Federated_learning codes used for the the paper "Evaluation of Federated Learning Aggregation Algorithms" and "A Federated Learning Aggregation Algorithm for Pervasive Computing: Evaluation and Comparison"

Federated Distance (FedDist)

This is the code accompanying the Percom2021 paper "A Federated Learning Aggregation Algorithm for Pervasive Computing: Evaluation and Comparison" and the code of federated learning experiments by Sannara Ek during his master thesis.

Overview

This experiments compares 3 federated learning algorithms along with a new one, FedDist. The FedDist algorithm incorporates a pair-wise distance scheme for identifying outlier-like neurons/filters. These outlier-like neurons/filter may be in fact features learned from sparse data and so it is directly added to the server model for the next round of training.

Core Dependencies (tested and stable)

• Tensorflow 2.2.2
• PyTorch 1.1
• scikit-learn 0.22.1

All the working scripts are presented in a Jupiter notebook file format.

There is an array of 3rd party packages that is necessary for the entirety of the scripts to run. It is recommended to run command "pip3 install -r requirements.txt" in your virtual environment and working directory to replicate the environments used in this experiment.

!Note! Visual Studio is required to solve dependency problems when working on a Windows Machine

Data Preparation

"DATA_UCI.ipynb" and "DATA_REALWORLD_SPLITSUB.ipynb" are respectively used to prepare the UCI and REALWORLD dataset for training. Simply run all cells in a Jupyter notebook. The formatted dataset will be placed in a new directory "datasetStand"

FL script implementations

The FedAvg and FedPer implementations are found in the file "FedAvg_FedPer.ipynb". You must specify which algorithm you which to run in the third cell of the notebook by changing the "algorithm" variable to either "FEDAVG" or "FEDPER"

FedDist is found in the "FedDist.ipynb" file.

FedMA is found in the "FedMA.ipynb" file.

For all the federated algorithms, the third cell gives a variety of options and testing environment to choose from. We recommend leaving the configuration in default other than changing the "algorithm" variable and specifying the GPU/CPU to use. Simply run all cells to start training.

If preferred to run as a python script, convert the files to a .py format VIA Jupiter notebook (FILES -> Download as -> Python (.py)).

Additionally with the command below from a console achieves the same result:

jupyter nbconvert --to script '[ScriptName].ipynb'

Simply specify the wanted parameters in the third cell beforehand.

Results Interpretability

All results of each experiments shall generate the "savedModels" folder. Within this folder will contain subfolders with the name of the chosen configuration and model architecture of the experiment. Additionally, within each model architecture folder will contain the another subfolder with the name of the dataset used for the experiment. E.g a directory should appear like:

./savedModels/FED_5C_10LE_50CR_400D_100D_BALANCED/UCI

Now within this folder:

The final server model is saved in a .h5 format. The recorded training statistics foreach communication round, such as the accuracy and loss of the clients model and server model, are stored in the trainingStats folder. The results regarding the Global accuracy and the detail of the server model can be found on the generated Server-Measure.csv file. Results for the Personalization accuracy can be found in the indivualClients Measure.csv file and finally the Generalization accuracy can be found at the AllClientsMeasure.csv file.

Sample script sequence:

An example of execution would be to first download and format the dataset (UCI and REALWORLD) then execute one of the FL algorithms (requires several days on CPU).

1.DATA_UCI.ipynb
2.DATA_REALWORLD_SPLITSUB.ipynb
3.FedAvg_FedPer.ipynb/FedDist.ipynb/FedMA.ipynb

Citing this work:

@INPROCEEDINGS{Lala2103:Federated,
AUTHOR="Sannara Ek and François Portet and Philippe Lalanda and German Vega",
TITLE="A Federated Learning Aggregation Algorithm for Pervasive Computing:
Evaluation and Comparison",
BOOKTITLE="2021 IEEE International Conference on Pervasive Computing and
Communications (PerCom) (PerCom 2021)",
ADDRESS="Kassel, Germany",
DAYS=21,
MONTH=mar,
YEAR=2021,
KEYWORDS="Federated Learning; Edge Computing; Human activity recognition"
}

Contact:

Please contact the authors by [firstname].[lastname]@univ-grenoble-alpes.fr if you have issues with the code.

To contact Sannara Ek, Please use [firstname].[lastname]@gmail.com