Non-IID Degree Aware Adaptive Federated Learning Procedure Selection Scheme for Edge-Enabled IoT Network

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The main weaknesses of the paper are:

 

W1: The main contributions are not introduced in the introduction Section. It is better to add main contributions.

 

W2: The Related Works Section didn't introduce neural networks. It is better to add reference neural networks such as Neural network gain scheduling design for large envelope curve flight control law.

 

W3: It is better to add a background Section with more introduction and merge the Related Works Section into it.

 

W4: Why the conclusion Section is split by figures?

Author Response

We really appreciate your time and valuable comments. We have carefully revised our paper taking into consideration of your comments and suggestions. The following are detailed responses to your comments (C: Comment, A: Answer). Please note that modifications are marked as “bold” in the revised manuscript.

(C1) The main contributions are not introduced in the introduction Section. It is better to add main contributions.

(A1) Thank you for your comment. In this revision, we have clarified and emphasized our main contributions in the Introduction section (see page 2). Specifically, we have restructured the paragraph to explicitly highlight the novelty of our work, i.e., 1) the empirical analysis of the non-IID degree, 2) the practical applicability of the proposed scheme, and 3) extensive experimental evaluation.

 

(C2) The Related Works Section didn't introduce neural networks. It is better to add reference neural networks such as Neural network gain scheduling design for large envelope curve flight control law.

(A2)  Thank you for your comment. In this revision, we have added the reference "Neural Network Gain Scheduling Design for Large Envelope Curve Flight Control Law" [13]  and included other recent studies related to neural network techniques in federated learning [11,12]. These additions help to clarify the context and distinguish our contributions from prior work (see page 2-3).

[11] Çiplak, Z.; Yıldız, K.; Altınkaya, FEDetect: A Federated Learning-Based Malware Detection and Classification Using Deep 317 Neural Network Algorithms. Arabian Journal for Science and Engineering 2025, pp. 1–28. 

[12] Nazir, S.; Kaleem, M. Federated learning for medical image analysis with deep neural networks. Diagnostics 2023, 13, 1532. 319

[13] Zhang, P.; Yang, X.; Chen, Z. Neural network gain scheduling design for large envelope curve flight control law. Journal of Beijing 320 University of Aeronautics and Astronautics 2005, 31, 604–608. 

 

(C3) It is better to add a background Section with more introduction and merge the Related Works Section into it.

(A3) Thank you for your comment. As per your suggestion, we have added a Background section and merged the Related Works section into it. In this section, we describe the fundamentals of the federated learning procedure and the non-IID problem. Also, we have revised the description of the related works to better clarify their key contributions and limitations, which are addressed by our proposed scheme (see pages 2-3).

 

(C4) Why the conclusion Section is split by figures?

(A4) Thank you for your comment. Due to a layout problem during editing, the Conclusion section was unintentionally split by figures. This issue has been corrected in the revised manuscript (see Conclusion Section). We appreciate the reviewer’s understanding, and thank you for pointing it out.

 

 

 

Reviewer 2 Report

Comments and Suggestions for Authors

The authors propose an Adaptive FL Procedure Selection (AFLS) scheme that dynamically chooses between traditional and sequential FL methods based on the level of non-IID data across IoT devices. The idea is interesting and the paper is generally clear, but a few issues need to be addressed:

1. The background and motivation are not clearly explained. It would help to better highlight the problem being solved and why it matters.
2. The related work section is too brief and doesn’t provide a clear summary of existing methods for selecting FL procedures. More discussion is needed to show how this work is different.
3. In Section 4, please include an analysis of the complexity of the AFLS algorithm to better show its feasibility.
4. The experiments are too simple. More detailed results and comparisons with recent methods would make the evaluation more convincing.
5. The references are limited and miss recent work in this area. Adding more up-to-date studies would improve the paper.

Author Response

We really appreciate your time and valuable comments. We have carefully revised our paper taking into consideration of your comments and suggestions. The following are detailed responses to your comments (C: Comment, A: Answer). Please note that modifications are marked as “bold” in the revised manuscript.

 

(C1) The background and motivation are not clearly explained. It would help to better highlight the problem being solved and why it matters.

(A1) Thank you for your comment. As per your suggestion, we have clarified the background and motivation in the newly added Background section, which has been merged with the Related Works section for better coherence (see page 2).

 

(C2) The related work section is too brief and doesn’t provide a clear summary of existing methods for selecting FL procedures. More discussion is needed to show how this work is different.

