All code used for Federated Learning Model training was provided by Bo Li and adapted by us.

Requirements:

• torch
• torchvision

Furthermore, because Federated Learning is very computationally expensive, we ran CIFAR-10 experiments exclusively on DTU's HPC.

To run it on HPC, first clone the repository:

git clone https://github.com/aerte/DFL.git

Then create a virtual environment called torch_dl or change the source in the job script submit_job to your desired environment.

Then just submit the job via:

bsub < submit_job.sh

All the relevant settings like the number of local epochs, alpha or what type of model to use (MLP,CNN or VGG11) can be changed in run_cifar.sh.

Specifically for the MNIST part of our report, please refer to MNIST-train which contains the corresponding code. This notebook can be run like any notebook locally or on Colab for example.

For post-processing of the results we used the three notebooks DataBinning, DataPlots and ExamplePlots. Please refer to the notebooks and our report for insight into how the uncertainty estimation was conducted.