goldenskygiang/flower-fhe

Flower framework for Federated Learning, with Fully Homomorphic Encryption integrated

Flower + OpenFHE

Getting started

python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Install OpenFHE

sudo apt-get install -y clang libomp5 libomp-dev cmake libgmp3-dev libntl-dev libomp-dev autoconf

export CC=/usr/bin/clang
export CXX=/usr/bin/clang++

cd ./openfhe-development
mkdir build 
cd build
cmake .. -DMATHBACKEND=6 -DWITH_OPENMP=ON -DWITH_NATIVEOPT=ON -DBUILD_SHARE=ON
make -j 8
sudo make install

Install OpenFHE-python binding

cd ./openfhe-python
mkdir build
cd build
cmake ..
make -j 8
make install

Observability

Observability is currenly only available when running simulations with python simul.py.

Follow the guide here. In summary:

• Install Prometheus and Grafana.
  • For Ubuntu Linux, brew services start grafana may not work. Install Grafana according to this guide and use sudo systemctl start grafana-server or grafana-server --homepath HOME_PATH --config CONF_PATH.
• Modify Grafana configuration to accept dashboards and datasources from Ray.
  • Allow anonymous view and set provisioning path to Ray folder.
• Start the services, they are accessible at:
  • Ray dashboard: localhost:8265
  • Prometheus: localhost:9090
  • Grafana: localhost:3000

Notes

• The transmission of the keys is done through model updates, not through modification of Flower's protocols.

• In the FHE scheme, the clients and the server will share the same crypto context, including its public and private keys. Hence, the server will send to clients the crypto context and the 2 keys generated by OpenFHE directly. These can be further encrypted by introducing the RSA scheme and adding RSA keys on top of FHE keys, or by conducting key exchange/derivation processes.

How to customize it to your own model?

• Make sure your model is written in PyTorch and it can be dissected layer by layer (the crytographic services encrypt/decrypt the model by layers, i.e. model.state_dict()).

• Make necessary modifications to the code files under the dataset and models modules. Other modules should remain untouched (if you do not change the function signatures).

Server

python server.py [options]

Server options

$ python server.py --help

usage: server.py [-h] --mode {fhe,sym} [--https] [--gpu] [--localhost] [--port PORT] --data_path DATA_PATH [--num_partitions NUM_PARTITIONS] [--batch_size BATCH_SIZE]
                 [--server_rounds SERVER_ROUNDS] [--fraction_fit FRACTION_FIT] [--fraction_evaluate FRACTION_EVALUATE] [--min_available_clients MIN_AVAILABLE_CLIENTS]
                 [--min_evaluate_clients MIN_EVALUATE_CLIENTS] [--min_fit_clients MIN_FIT_CLIENTS]

Run Federated Learning server with model encryption mechanisms

options:
  -h, --help            show this help message and exit
  --mode {fhe,sym}      Choose either Fully Homomorphic Encryption (fhe) or Symmetric Encryption (sym) mode
  --gpu                 Enable GPU for global evaluation (optional)
  --localhost           Run localhost only (optional)
  --port PORT           Port number (default is 8080)

Data Configuration:
  --data_path DATA_PATH
                        Path to data folder
  --num_partitions NUM_PARTITIONS
                        Number of data partitions (default is 100)
  --batch_size BATCH_SIZE
                        Data batch size (default is 32)

Federated Learning Configuration:
  --server_rounds SERVER_ROUNDS
                        Number of server rounds (default is 5)
  --fraction_fit FRACTION_FIT
                        Fraction of fitting clients (default is 0.2, limit double value from 0 to 1)
  --fraction_evaluate FRACTION_EVALUATE
                        Fraction of evaluating clients (default is 0.2, limit double value from 0 to 1)
  --min_available_clients MIN_AVAILABLE_CLIENTS
                        Minimum number of available clients to run federated learning (default is 2)
  --min_evaluate_clients MIN_EVALUATE_CLIENTS
                        Minimum number of evaluating clients (default is 2)
  --min_fit_clients MIN_FIT_CLIENTS
                        Minimum number of fitting clients (default is 2)

Client

python client.py [options]

Client options

$ python client.py --help

usage: client.py [-h] --mode {fhe,sym} [--https] [--gpu] [--localhost] [--host HOST] [--port PORT] --data_path DATA_PATH
                 [--num_partitions NUM_PARTITIONS] [--batch_size BATCH_SIZE]

Run Federated Learning client with model encryption mechanisms

options:
  -h, --help            show this help message and exit
  --mode {fhe,sym}      Choose either Fully Homomorphic Encryption (fhe) or Symmetric Encryption (sym) mode
  --gpu                 Enable GPU training & inference (optional)
  --localhost           Run localhost only (optional)
  --host HOST           IP address or hostname of the server; cannot be used together with localhost
  --port PORT           Port number (default is 8080)

Data Configuration:
  --data_path DATA_PATH
                        Path to data folder
  --batch_size BATCH_SIZE
                        Data batch size (default is 32)