(A2) Thank you for your comment. We have revised the Background Section to provide a clearer summary of existing methods for selecting FL procedures and to clarify the differences between previous studies and our proposed work (see pages 2~3). In addition, we have added more recent references to highlight the novelty and contributions of our work (see pages 2~3) [11, 12, 13].

[11] Çiplak, Z.; Yıldız, K.; Altınkaya, FEDetect: A Federated Learning-Based Malware Detection and Classification Using Deep 317 Neural Network Algorithms. Arabian Journal for Science and Engineering 2025, pp. 1–28.

[12] Nazir, S.; Kaleem, M. Federated learning for medical image analysis with deep neural networks. Diagnostics 2023, 13, 1532. 319

[13] Zhang, P.; Yang, X.; Chen, Z. Neural network gain scheduling design for large envelope curve flight control law. Journal of Beijing 320 University of Aeronautics and Astronautics 2005, 31, 604–608. 321

 

(C3) In Section 4, please include an analysis of the complexity of the AFLS algorithm to better show its feasibility.

(A3) Thank you for your comment. We have analyzed the complexity of the AFLS algorithm and confirmed that it operates in polynomial time (i.e., low complexity). This has been added to Section 4 of the revised manuscript (see pages 5~6). In addition,  we have highlighted that the low complexity of AFLS makes it feasible for practical implementation in real-world systems.

 

(C4) The experiments are too simple. More detailed results and comparisons with recent methods would make the evaluation more convincing.

(A4)  Thank you for your comment. We have revised the description of the simulation results to include a more detailed comparison with state-of-the-art FL methods (see pages 5-12). 

 

(C5) The references are limited and miss recent work in this area. Adding more up-to-date studies would improve the paper.

(A5)  Thank you for your comment. We have added more up-to-date studies to the Background Section  (see page 2).

Reviewer 3 Report

Comments and Suggestions for Authors

Overall a good work but requires major revisions:

Lack of novelty and contribution clarity: The manuscript does not clearly demonstrate how it advances the state of the art. Authors should elaborate on the novelty of their approach compared to existing methods.

Insufficient experimental validation: The performance evaluation lacks comprehensive benchmarking against relevant baseline models or recent literature. More extensive experiments are necessary.

Poor organization and clarity: The structure of the paper is disjointed in places, with unclear transitions between sections. Technical descriptions need refinement for better readability.

Incomplete methodology: Key implementation details are missing, such as specific parameters, training procedures, or datasets used. This hinders reproducibility.

Limited discussion of results: The analysis of the results is superficial. A deeper interpretation and discussion of why certain results occur would strengthen the paper.

Figures and tables need improvement: Some figures lack proper labelling or are difficult to interpret. Improve visual presentation for better comprehension.

Author Response

We really appreciate your time and valuable comments. We have carefully revised our paper taking into consideration of your comments and suggestions. The following are detailed responses to your comments (C: Comment, A: Answer). Please note that modifications are marked as “bold” in the revised manuscript.

(C1) Lack of novelty and contribution clarity: The manuscript does not clearly demonstrate how it advances the state of the art. Authors should elaborate on the novelty of their approach compared to existing methods.

(A1)  Thank you for your comment. To clearly demonstrate how our work advances the state of the art, we have added a detailed summary of our key contributions in the revised manuscript. Furthermore, we have incorporated additional state-of-the-art references to provide a more comprehensive comparison. In addition, we have revised the manuscript to emphasize the limitations of existing approaches and the necessity of our proposed method in addressing those gaps (see pages 2~3).

 

(C2) Insufficient experimental validation: The performance evaluation lacks comprehensive benchmarking against relevant baseline models or recent literature. More extensive experiments are necessary.

(A2) Thank you for your valuable comment. We sincerely agree with your point. In this work, we selected baseline schemes that share the same underlying principles as AFLS in order to fairly evaluate its decision-making process between two fundamental FL procedures (i.e., traditional and sequential FL). Other methods, such as those based on personalization, regularization, or contrastive learning, pursue different objectives, which makes direct and fair comparison with AFLS difficult within our current evaluation framework.

Nevertheless, we agree that benchmarking AFLS against more advanced FL schemes through more extensive experiments would provide a broader understanding of its performance. We have acknowledged this limitation in a footnote in Section 4 (see page 8), and we consider it an important direction for future work.

 

(C3)Poor organization and clarity: The structure of the paper is disjointed in places, with unclear transitions between sections. Technical descriptions need refinement for better readability.

(A3) Thank you for your valuable comment. To improve the overall organization and clarity of the manuscript, we have restructured its flow. Specifically, the original Related Works section has been revised and incorporated into a newly created Background section, which now includes fundamental concepts and a more thorough discussion of the limitations of previous studies. Additionally, we have merged the System Model and Proposed Scheme sections to enhance coherence and improve the clarity of technical descriptions. These changes were made to ensure better readability and a smoother transition between sections for the reader.

