Centrality Matters: Lightweight Graph-centric Aggregation for Moderately Heterogeneous Federated Learning

SETUP AND INSTALLATION

The work was run on

• Python 3.8.16
• Ubuntu 22.04

To set up the environment and install all the packages, run:

conda env create -f environment.yml

Activate the imported environment:

conda activate fxg

RUN THE PROGRAM

SOTA BASELINES

Incuding customized FedDyn, FedGraB, FedPer and FedRep.

Follow the README.md in each algorithm folder under baseline/ folder to run.

FXG & OTHER BASELINES

Including: fxg- variant, FedAvg, FedProx, GNN, Avg+GNN and Avg+Deg.

DEFAULT VALUES:

Params	Type	Default Value
--strategy	string	FedAvg
--mu	float	0.1
--alpha	float	0.1
--dataset	string	CIFAR10
--rounds	int	500
--num_clients	int	10
--quantile	int	20
--fuse	int	1

FXG VARIANTS SUMMARY:

FXG- variant	Centralities (+ Anchor)
DEG	Degree
EIG	Eigenvector
LAP	Laplacian
BIH	string
EIGLAP	Eigenvector + Laplacian
EIGBIH	Eigenvector + Biharmonic
LAPBIH	Laplacian + Biharmonic
EIGLAPBIH	Eigenvector + Laplacian + Biharmonic
FULL	Degree + Eigenvector + Laplacian + Biharmonic
GAF	GNN

USAGE:

python main.py --strategy [aggregation strategy] 
               --mu [proximal_mu for FedProx only] 
               --alpha [dirichlet alpha] 
               --dataset [dataset name] 
               --rounds [num of communication rounds] 
               --num_clients [number of clients]
               --quantile [top % edges] 
               --fuse [1 (fuse with anchor) | 0 (pure)]

EXAMPLES:

• FedProx (500 rounds/default) on Fashion_MNIST dataset with $\alpha=1$ and $\mu=0.01$

  python main.py --strategy FedProx --dataset Fashion_MNIST --alpha 1 --mu 0.01
  

• FedAvg (5 rounds) on CIFAR10 dataset with $\alpha=10$

  python main.py --strategy FedAvg --rounds 5 --dataset CIFAR10 --alpha 10
  

• Pure GNN (5 rounds) on CIFAR10 dataset with $\alpha=10$ and quantile 20% (NO fusion)

  python main.py --strategy GA --rounds 5 --dataset CIFAR10 --alpha 10 --quantile 20
  

• GNN fused (5 rounds) on CIFAR10 dataset with with 5 clients $\alpha=0.1$ and quantile 25%

  python main.py --strategy GAF --rounds 5 --num_clients 5 --dataset CIFAR10 --alpha 0.1 --quantile 25
  

• All four centralities fused (2 rounds) on CIFAR10 dataset with $\alpha=1$ and quantile 15%

  python main.py --strategy FULL --rounds 2 --dataset CIFAR10 --alpha 1 --quantile 15
  

Run and save output log to file using: python main.py [PARAMS CONFIGS] > [PATH_TO_RESULT_FILE] 2>&1

For helps and understand more the arguments, run: python [.PY FILE] -h