Regularizing Federated Learning via adversarial model perturbations

Procedure

We regularize FL algorithms using Adversarial Model Perturbations (AMP) regularizer

We implement 4 algorithm for FL from scratch

• FedAVG
• FedProx
• FedNTD
• SCAFFOLD

Dependencies are given in requirement.txt file

pip install -r requirements.txt

How to RUN

Choose the required YAML file and run the following commands

python scaffold.py configs/scaffold_c100_amp.yaml > logfiles/scaffold_c100_amp.log
python fedntd.py configs/fedntd_c100_amp.yaml > logfiles/fedntd_c100_amp.log
python fedprox.py configs/fedprox_c100_amp.yaml > logfiles/fedprox_c100_amp.log
python fedavg.py configs/fedavg_c100_amp.yaml > logfiles/fedavg_c100_amp.log

Results

Comparsion of FL algorithms on CIFAR10 dataset

Comparison without AMP regularizer

Comparison with AMP regularizer

Comparsion of FL algorithms on CIFAR100 dataset

Comparison without AMP regularizer

Comparison with AMP regularizer

Members

• Gautam Kumar (B19EE031)
• Nirbhay Sharma (B19CSE114)