venv/bin/activate
cd orchestrator
PYTHONPATH=$(pwd)/.. PYTHONUNBUFFERED=1 PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python python3 orchestrator.py

After that, launch the CIFAR-10 model owner using

source .venv/bin/activate
cd examples/cifar_10
PYTHONPATH=$(pwd)/../.. PYTHONUNBUFFERED=1 PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python python3 cifar_model_owner.py

Any finally start two Nettle Clients by executing the following commands twice

source .venv/bin/activate
cd examples/cifar_10
PYTHONPATH=$(pwd)/../.. PYTHONUNBUFFERED=1 PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python python3 cifar_client.py

IMPORTANT: Please make sure that the virtual environment is activated. See the setup section for details.

Dependencies are managed using pip-tools. They are declared in the requirements.in file. After an update the corresponding requirements.txt file can be generated using

pip-compile

Syncing installed packages in the virtual environment is done using

pip-sync

You can generate the GRPC-related code using

python3 -m grpc_tools.protoc -I protos --pyi_out=generated \
  --python_out=generated --grpc_python_out=generated \
  protos/model_training.proto

Carbyne Stack Nettle is open-sourced under the Apache License 2.0. See the LICENSE file for details.

For information on how license obligations for 3rd-party OSS dependencies are fulfilled see the README file of the Carbyne Stack repository.

Please see the Carbyne Stack Contributor's Guide