PromptSFL: Improving Visual Prompt Tuning For Split Federated Learning
Keywords: Federated Learning, Fine Tuning, Prompts
TL;DR: This paper proposes PromptSFL to improve the performance of visual prompt tuning in split federated learning.
Abstract: Conflict arises due to the disparity between the substantial resource demands of pre-trained models and the limited available resources of federated learning (FL) participants. Split learning presents a viable approach for adapting pre-trained models to FL, involving the allocation of a small portion of the pre-trained model to clients while deploying the remaining part on a server. Moreover, the application of Visual Prompt Tuning (VPT) to pre-trained models has shown state-of-the-art performances in parameter-efficient fine-tuning methods. However, VPT exhibits unsatisfactory performance in split federated learning (SFL) compared to its performance in centralized learning. In this paper, we first identify that VPT falls short of expectations in SFL due to the insufficient generalization capability of clients. To address this issue, we propose PromptSFL, which aligns the feature spaces of prompts between clients and the server to adapt VPT for SFL. PromptSFL transmits the final prompts in clients, termed skip prompts, to the first prompts in the server, enabling clients to extract more common features from the server. Additionally, we introduce a linear layer to map the prompts from clients to the feature space in the server during this skipping process, preventing the prompts of clients from overfitting to local datasets. Moreover, to enhance the convergence speed of SFL, PromptSFL employs an adaptive learning rate for clients. Extensive experiments demonstrate the effectiveness and efficiency of PromptSFL.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation 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: 1487
Loading