Collaborative Fairness in Federated Learning

This repository is the official implementation of Collaborative Fairness in Federated Learning . Accepted to FL-IJCAI 2020.

📋 In this work, we propose a Collaborative Fair Federated Learning (CFFL) framework which modifies the existing FL paradigm in order to encourage collaborative fairness (defined in our paper) and still maintain a competitive performance.

Citing

If you have found our work to be useful in your research, please consider citing it with the following bibtex:

@Inbook{Lyu2020,
    author="Lyu, Lingjuan
    and Xu, Xinyi
    and Wang, Qian
    and Yu, Han",
    title="Collaborative Fairness in Federated Learning",
    bookTitle="Federated Learning: Privacy and Incentive",
    year="2020",
    publisher="Springer International Publishing",
    address="Cham",
    pages="189--204",
}

Requirements

To install requirements:

conda env create -f environment.yml

📋 We recommend managing your environment using Anaconda, both for the versions of the packages used here and for easy management.

📋 Our code automatically detects GPU(s) through NVIDIA driver, and if not available it will use CPU instead. You can manually disable the use of GPU through changing the arguments.py or in the execution script main.py or test.py.

Training

To run the code in the paper, run this command:

python main.py or python.test.py

📋 Note that the arguments for experiments are specified within the main.py or test.py file. Alternatively, you can make changes to arguments.py where the default arguments are stored.

📋 The execution script is main.py or test.py. Running main.py starts the full-fledged experiments and it creates and writes to corresponding directories. Running test.py on the other hand only executes the code without creating directories or writing to them, but prints out the message to terminal.

Evaluation

To produce the collated accuracy and fairness results from complement execution of the code, run:

python examine_results.py --experiment_dir <experiment_dir>

📋 For each experiment directory created from running main.py, you can run the above command to get the corresponding accuracy and fairness results, along with the figures showing the convergence behaviors.

Results

We expect the algorithm to perform consistently on the same machine, but might achieve slightly different results over different machines.

Our CFFL achieves the following performance on:

MNIST

Fairness

Accuracy

Adult

Fairness

Accuracy

📋 Include a table of results from your paper If your main result is a figure, include that figure and link to the command or notebook to reproduce it.

Contributing

📋 Suggestions and questions are welcome through issues. All contributions welcome! All content in this repository is licensed under the MIT license.