Towards Addressing Label Skews in One-Shot Federated Learning

Authors: Yiqun Diao, Qinbin Li, Bingsheng He

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our extensive experiments show that Fed OV can significantly improve the test accuracy compared to state-of-the-art approaches in various label skew settings.
Researcher Affiliation Academia Yiqun Diao, Qinbin Li & Bingsheng He National University of Singapore {yiqun,qinbin,hebs}@comp.nus.edu.sg
Pseudocode Yes Algorithm 1: The Fed OV algorithm.
Open Source Code Yes Code is available at https://github.com/Xtra-Computing/Fed OV.
Open Datasets Yes Datasets We conduct experiments on MNIST, Fashion-MNIST, CIFAR-10 and SVHN datasets. We use the data partitioning methods in Li et al. (2021b) to simulate different label skews.
Dataset Splits Yes In each task, we use a half of the test dataset as the public dataset for distillation for Fed KT and Fed DF and the remaining for testing.
Hardware Specification Yes All experiments are conducted on a single 3090 GPU.
Software Dependencies No The paper mentions 'Adam optimizer' and 'Re Lu as the activation function' but does not provide specific version numbers for any software dependencies like programming languages or libraries.
Experiment Setup Yes For local training, we run 200 local epochs for each client. We set batch size to 64 and learning rate to 0.001. For PROSER, we choose β = 0.01, γ = 1, according to the default trade-off parameter setting in the official code1. For adversarial learning, we set 5 local steps and each step size 0.002.