FedRC: Tackling Diverse Distribution Shifts Challenge in Federated Learning by Robust Clustering
Authors: Yongxin Guo, Xiaoying Tang, Tao Lin
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that Fed RC significantly outperforms other SOTA cluster-based FL methods. Our code is available at https: //github.com/LINs-lab/Fed RC. |
| Researcher Affiliation | Academia | 1School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen, Guangdong, 518172, P.R. China 2The Shenzhen Institute of Artificial Intelligence and Robotics for Society 3The Guangdong Provincial Key Laboratory of Future Networks of Intelligence 4Research Center for Industries of the Future, Westlake University 5School of Engineering, Westlake University. |
| Pseudocode | Yes | In Algorithm 1 and Figure 10, we summarize the whole process of Fed RC. ... Algorithm 1 Fed RC Algorithm Framework ... Algorithm 3 Check and remove model in Fed RC |
| Open Source Code | Yes | Our code is available at https: //github.com/LINs-lab/Fed RC. |
| Open Datasets | Yes | We construct a scenario that encompasses label shift, feature shift, and concept shift issues. ... For feature shift, we employ the idea of constructing Fashion MNISTC (Weiss & Tonella, 2022), CIFAR10-C, CIFAR100-C, and Image Net-C (Hendrycks & Dietterich, 2019). ... We directly use the test datasets provided by Fashion MNIST, CIFAR10, and CIFAR100 datasets, and construct 3 test clients. ... Airline (Ikonomovska, 2020) ... Electricity (Harries et al., 1999) |
| Dataset Splits | Yes | We split the Fashion MNIST dataset into 300 groups of clients using LDA with a value of 1.0 for alpha. ... For the Mobile Net V2 model, we divided the dataset into 300 smaller groups of data (called 'clients'). ... Then, we used a technique called Latent Dirichlet Allocation (LDA) with a parameter of α = 1.0 to divide the dataset into the 300 clients. For the Res Net18 model, we also used LDA with α = 1.0, but we divided the dataset into 100 clients without making any copies. ... Nonparticipating clients: ... using the test sets provided by each dataset. |
| Hardware Specification | Yes | For all experiments, we use NVIDIA Ge Force RTX 3090 GPUs. Each simulation trail with 200 communication rounds and 3 clusters takes about 9 hours. |
| Software Dependencies | No | The paper mentions using SGD optimizer and Adam, and models like Mobile Net V2 and Res Net18, but does not provide specific version numbers for any software libraries or dependencies (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | We use a three-layer CNN for the Fashion MNIST dataset, and use pre-trained Mobile Net V2 (Sandler et al., 2018) for CIFAR10 and CIFAR100 datasets. We set the batch size to 128, and run 1 local epoch in each communication round by default. We use SGD optimizer and set the momentum to 0.9. The learning rates are chosen in [0.01, 0.03, 0.06, 0.1], and we run each algorithm for 200 communication rounds, and report the best result of each algorithm. |