Federated Learning in Streaming Subspace

ICLR 2025 Conference Submission3087 Authors

Keywords: Federated Learning, communication, subspace

Abstract:

Federated learning (FL) has received widespread attention due to its distributed training and privacy protection. However, existing federated learning methods encounter significant challenges, such as increased communication costs and degraded model performance, when processing non-independently and identically distributed (non-IID) data. This paper jointly alleviates these problems by analyzing and exploiting the low-rank properties of global model trajectories.

Primarily, we introduce a streaming subspace update strategy and then propose a general federated learning framework, $\\textbf{F}$erated $\\textbf{L}$earning in $\\textbf{S}$treaming $\\textbf{S}$ubspace ($\\texttt{FLSS}$). In $\\texttt{FLSS}$, local model updates are restricted to the global streaming subspace, resulting in low-dimensional trajectories. The server then aggregates these trajectories to update the global model. Comprehensive experiments verify the effectiveness of our framework. In Cifar100, the $\\texttt{FLSS}$-equipped FL method outperforms the baseline by 2.14$\\%$ and reduces the communication cost by 80$\\%$. $\\texttt{FLSS}$ utilizes the early training information of the global model to simultaneously improve the performance and communication efficiency of federated learning.

Primary Area: optimization

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.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

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: 3087