Data-Free Black-Box Federated Learning via Zeroth-Order Gradient Estimation
Authors: Xinge Ma, Jin Wang, Xuejie Zhang
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on large-scale image classification datasets and network architectures demonstrate the superiority of Fed ZGE in terms of data heterogeneity, model heterogeneity, communication efficiency, and privacy protection. Experimental Setup Datasets. The evaluation is conducted on two image classification datasets commonly used in FL research: 1) CIFAR-10 (Krizhevsky 2009); 2) CIFAR-100 (Krizhevsky 2009). |
| Researcher Affiliation | Academia | Xinge Ma, Jin Wang*, Xuejie Zhang School of Information Science and Engineering Yunnan University Kunming, China EMAIL, EMAIL |
| Pseudocode | Yes | See Appendix A for detailed algorithmic procedures of the proposed Fed ZGE framework. |
| Open Source Code | Yes | Code https://github.com/maxinge8698/Fed ZGE |
| Open Datasets | Yes | The evaluation is conducted on two image classification datasets commonly used in FL research: 1) CIFAR-10 (Krizhevsky 2009); 2) CIFAR-100 (Krizhevsky 2009). |
| Dataset Splits | Yes | To simulate data heterogeneity among clients, we follow prior work (Hsu, Qi, and Brown 2019) to heterogeneously partition the training set of each dataset among clients using a Dirichlet distribution Dir(α), where α is a concentration parameter that controls the degree of non-IID, with smaller values indicating more heterogeneous data distribution. To ensure reliable performance evaluation, we run the experiments three times with different random seeds and report the average accuracy with standard deviation of the global model on the original test set. |
| Hardware Specification | No | The paper does not provide specific hardware details (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments. |
| Software Dependencies | No | The paper does not provide specific ancillary software details (e.g., library or solver names with version numbers) needed to replicate the experiment. |
| Experiment Setup | Yes | Settings. The experiments are performed under two distinct FL settings with the key hyperparameters α={1, 0.1}, K=10, ε=1, T=100, B=500, and q=10: 1) homogeneous FL setting, where clients are forced to replicate homogeneous local models with the same architecture as the global model. We employ three types of network architectures to explore the effects of model scaling, including Res Net-18, Res Net34, and Res Net-50 (He et al. 2016); 2) heterogeneous FL setting, where clients are allowed to independently design heterogeneous local models. We employ Res Net-50 as the global model and allocate Res Net-18, Res Net-34, and Res Net-50 as the local models to clients in a ratio of 3:3:4. See Appendix C for full implementation details. |