PromptSFL: Improving Visual Prompt Tuning For Split Federated Learning

The novelty is limited. Integrating prompt tuning with SFL is straightforward and the additional design is trivial and easy to implement.

The meaning of the "skip" is unclear. It would be better for the authors to explain in detail what is being skipped, or what is the intuition behind the design of the skip prompt.

Another problem is that the experimental settings for some motivation/preliminary experiments are unclear. What are the datasets and data distribution used in Figure 5? And how about the settings on data distribution, model, etc?

How about the performance of PromptSFL under different settings of clients, i.e., the number of clients ranges from 50 to 100/200, etc?

According to Table 1, the performance upgrade by the PromptSFL is marginal compared to the VPF-SFL, a more detailed analysis would be helpful to convince the reviewer.

The reason why we need to do visual prompt fine-tuning over split federated learning is not elaborated. Since visual prompt fine-tuning performs badly over split federated learning, why do we still need to stick to it? In other words, why do authors not optimize more promising solutions over SFL. It is confused to choose an inferior solution in a particular research question and then improve it to the average bar.

Sending prompts to the centralized server completely breaks the privacy protocol of either split learning, federated learning, or split federated learning. In any cases, the prompt, which belongs to the user data should be sent to the server.

The evaluation is over image classification, which is already well-explored in the scope of federated learning and split federated learning. This back to my first point. Previous research can already achieve good performance over image classification. The authors first introduce visual prompt fine-tuning and decrease performance. And then the authors improve on this. Improvement should be based on existing state-of-the-art solutions and results. In fact, the reviewer does not think visual prompt fine-tuning should be a proper solution to choose. The reason is apparent as federated devices can suffer from heterogeneous data distribution, different centralized schemes. Visual prompt fine-tuning direct transfer knowledge between different devices can deviate the direction of correct gradients for the global model, if directly applied.

In Figure 2a, if a fixed dataset is used, the amount of data per client decreases as the number of clients increases. It raises the question of whether the lower performance of VPT compared to FFT is due to reduced data volume or if VPT is inherently less effective than FFT. Further experimentation is needed to support the conclusions drawn in Figure 2.

The idea of adaptively adjusting the learning rate based on loss values is interesting, but I am concerned about its suitability in the federated learning context, as it requires sharing loss values between the server and clients.

The claim of "improving the convergence speed" mentioned in the contributions is not clearly validated in the experiments.

This paper focuses on visual classification tasks and visual prompt tuning, which is quite limited in scope. The reviewer would like to know if and how PromptSFL could be applied to other types of tasks, such as NLP.

PromptSFL introduces additional computation and communication costs. A comparison of the computational overhead and communication cost of PromptSFL with other methods should be added.

Overall, this method seems straightforward and requires training a learnable weight and linear layer (as in Equation (8)). Is this process done on the server or on the clients?