Federated Domain Generalization with Data-free On-server Matching Gradient
Authors: Binh Nguyen, Minh-Duong Nguyen, Jinsun Park, Viet Pham, Won-Joo Hwang
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experimental evaluations on various settings demonstrate the robustness of Fed OMG compared to other FL/FDG baselines. Our method outperforms recent SOTA baselines on four FL benchmark datasets (MNIST, EMNIST, CIFAR-10, and CIFAR-100), and three FDG benchmark datasets (PACS, VLCS, and Office Home). The reproducible code is publicly available. |
| Researcher Affiliation | Academia | 1 Pusan National University, Republic of Korea, 2 Trinity College Dublin, Ireland |
| Pseudocode | Yes | Algorithm 1: Federated Learning via On-server Matching Gradient |
| Open Source Code | Yes | The reproducible code is publicly available 1. 1https://github.com/skydvn/fedomg |
| Open Datasets | Yes | D.1 DATASETS MNIST (Lecun et al., 1998) consists of 70,000 grayscale images of handwritten digits and ten classes EMNIST (Cohen et al., 2017) extension version of original MNIST dataset consists of 70,000 samples and 10 classes CIFAR10 (Krizhevsky, 2012) consists of 60,000 images across 10 classes "Airplane", "Automobile", "Bird", "Cat", "Deer", "Dog", "Frog", "Horse", "Ship" and "Truck" CIFAR-100 (Krizhevsky, 2012) is a more challenging extension of CIFAR-10, consisting of 60,000 images distributed across 100 distinct classes. VLCS (Torralba & Efros, 2011) includes 10,729 images from four domains: "VOC2007", "Label Me", "Caltech101", "SUN09" and five classes ( bird , car , chair , dog and person ). PACS (Li et al., 2017) includes 9,991 images from four domains: Photos , Art , Cartoons , and Sketches and seven classes ( dog , elephant , giraffe , guitar , horse , house , and person ). Office Home (Venkateswara et al., 2017) includes four domains: Art , Clipart , Product , and Real . The dataset contains 15,588 samples and sixty five classes. |
| Dataset Splits | Yes | D.6 IMPLEMENTATION DETAILS To evaluate the performance of our proposed FDG method, we adopt the methodologies described in Lin et al. (2020); Acar et al. (2021) to simulate non-IID data. We utilize the Dirichlet distribution with two levels of data heterogeneity, specifically α = 0.1 and α = 0.5. Our experimental setup comprises three configurations of 100 clients, with join rate ratios of 1, 0.6, and 0.4, respectively, and global communication rounds set at 800. Each client undergoes 5 local training rounds using a local learning rate of 0.005, an SGD optimizer, and a batch size of 16. To ensure a fair comparison, all methods are evaluated with the same network architecture and settings. For FDG scenarios, we evaluate model performance using the conventional leave-one-domain-out method (Guo et al., 2023; Fan et al., 2024), where one domain is designated as the test domain and all other domains are used for training. |
| Hardware Specification | No | The paper does not explicitly describe the specific hardware (e.g., GPU/CPU models, memory, etc.) used for running its experiments. It mentions using a pretrained ResNet-18 backbone model and general training parameters but no hardware details. |
| Software Dependencies | No | The paper mentions using "Py Torch" for the pretrained ResNet-18 model, but does not specify a version number. It also mentions "SGD optimizer" without a version. No other specific software dependencies with version numbers are provided. |
| Experiment Setup | Yes | D.5 HYPER-PARAMETERS In this section, we present the hyper-parameters chosen for Fed OMG across different datasets. Table 4: Hyper-parameter Summary on PACS, VLCS, and Office Home Method Hyper-parameter PACS VLCS Office Home Global lr 5 10 2 5 10 2 5 10 2 Training lr 25 25 25 Iteration 21 21 21 Momentum 0.5 0.5 0.5 Searching radius 0.5 0.5 0.5 ... D.6 IMPLEMENTATION DETAILS ...Each client undergoes 5 local training rounds using a local learning rate of 0.005, an SGD optimizer, and a batch size of 16. ... Our experiments involve 100 global communication rounds with 5 local training rounds, utilizing the SGD optimizer with a learning rate of 0.001 and a training batch size of 16. |