This repository is the official implementation of the following paper accepted by the 12th International Conference on Learning Representations (ICLR) 2024:
Incentive-Aware Federated Learning with Training-Time Model Rewards.
Zhaoxuan Wu, Mohammad Mohammadi Amiri, Ramesh Raskar, Bryan Kian Hsiang Low
To install requirements:
conda env create -f environment.yml
We use MNIST, FMNIST, SVHN, CIFAR-10, CIFAR-100, STT as benchmark datasets. The downloading and processing of the datasets will be handled by the code automatically.
We need to create data partitions for the heterogenous clients first before running federated learning algorithms on them.
The types of partitions that we support are
- Distribution-based label distribution skew:
noniid-labeldir - Quantity-based label distribution skew:
noniid-#label3(the number of labels can be modified) - Noise-based feature distribution skew:
gaussian_0.1(the sigma=0.1 can be modifed) - Quantity skew:
iid-diff-quantity - Homogeneous partition:
homo
We give one example here:
python partition.py -D mnist -N 50 -split noniid-labeldir -seeds 10At the beginning of the main_IAFL.py file, there are descriptions for the options required to run the code.
We give one example here:
python main_IAFL.py -D mnist -model cnn -split noniid-labeldir -N 50 -seed 0 -gpu 0The standalone accuracies are used as a contribution measure in the baseline experiments. To obtain them, run something like
python standalone.py -D mnist -model cnn -split noniid-labeldir -N 50 -seed 0 -gpu 0Then, we have already provided a commented code snippet in main_IAFL.py to use standalone accuracies as contributions. Uncomment that section of code and run the main_IAFL.py file.
We implemented FedAvg finetune (FedAvg-FT), local global FedAvg (LG-FedAvg), cosine gradient Shapley value (CGSV) and rank as baseline methods for comparison.
They can be run similarly as the main_IAFL.py file. The implementations are in main_fedavgft.py, main_lgfedavg.py, main_cgsv.py and main_rank.py, respectively.