SemiDFL: A Semi-Supervised Paradigm for Decentralized Federated Learning

Authors: Xinyang Liu, Pengchao Han, Xuan Li, Bo Liu

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through extensive experimentation, we demonstrate the remarkable performance superiority of the proposed DFL-Semi method over existing CFL and DFL schemes in both IID and non-IID SSL scenarios. Experiments Experimental Setup We use the decentralized communication topology in Figure 1(a) as an example to evaluate Semi DFL on different datasets, various labeled data ratios, and non-IID degrees.
Researcher Affiliation Academia 1Shenzhen Institute of Artificial Intelligence and Robotics for Society, China 2Department of Aeronautical and Aviation Engineering, The Hong Kong Polytechnic University, China 3School of Information Engineering, Guangdong University of Technology, China 4School of Information Science and Engineering, Southeast University, China EMAIL, EMAIL, EMAIL, EMAIL
Pseudocode Yes Algorithm 1: Semi DFL
Open Source Code Yes Code https://github.com/ez4lionky/Semi DFL
Open Datasets Yes We evaluate Semi DFL on MNIST (Le Cun, Cortes, and Burges 1998) and Fashion MNIST (Xiao, Rasul, and Vollgraf 2017) using Convolutional Neural Network (CNN), and on CIFAR10 (Krizhevsky, Nair, and Hinton 2010) using Res Net-18.
Dataset Splits Yes We evaluate Semi DFL on MNIST (Le Cun, Cortes, and Burges 1998) and Fashion MNIST (Xiao, Rasul, and Vollgraf 2017) using Convolutional Neural Network (CNN), and on CIFAR10 (Krizhevsky, Nair, and Hinton 2010) using Res Net-18.
Hardware Specification Yes All experiments are conducted using Py Torch 2.0 on a machine with 2 RTX 4090 GPUs.
Software Dependencies Yes All experiments are conducted using Py Torch 2.0 on a machine with 2 RTX 4090 GPUs.
Experiment Setup Yes We train models for 500 global rounds on all datasets. In each global training round, each client performs E (25 for MNIST, 50 for Fashion-MNIST and CIFAR-10) iterations of local training via mini-batch SGD with a batch size of B = 10. Other hyper-parameters during local model training are inherited from the default settings of Adam (Kingma and Ba 2014) and for all datasets, we use a learning rate of 0.05.