federated is the source code for the Bachelor's Thesis.

Privacy-Preserving Federated Learning Applied to Decentralized Data (Spring 2021, NTNU)

Federated learning (also known as collaborative learning) is a machine learning technique that trains an algorithm across multiple decentralized edge devices or servers holding local data samples, without exchanging them. In this project, the decentralized data is the MIT-BIH Arrhythmia Database.


  • ML pipelines using centralized learning or federated learning.
  • Support for the following aggregation methods:
    • Federated Stochastic Gradient Descent (FedSGD)
    • Federated Averaging (FedAvg)
    • Differentially-Private Federated Averaging (DP-FedAvg)
    • Federated Averaging with Homomorphic Encryption
    • Robust Federated Aggregation (RFA)
  • Support for the following models:
    • A simple softmax regressor
    • A feed-forward neural network (ANN)
    • A convolutional neural network (CNN)
  • Model compression in federated learning.



Initial Setup

1. Cloning federated

$ git clone https://github.com/dilawarm/federated.git
$ cd federated

2. Getting the Dataset

To download the MIT-BIH Arrhythmia Database dataset used in this project, go to https://www.kaggle.com/shayanfazeli/heartbeat and download the files

  • mitbih_train.csv
  • mitbih_test.csv

Then write:

mkdir data
mkdir data/mitbih

and move the downloaded data into the data/mitbih folder.

Installing federated locally

1. Install the Python development environment

On Ubuntu:

$ sudo apt update
$ sudo apt install python3-dev python3-pip  # Python 3.8
$ sudo apt install build-essential          # make
$ sudo pip3 install --user --upgrade virtualenv

On macOS:

$ /usr/bin/ruby -e "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/master/install)"
$ export PATH="/usr/local/bin:/usr/local/sbin:$PATH"
$ brew update
$ brew install python  # Python 3.8
$ brew install make    # make
$ sudo pip3 install --user --upgrade virtualenv

2. Create a virtual environment

$ virtualenv --python python3 "venv"
$ source "venv/bin/activate"
(venv) $ pip install --upgrade pip

3. Install the dependencies

(venv) $ make install

4. Test TensorFlow Federated

(venv) $ python -c "import tensorflow_federated as tff; print(tff.federated_computation(lambda: 'Hello World')())"

Installing with Docker (optional)

Build and run image from Dockerfile

$ make docker

Running experiments with federated

federated has a client program, where one can initialize the different pipelines and train models with centralized or federated learning. To run this client program:

(venv) $ make help

This will display a list of options:

usage: python -m federated.main [-h] -l  -n  [-e] [-op] [-b] [-o] -m  [-lr]

Experimentation pipeline for federated 🚀

optional arguments:
  -b , --batch_size     The batch size. (default: 32)
  -e , --epochs         Number of global epochs. (default: 15)
  -h, --help            show this help message and exit
  -l , --learning_approach 
                        Learning apporach (centralized, federated). (default: None)
  -lr , --learning_rate 
                        Learning rate for server optimizer. (default: 1.0)
  -m , --model          The model to be trained with the learning approach (ann, softmax_regression, cnn). (default: None)
  -n , --experiment_name 
                        The name of the experiment. (default: None)
  -o , --output         Path to the output folder where the experiment is going to be saved. (default: history)
  -op , --optimizer     Server optimizer (adam, sgd). (default: sgd)

Here is an example on how to train a cnn model with federated learning for 10 global epochs using the SGD server-optimizer with a learning rate of 0.01:

(venv) $ python -m federated.main --learning_approach federated --model cnn --epochs 10 --optimizer sgd --learning_rate 0.01 --experiment_name experiment_name --output path/to/experiments

Running the command illustrated above, will display a list of input fields where one can fill in more information about the training configuration, such as aggregation method, if differential privacy should be used etc. Once all training configurations have been decided, the pipeline will be initialized. All logs and training configurations will be stored in the folder path/to/experiments/logdir/experiment_name.

Analyzing experiments with federated


To analyze the results with TensorBoard:

(venv) $ tensorboard --logdir=path/to/experiments/logdir/experiment_name --port=6060


Jupyter Notebook

To analyze the results in the ModelAnalysis notebook, open the notebook with your editor. For example:

(venv) $ code notebooks/ModelAnalysis.ipynb

Replace the first line in this notebook with the absolute path to your experiment folder, and run the notebook to see the results.


The documentation can be found here.

To generate the documentation locally:

(venv) $ cd docs
(venv) $ make html
(venv) $ firefox _build/html/index.html


The unit tests included in federated are:

  • Tests for data preprocessing
  • Tests for different machine learning models
  • Tests for the training loops
  • Tests for the different privacy algorithms such as RFA.

To run all the tests:

(venv) $ make tests

To generate coverage after running the tests:

(venv) $ coverage html
(venv) $ firefox htmlcov/index.html