Advancing Federated Learning in 6G: A Trusted Architecture with Graph-based Analysis

Official implementation for the paper Advancing Federated Learning in 6G: A Trusted Architecture with Graph-based Analysis accepted at GlobeCom2023

A pre-print of our paper can be viewed on arXiv

Proposed architecture

A trusted architecture for FL support with graph-based analysis

Environment setup

install pytorch, see PyTorch

conda install pytorch torchvision -c pytorch

PyTorch Geometric, see PyG

pip install torch-scatter torch-sparse torch-cluster torch-spline-conv torch-geometric -f https://data.pyg.org/whl/torch-1.13.0+cpu.html

Encryption

See paillier

pip install phe

Other packages

pip install tensorboard pip install jupyterlab

Graph dataset creation

The project is not aimed at training a CNN with SOTA result, but to analyze the performance of federated learning. Therefore, we train a CNN with a simple architecture. The example script is e.g.

python train_mnist_cnns.py --create_gnn --aggrPergraph 10 --modeslPeraggr 5 --local_epoch 5 --global_epoch 20 --model_perturb label --ascent_steps 3 --perturb_rate 0.5 --seed 42

which will give a set of models trained on different digits with given initialization seeds. The trained models will be used as a dataset.

For FEMNIST dataset, we use the LEAF repo, please see to the official repo and create FEMNIST dataset. After that, simply copy the train and test folder under ./datasets/FEMNIST.

Train GNN and MLP baseline

Once you have created the graph datasets, you can train a heterogeneous GNN on them. Simply run train_gnn.py or mlp_baseline.py with deep learning hyperparameters. Don't forget to include your created datasets :)

Dynamic filtering

After training the GNN, we can filter out malicious parameter nodes. For comparison with/without filtering, run e.g.

python train_perturb_mnist_cnns.py --global_epoch 20 --seed 42 --model_perturb label --ascent_steps 3 --perturb_rate 0.5

python train_perturb_mnist_cnns.py --global_epoch 20 --seed 42 --model_perturb label --ascent_steps 3 --perturb_rate 0.5 --filter_models --modelpath trained_gnns/gnn0.pt --aggrPergraph 10 --modeslPeraggr 5