FedPS

FedPS is a Python module designed for data preprocessing in Federated Learning, primarily leveraging aggregated statistics. The preprocessing workflow involves the following five steps:

1. Local Statistics Estimation: Clients estimate local statistics from their local data.
2. Aggregation: The server receives the local statistics and performs aggregation.
3. Global Parameter Calculation: The server calculates the global preprocessing parameters.
4. Parameter Distribution: The global parameters are then sent back to the clients.
5. Data Preprocessing: Clients apply the preprocessing to their local data.

Installation

Dependencies

• Python (>= 3.9)
• Scikit-learn (>= 1.4)
• NumPy (>= 1.20)
• DataSketches
• PyZMQ

Building from source

1. Create a Python env

conda create --name fedps python=3.9
conda activate fedps

2. Clone this project

git clone https://github.com/xuefeng-xu/fedps.git

3. Build the project

cd fedps
pip install .

Usage

1. Set up communication channels

# Client1 channel
from fedps.channel import ClientChannel

channel = ClientChannel(
    local_ip="127.0.0.1", local_port=5556,
    remote_ip="127.0.0.1", remote_port=5555,
)

# Client2 channel
from fedps.channel import ClientChannel

channel = ClientChannel(
    local_ip="127.0.0.1", local_port=5557,
    remote_ip="127.0.0.1", remote_port=5555,
)

# Server channel
from fedps.channel import ServerChannel

channel = ServerChannel(
    local_ip="127.0.0.1", local_port=5555,
    remote_ip=["127.0.0.1", "127.0.0.1"],
    remote_port=[5556, 5557],
)

2. Specify FL_type and role in the preprocessor

• FL_type: "H" (Horizontal) or "V" (Vertical)

• role: "client" or "server"

# Client1 code example
from fedps.preprocessing import MinMaxScaler

X = [[-1, 2], [-0.5, 6]]
est = MinMaxScaler(FL_type="H", role="client", channel=channel)
Xt = est.fit_transform(X)
print(Xt)

# Client2 code example
from fedps.preprocessing import MinMaxScaler

X = [[0, 10], [1, 18]]
est = MinMaxScaler(FL_type="H", role="client", channel=channel)
Xt = est.fit_transform(X)
print(Xt)

# Server code example
from fedps.preprocessing import MinMaxScaler

est = MinMaxScaler(FL_type="H", role="server", channel=channel)
est.fit()

3. Run the script

# Run in three terminals
python client1.py
python client2.py
python server.py

PS: See more cases in the example folder.

Available preprocessing modules

• Discretization

  • KBinsDiscretizer

• Encoding

  • LabelEncoder
  • LabelBinarizer
  • MultiLabelBinarizer
  • OneHotEncoder
  • OrdinalEncoder
  • TargetEncoder

• Scaling

  • MaxAbsScaler
  • MinMaxScaler
  • Normalizer
  • RobustScaler
  • StandardScaler

• Transformation

  • PowerTransformer
  • QuantileTransformer
  • SplineTransformer

• Imputation

  • IterativeImputer (experimental)
  • KNNImputer
  • SimpleImputer

Differences from Scikit-learn

• Currently, this library does not support sparse data.
• KBinsDiscretizer, StandardScaler, and SplineTransformer cannot set the sample_weight parameter in their fit methods.
• IterativeImputer does not support the sample_posterior and n_nearest_features parameters.
• KNNImputer does not support custom weight funtion and distance metric.

Acknowledgement

This project is build on Scikit-learn.