Federated Random Forest

Using Flower federated learning with sklearn Random Forest Classifier

This project implements a Federated Random Forest (FRF) using the federated learning library Flower and the sklearn random forest classifier. It is demonstrated through three clients as an example.

Getting Started

1. This project contains following source code files:

   helper.py
   server.py
   client1.py
   client2.py
   client3.py
   run.bat
   

2. Changes to be made in the code before you start:

   1. helper.py:
      Path of the dataset in the load_dataset() method.

      df = pd.read_csv('data.csv')

      If you want to add more clients, you need to modify the load_dataset() method so that the dataset is split into more sub-datasets.
   2. server.py:
      Modify the global values so that it meets your requirements.

      NUM_CLIENTS = 3
      ROUNDS = 2

   3. client1.py, client2.py, client3.py:
      Duplicate more client files if you need to, change the Client ID in the main().

      client_id = 1  # Client 1

   4. (Optional) server.py, client(1,2,3).py:
      If the run process results in a server address error, try the default address. Modify server_address in the server file and the client files respectively.

      server_address="0.0.0.0:8080"

Running

1. Install Required Libraries:

   pip install -r requirements.txt

2. Running the Code:

   You have two options to run the code:

   • Option 1: Using run.bat:

     Double-click on run.bat file. This will automatically execute the necessary scripts.

   • Option 2: Using Command Line:

     If you prefer to run the scripts manually, execute the following commands in your terminal:

     1. Loading the helper helper.py:

        python helper.py

     2. Start the server server.py:

        python server.py

     3. Run three clients in three terminals:

        python client1.py

        python client2.py

        python client3.py

Evaluating

1. Example of running results on the server:

    Server:
   
    INFO flwr 2023-11-23 17:18:58,487 | app.py:162 | Starting Flower server, config: ServerConfig(num_rounds=2, round_timeout=None)
    INFO flwr 2023-11-23 17:18:58,487 | app.py:175 | Flower ECE: gRPC server running (2 rounds), SSL is disabled
    INFO flwr 2023-11-23 17:18:58,487 | server.py:89 | Initializing global parameters
    INFO flwr 2023-11-23 17:18:58,487 | server.py:276 | Requesting initial parameters from one random client
    INFO flwr 2023-11-23 17:19:40,389 | server.py:280 | Received initial parameters from one random client
    INFO flwr 2023-11-23 17:19:40,389 | server.py:91 | Evaluating initial parameters
    INFO flwr 2023-11-23 17:19:40,389 | server.py:104 | FL starting
    DEBUG flwr 2023-11-23 17:19:41,214 | server.py:222 | fit_round 1: strategy sampled 3 clients (out of 3)
    DEBUG flwr 2023-11-23 17:20:22,909 | server.py:236 | fit_round 1 received 3 results and 0 failures
    DEBUG flwr 2023-11-23 17:20:22,909 | server.py:173 | evaluate_round 1: strategy sampled 3 clients (out of 3)
    DEBUG flwr 2023-11-23 17:20:23,613 | server.py:187 | evaluate_round 1 received 3 results and 0 failures
    DEBUG flwr 2023-11-23 17:20:23,613 | server.py:222 | fit_round 2: strategy sampled 3 clients (out of 3)
    DEBUG flwr 2023-11-23 17:21:04,206 | server.py:236 | fit_round 2 received 3 results and 0 failures
    DEBUG flwr 2023-11-23 17:21:04,206 | server.py:173 | evaluate_round 2: strategy sampled 3 clients (out of 3)
    DEBUG flwr 2023-11-23 17:21:04,914 | server.py:187 | evaluate_round 2 received 3 results and 0 failures
    INFO flwr 2023-11-23 17:21:04,914 | server.py:153 | FL finished in 84.52697419992182
    INFO flwr 2023-11-23 17:21:04,914 | app.py:225 | app_fit: losses_distributed [(1, 2.2396674950917563), (2, 2.224978526433309)]
    INFO flwr 2023-11-23 17:21:04,914 | app.py:226 | app_fit: metrics_distributed_fit {'Accuracy': [(1, 0.0), (2, 0.0)], 'Precision': [(1, 0.0), (2, 0.0)], 'Recall': [(1, 0.0), (2, 0.0)], 'F1_Score': [(1, 0.0), (2, 0.0)]}
    INFO flwr 2023-11-23 17:21:04,914 | app.py:227 | app_fit: metrics_distributed {'Accuracy': [(1, 0.937862), (2, 0.93827)], 'Precision': [(1, 0.937341), (2, 0.93773)], 'Recall': [(1, 0.937862), (2, 0.93827)], 'F1_Score': [(1, 0.932165), (2, 0.932682)]}
    INFO flwr 2023-11-23 17:21:04,914 | app.py:228 | app_fit: losses_centralized []
    INFO flwr 2023-11-23 17:21:04,914 | app.py:229 | app_fit: metrics_centralized {}

