Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Capture Global Feature Statistics for One-Shot Federated Learning
Authors: Zenghao Guan, Yucan Zhou, Xiaoyan Gu
AAAI 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experimental results demonstrate the effectiveness of our methods across diverse data-heterogeneity settings. To show the effectiveness of our proposed Fed CGS, we conduct experiments on both the global one-shot FL and the personalized one-shot FL. More details and extra results are included in the appendix. |
| Researcher Affiliation | Academia | 1Institute of Information Engineering, Chinese Academy of Sciences 2School of Cyber Security, University of Chinese Academy of Sciences 3Key Laboratory of Cyberspace Security Defense EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1 summarizes the procedure of Fed CGS with pseudocode. Algorithm 1: Procedure of Fed CGS |
| Open Source Code | Yes | Code https://github.com/Yuqin-G/Fed CGS |
| Open Datasets | Yes | Our experiments are conducted for classification task on the following image datasets: SVHN (Netzer et al. 2011), CIFAR10 (Krizhevsky, Nair, and Hinton 2009b), CIFAR100 (Krizhevsky, Nair, and Hinton 2009a), PACS (Li et al. 2017), and Office Home (Venkateswara et al. 2017). |
| Dataset Splits | Yes | For label shift scenario, we use Dirichlet distribution to generate disjoint non-IID client training datasets as same as other global one-shot FL methods (Zhang et al. 2022; Heinbaugh, Luz-Ricca, and Shao 2023; Dai et al. 2024) for fair comparison. For feature shift scenario, we follow the domain generalization settings in (Bai et al. 2024). Specifically, we select three domains for training and distribute their data across M clients. For data partitions, we follow the previous personalized FL methods (Zhang et al. 2020; Huang et al. 2021; Xu, Tong, and Huang 2023) that all clients have same data size, owning s% of data (20% by default) uniformly sampled from all classes and (100 s)% from a set of dominant classes. |
| Hardware Specification | No | No specific hardware details (GPU/CPU models, memory, etc.) are provided in the paper for running experiments. |
| Software Dependencies | No | The paper mentions optimizers like 'Stochastic Gradient Descent (SGD) optimizer' and 'Adam optimizer', and models like 'Res Net18' but does not provide specific version numbers for any software, libraries, or frameworks used in the implementation. |
| Experiment Setup | Yes | For methods involving backpropagation training (Fed Avg, Ensemble, DENSE, Co-Boosting, Fed PFT), we set the batch size to 128, the number of epochs to 50, and use the Stochastic Gradient Descent (SGD) optimizer with momentum = 0.9 and the learning rate = 0.01. For data-free knowledge distillation, we use the same generator as in (Dai et al. 2024; Zhang et al. 2022). It is trained with the Adam optimizer, a learning rate of 1e-3, for 30 epochs. During local training phase for each client, we employ mini-batch SGD as the local optimizer and set the batch size to 128, the local epoch to 1 for traditional personalized FL, 200 for Local-only and ours. The momentum is set to 0.5, the learning rate is set to 0.01, the weight decay is set to 5e-4 as (Xu, Tong, and Huang 2023). The number of global communication rounds for traditional personalized FL is set to 100 across all datasets. |