LanXiaoPang613/FedLN

Federated Learning with Noisy Labels

Labeling Chaos to Learning Harmony: Federated Learning with Noisy Labels

Federated Learning (FL) is a distributed machine learning paradigm that enables learning models from decentralized private datasets while keeping the data on users' devices. However, most existing FL approaches assume high-quality labels are readily available, which is often not the case in real-world scenarios. In this paper, we propose FedLN, a framework for addressing label noise in different FL training stages, including initialization, on-device model training, and server model aggregation. FedLN computes per-client noise-level estimation in a single federated round and improves the models' performance by correcting or limiting the effect of noisy samples. Extensive experiments on various publicly available vision and audio datasets demonstrate a 24% improvement on average compared to other existing methods for a label noise level of 70%. We further validate the efficiency of FedLN in human-annotated real-world noisy datasets and report a 9% increase on average in models' recognition performance, emphasizing that FedLN can be useful for improving FL services provided to everyday users.

A complete description of our work can be found in our paper.

git clone https://github.com/FederatedML/FedLN
cd FedLN/src

Dependencies

Create a new Python enviroment (virtualenvs, anacoda, etc.) and install all required packages via:

foo@bar:~$ pip install -r requirements.txt

Executing experiments

From the src (root) directory of this repo, run:

# FedLN-NNC
foo@bar:~$ ./run.sh nnc --dataset_name cifar10 --model_name resnet20 --num_rounds 100 --num_clients 10 --noisy_frac 0.8 --noise_level 0.4 --noise_sparsity 0.7
# FedLN-NA_FedAvg
foo@bar:~$ ./run.sh na_fedavg --dataset_name cifar10 --model_name resnet20 --num_rounds 100 --num_clients 10 --noisy_frac 0.8 --noise_level 0.4 --noise_sparsity 0.7
# FedLN-AKD
foo@bar:~$ ./run.sh akd --dataset_name cifar10 --model_name resnet20 --num_rounds 100 --num_clients 10 --noisy_frac 0.8 --noise_level 0.4 --noise_sparsity 0.7

To execute any of the baselines, run:

# FedAvg
foo@bar:~$ ./run.sh fedavg --dataset_name cifar10 --model_name resnet20 --num_rounds 100 --num_clients 10 --noisy_frac 0.8 --noise_level 0.4 --noise_sparsity 0.7
# FedAvg + Bi-Tempered Logistic Loss
foo@bar:~$ ./run.sh bi_tempered_loss --dataset_name cifar10 --model_name resnet20 --num_rounds 100 --num_clients 10 --noisy_frac 0.8 --noise_level 0.4 --noise_sparsity 0.7
# FedAvg + Label Smoothing
foo@bar:~$ ./run.sh label_smoothing --dataset_name cifar10 --model_name resnet20 --num_rounds 100 --num_clients 10 --noisy_frac 0.8 --noise_level 0.4 --noise_sparsity 0.7
# FedAvg + Confidence Learning
foo@bar:~$ ./run.sh confidence_learning --dataset_name cifar10 --model_name resnet20 --num_rounds 100 --num_clients 10 --noisy_frac 0.8 --noise_level 0.4 --noise_sparsity 0.7

You can configure all federated parameters (i.e. number of federated rounds, number of clients, percentage of noisy data, dataset, model, etc.,) by passing them as command line arguments.

Reference

If you use this repository, please consider citing:

@misc{tsouvalas2022federated,
	title={Labeling Chaos to Learning Harmony: Federated Learning with Noisy Labels}, 
	author={Vasileios Tsouvalas and Aaqib Saeed and Tanir Ozcelebi and Nirvana Meratnia},
	year={2022},
	eprint={2208.09378},
	archivePrefix={arXiv},
	primaryClass={cs.LG}
}

@inproceedings{tsouvalas2022federatedworkshop,
	title={Federated Learning with Noisy Labels: Achieving Generalization in the Face of Label Noise},
	author={Vasileios Tsouvalas and Aaqib Saeed and Tanir {\"O}z{\c{c}}elebi and Nirvana Meratnia},
	booktitle={First Workshop on Interpolation Regularizers and Beyond at NeurIPS 2022},
	year={2022},
	url={https://openreview.net/forum?id=gNHMC4I0Pva}
}