anonymous429/fedkemf

FedKEMF: Resource-aware Federated Learning using Knowledge Extraction and Multi-model Fusion

FedKEMF

FedKEMF: Resource-aware Federated Learning using Knowledge Extraction and Multi-model Fusion

Dependencies

Current code base is tested under following environment:

1. Python 3.9
2. PyTorch 1.12.1 (cuda 11.3)
3. torchvision 0.13.1 4.scikit-learn 1.1.2
4. tensorboard 2.10.0
5. matplotlib 3.5.3

FedKEMF Overview –– Efficient federated learning

FedKEMF is a resource-aware FL algorithm, which aggregate an ensemble of local knowledge extracted from edge models. Different from exsisting works which aggregate the weights of each local model, FEDKEMF is distilled into a robust global knowledge as the server model through knowledge distillation.

Client local updates Download knowledge network from cloud server and optimize it jointly with local model.

Multi-model Fusion in cloud Could server ensemble all the selected client models, and distillate the ensemble model's knowledge to cloud model.

In this repository, we provide efficient federated learning experimental evaluation using FedKEMF. We test FedKEMF on ResNet20, ResNet32, ResNet44, VGG-11/16, and 2-layer simple CNN on various benchmark Non-IID federated learning settings.

Usage

Running experiments through using Docker container

The instruction for build the docker image for FedKEMF can be find in Docker/README.md. Please follow the requirements to build the docker image. To reproducing the experiments, please follow the instruction.

Running experiments through Python scripts

We highly recommend you create a conda virtual environment before you start the experiment. Instructions can be found in Anaconda.

After creating the environment, installing the dependencies with the correct versions:

• Installing PyTorch 1.12.1 (cuda11.3)

conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch

• Installing support packages

pip install -r requirements.txt

Experiment on Non-IID CIFAR-10

After configured all the dependencies, we can conduct the experiment.

In this subsection, clients are trained on CIFAR-10 with Non-IID settings.

Train ResNet-32 200 rounds with 10 clients and sample ratio = 1:

python3 knowlege_aggregation.py --comm_round=400 --k_model='resnet20' --model='resnet32' --dataset=cifar100 --batch-size=128 --epochs=20 --n_parties=10 --sample=0.7 --logdir='./logs/'

Train vgg-11 200 rounds with 30 clients and sample ratio = 0.7:

python3 knowlege_aggregation.py --comm_round=400 --k_model='resnet20' --model='resnet32' --dataset=cifar100 --batch-size=128 --epochs=20 --n_parties=10 --sample=0.7 --logdir='./logs/'

Multi-model experiment

Multi-model experiment will randomly initialize the FL system with different type of client models. To run the multi-model experiments, you need enable the argument --env=multi-model:

python3 knowlege_aggregation.py --comm_round=400 --k_model='resnet20' --env=multi-model --dataset=cifar100 --batch-size=128 --epochs=20 --n_parties=10 --sample=0.7 --logdir='./logs/'