FedPCE: Federated Personalized Client Embeddings

Soma Hansel, Erich Kobler, Alexander Effland

Keywords: federated learning, deep learning, computer vision, transfer learning, model personalization, image classification, domain adaptation

Abstract:

Despite recent efforts, federated learning (FL) still faces performance challenges due to non-IID data distributions among clients.
This distribution shift complicates the addition of new clients and the transfer of federally learned models to unseen data.
Inspired by the adaptation ability of normalization layer parameters, we first demonstrate the effectiveness of models trained using FedBN when being adapted to so far unseen data.
Specifically, we extend the adaptation method based on a visual analysis of the normalization layer feature vectors.
We introduce Federated Personalized Client Embeddings (FedPCE), which utilizes local embeddings to capture the underlying structure of the normalization feature vectors and, by extension, the dataset.
Our results show that FedPCE performs comparably to other common FL algorithms during both training and adaptation.
Notably, FedPCE achieves this performance using only a fraction of the parameters during fine-tuning (32 parameters in our experiments) compared to other methods.

Primary Area: transfer learning, meta learning, and lifelong learning

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: 11770