Fake It Till Make It: Federated Learning with Consensus-Oriented Generation
Authors: Rui Ye, Yaxin Du, Zhenyang Ni, Yanfeng Wang, Siheng Chen
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on classical and real-world FL datasets show that Fed COG consistently outperforms state-of-the-art methods. |
| Researcher Affiliation | Academia | Rui Ye1,3, Yaxin Du1,3, Zhenyang Ni1,3, Yanfeng Wang2,1, Siheng Chen1,2,3B 1Shanghai Jiao Tong University, 2Shanghai AI Laboratory 3Multi-Agent Governance & Intelligence Crew (MAGIC) {yr991129,dorothydu,0107nzy,wangyanfeng,sihengc}@sjtu.edu.cn |
| Pseudocode | Yes | also see details in Algorithm 1 and illustration in Figure 1. |
| Open Source Code | Yes | Code is available at https://github.com/rui-ye/FedCOG |
| Open Datasets | Yes | We conduct comprehensive comparisons with many FL baselines on multiple datasets, including real-world FL dataset FLAIR (Song et al., 2022). Overall, we consider three classical datasets and one real-world FL dataset. 1) For the three classical datasets, including Fashion-MNIST (Xiao et al., 2017), CIFAR-10/100 (Krizhevsky et al., 2009)... 2) FLAIR (Song et al., 2022) is a recently released real-world FL multi-label dataset... |
| Dataset Splits | No | The paper uses standard datasets (Fashion-MNIST, CIFAR-10/100, FLAIR) which often have predefined splits. However, the paper does not explicitly state the train/validation/test split percentages, sample counts, or refer to predefined splits with specific citations or file names needed for full reproducibility of the data partitioning strategy. It mentions 'uniform test dataset' and 'personalized test dataset' but without detailed split ratios. |
| Hardware Specification | No | The paper discusses 'Local time of clients' and 'Computational cost' in Table 4, and mentions 'using a simple CNN' and 'Res Net18 (He et al., 2016) model', but does not specify the exact hardware (e.g., GPU/CPU models, memory) used to run the experiments. |
| Software Dependencies | No | For FLAIR, we use Res Net18 (He et al., 2016) from Py Torch API. However, no specific version numbers for PyTorch or other software dependencies are provided. |
| Experiment Setup | Yes | Training setups. For the three classical datasets, we consider 70 communication rounds, K = 10 clients, τ = 400 iterations of model training, and apply a simple CNN (Li et al., 2021a). For the more challenging FLAIR (Song et al., 2022) dataset, we consider 400 communication rounds and apply Res Net18 (He et al., 2016) model. Besides, we sample K = 200 clients in total and 10 clients in each FL round, and set τ = 10 because each client holds relatively few data samples. For all datasets, we use SGD as the optimizer with learning rate 0.01. For Fed COG, we apply the most computation-efficient way by setting the inputs as learnable, supervised by 256 pre-defined target labels. The default settings of λdis and λkd are 0.1 and 0.01, respectively. |