Salvaging Federated Learning by Local Adaptation

Authors:

• Tao Yu
• Eugene Bagdasaryan
• Vitaly Shmatikov

Introduction

This repository contains official code and models for the paper, Salvaging Federated Learning by Local Adaptation.

Dependencies

Our implementation works with Python >=3.7 and PyTorch>=1.2.0. Install other dependencies: $ pip install -r requirement.txt

Datasets

We use two datasets in the paper:

• CIFAR-10 through torchvision datasets
• Reddit data, fetch following shared data and unzip files correspondingly,
  • test_data.json: https://drive.google.com/file/d/1X10JcpVGuRYqhUiwMPRCBJ6k-g9xhL3p/view?usp=sharing
  • Whole dataset: https://drive.google.com/file/d/1yAmEbx7ZCeL45hYj5iEOvNv7k9UoX3vp/view?usp=sharing
  • Dictionary: https://drive.google.com/file/d/1gnS5CO5fGXKAGfHSzV3h-2TsjZXQXe39/view?usp=sharing

Usage

1. For the federated learning model, configure the parameters using utils/params.yaml, to train a federated learning model on the Reddit Corpus, run:

$ python training.py --name text --params utils/words.yaml

2. For the adaptation of the federated learning model, configure the parameters using utils/adapt_text.yaml or utils/adapt_image.yaml, to adapt a federated learning model on the Reddit Corpus, run:

$ python adapt.py --name text --params utils/adapt_text.yaml

Similarly, change text, words.yaml and adapt_text.yaml into image, params.yaml and adapt_text.yaml to train and adapt the federated learning model on CIFAR.

Citation

If you use our code or wish to refer to our results, please use the following BibTex entry:

@misc{yu2020salvaging,
    title={Salvaging Federated Learning by Local Adaptation},
    author={Tao Yu and Eugene Bagdasaryan and Vitaly Shmatikov},
    year={2020},
    eprint={2002.04758},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}