FedKSS: One-shot federated learning with kernel space statistics.
Saved in:
| Title: | FedKSS: One-shot federated learning with kernel space statistics. |
|---|---|
| Authors: | Ding, Jiaman1,2 (AUTHOR) jiamanding@kust.edu.cn, Jiang, Linglin1,2 (AUTHOR) jiangll2024@126.com, Jia, Lianyin1,2 (AUTHOR) lianyinjia@kust.edu.cn, Fu, Xiaodong1,2 (AUTHOR) xdfu@kust.edu.cn, Jiang, Ying1,2 (AUTHOR) jy_910@163.com |
| Source: | Information Fusion. Nov2026, Vol. 135, pN.PAG-N.PAG. 1p. |
| Subjects: | Federated learning, Feature selection, Machine learning, Heterogeneity |
| Abstract: | • Proposes FedKSS, a training-free one-shot FL method via kernel statistics, cutting compute costs. • Introduces adaptive optimal layer selection to pick discriminative features, boosting performance. • Adds kernel feature alignment for personalization, needing one extra round for efficient FL. • Slashes comms and boosts privacy by uploading statistics, not full model parameters. • Beats SOTA by 15.71% global acc and 5.52% personalized acc under high heterogeneity. One-shot Federated Learning (FL) has emerged as a promising learning paradigm, capable of training a global model through a single communication round. However, existing one-shot FL methods still face the problem of high computational cost on the server or clients and struggle to handle non-IID (Independent and Identically Distributed) data stably and effectively. To address these issues, we propose FedKSS, a novel federated learning method based on kernel space statistics. Specifically, FedKSS maps features into a high-dimensional kernel space using Random Fourier Features, calculates local statistics on each client, and uploads these statistics to the server. The server then aggregates these local statistics to obtain global statistics, and based on this, computes the global prototype and covariance, thereby directly constructing a kernelized training-free classifier. To further enhance the discriminative power of the classifier, we introduce an optimal layer selection mechanism that uses features from different layers of the pre-trained model to construct the classifier and evaluate its performance, thereby determining the optimal feature layer for the current data distribution. In addition, we extend this method to personalization scenario, where clients only need one extra round of communication to download the global prototype from the server and achieve personalized learning through feature alignment. Experiments under high heterogeneity show that, compared with state-of-the-art methods, FedKSS improves global accuracy by 15.71% and local personalization accuracy by 5.52% across multiple benchmark datasets. These results confirm the effectiveness of FedKSS in data-heterogeneity settings. [ABSTRACT FROM AUTHOR] |
| Copyright of Information Fusion is the property of Elsevier B.V. and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.) | |
| Database: | Engineering Source |
Be the first to leave a comment!
