Exploring One-Shot Federated Learning by Model Inversion and Token Relabel with Vision Transformers
Keywords: One-Shot Federated Learning, Model Inversion, Token Relabel
Abstract: One-Shot Federated Learning, where a central server learns a global model over a network of federated devices in a single round of communication, has recently emerged as a promising approach. For extremely Non-IID data, training models separately on each client results in poor performance, with low-quality generated data that are poorly matched with ground-truth labels. To overcome these issues, we propose a novel Federated Model Inversion and Token Relabel (FedMITR) framework, which trains the global model by better utilizing all patches of the synthetic images. FedMITR employs model inversion during the data generation process, selectively inverting semantic foregrounds while gradually halting the inversion process of uninformative backgrounds. Due to the presence of semantically meaningless tokens that do not positively contribute to ViT predictions, some of the generated pseudo-labels can be utilized to train the global model using patches with high information density, while patches with low information density can be relabeled using ensemble models. Extensive experimental results demonstrate that FedMITR can substantially outperform existing baselines under various settings.
Supplementary Material: zip
Primary Area: other topics in machine learning (i.e., none of the above)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 9305
Loading