2. Example of running results on the client:

   Client 1:
   
   Label distribution in the training set: {0: 149362, 1: 24035}
   Label distribution in the testing set: {0: 37175, 1: 6175}
   
   INFO flwr 2023-11-23 17:19:40,377 | grpc.py:49 | Opened insecure gRPC connection (no certificates were passed)
   DEBUG flwr 2023-11-23 17:19:40,389 | connection.py:42 | ChannelConnectivity.IDLE
   DEBUG flwr 2023-11-23 17:19:40,389 | connection.py:42 | ChannelConnectivity.CONNECTING
   DEBUG flwr 2023-11-23 17:19:40,389 | connection.py:42 | ChannelConnectivity.READY
   Client 1 received the parameters.
   Parameters before setting:  [array(100), array(40), array(2), array(1)]
   Parameters after setting:  {'bootstrap': True, 'ccp_alpha': 0.0, 'class_weight': 'balanced', 'criterion': 'entropy', 'max_depth': 40, 'max_features': 'sqrt', 'max_leaf_nodes': None, 'max_samples': None, 'min_impurity_decrease': 0.0, 'min_samples_leaf': 1, 'min_samples_split': 2, 'min_weight_fraction_leaf': 0.0, 'n_estimators': 100, 'n_jobs': None, 'oob_score': False, 'random_state': None, 'verbose': 0, 'warm_start': False}
   Training finished for round 1.
   Trained Parameters:  [100, 40, 2, 1]
   ---------------------
   Accuracy : 0.93718570
   Precision: 0.93695380
   Recall   : 0.93718570
   F1 Score : 0.93131784
   ---------------------
   Parameters before setting:  [array(100), array(40), array(2), array(1)]
   Parameters after setting:  {'bootstrap': True, 'ccp_alpha': 0.0, 'class_weight': 'balanced', 'criterion': 'entropy', 'max_depth': 40, 'max_features': 'sqrt', 'max_leaf_nodes': None, 'max_samples': None, 'min_impurity_decrease': 0.0, 'min_samples_leaf': 1, 'min_samples_split': 2, 'min_weight_fraction_leaf': 0.0, 'n_estimators': 100, 'n_jobs': None, 'oob_score': False, 'random_state': None, 'verbose': 0, 'warm_start': False}
   Training finished for round 2.
   Trained Parameters:  [100, 40, 2, 1]
   ---------------------
   Accuracy : 0.93757785
   Precision: 0.93744875
   Recall   : 0.93757785
   F1 Score : 0.93174330
   ---------------------
   DEBUG flwr 2023-11-23 17:21:04,932 | connection.py:139 | gRPC channel closed
   INFO flwr 2023-11-23 17:21:04,932 | app.py:215 | Disconnect and shut down

Author: Hongwei Zhang

Citation:

1. Flower Framework Documentation
2. scikit-learn RandomForestClassifier