 

(C4)Incomplete methodology: Key implementation details are missing, such as specific parameters, training procedures, or datasets used. This hinders reproducibility.

(A4) Thank you for your valuable comment. As per your suggestion, we have added detailed implementation information, including the dataset used, training parameters, and evaluation procedures, in the Evaluation Results section (see page 8). These additions aim to improve the reproducibility of our study and provide clearer guidance for future research.

 

(C5)Limited discussion of results: The analysis of the results is superficial. A deeper interpretation and discussion of why certain results occur would strengthen the paper.

(A5) Thank you for your valuable comment. To provide a deeper interpretation and discussion, we have added additional simulation results and revised the descriptions of the experimental findings in greater detail (see pages 8~11). These updates aim to clarify the reasons behind the observed behaviors and strengthen the overall analysis presented in the manuscript.

 

(C6)Figures and tables need improvement: Some figures lack proper labelling or are difficult to interpret. Improve visual presentation for better comprehension.

(A6) Thank you for your valuable comment. In response to your suggestion, we have improved the visual presentation of the figures and tables by enhancing labeling, readability, and clarity to facilitate better comprehension.

Reviewer 4 Report

Comments and Suggestions for Authors

This study proposes an adaptive FL procedure selection (AFLS) scheme that selects an appropriate FL procedure between the conventional FL procedure or the sequential FL procedure depending on the degree of non-IID between IoT devices to achieve both adequate learning accuracy and low convergence time. This overcomes the limitation of the sequential FL scheme that performs deep model learning in a serialized manner due to the lack of parallelism in the past, which leads to long convergence time. In addition, the proposed methodology introduces a device-to-device (D2D) based sequential FL scheme to further reduce the convergence time of the sequential FL procedure. It is shown that the proposed AFLS can reduce the convergence time by up to 16% compared to the sequential FL scheme, and improve the learning accuracy by up to 6–26% compared to the conventional FL scheme. Although this reviewer agrees that the proposed study performed well in a special experimental environment, the reviewer would like to revise the manuscript considering the following points.

1.  Section 2 seems to be a detailed explanation of the concept of FL and previous studies, but the current Section 2 seems to have been written without much meaning. It seems as if it was written in a way that the authors should look up the references themselves, but it is necessary to describe in more detail the methodological differences between the client selection or clustering approach in the traditional FL procedure and the sequential FL procedure introduced by the authors. It seems necessary to write a specific Background section to clearly explain the academic contributions of this study to readers interested in FL. When looking at the review of previous studies, there are only one or two lines about which researchers did what, but it does not describe why they did it and what the advantages and disadvantages are, making it difficult to identify the trend of FL technology development. In addition, it is necessary to explain how this part is connected to the improvement that the proposed methodology is intended to make, as presented in the abstract.

2. It is questionable whether Sections 3 and 4, which describe the proposed scheme, need to be separated, and it would be better to combine the two sections, with a structural and theoretical description of the proposed scheme in Section 3, experiments in Section 4, and additional discussion of the experimental results in Section 5. The experiments simply end with a performance comparison, and there is no discussion of how they are different or similar in relation to real life or existing theories.

3. In conducting the experiment, only the performance of 10 specific IoT devices was evaluated, and the authors seem to need to consider how to generalize this. In Section 6, it seems necessary to describe the limitations of the study, and explain the future research direction to overcome these limitations. In Section 5, it is necessary to add additional explanations about how the designed experimental plan and environment are fairly evaluated with existing FL methodologies and the proposed AFLS. The current experimental results section lacks analysis and discussion (how the obtained results affect and are applied to real-life or general cases).

Author Response

We really appreciate your time and valuable comments. We have carefully revised our paper taking into consideration of your comments and suggestions. The following are detailed responses to your comments (C: Comment, A: Answer). Please note that modifications are marked as “bold” in the revised manuscript.

(C1) Section 2 seems to be a detailed explanation of the concept of FL and previous studies, but the current Section 2 seems to have been written without much meaning. It seems as if it was written in a way that the authors should look up the references themselves, but it is necessary to describe in more detail the methodological differences between the client selection or clustering approach in the traditional FL procedure and the sequential FL procedure introduced by the authors. It seems necessary to write a specific Background section to clearly explain the academic contributions of this study to readers interested in FL. When looking at the review of previous studies, there are only one or two lines about which researchers did what, but it does not describe why they did it and what the advantages and disadvantages are, making it difficult to identify the trend of FL technology development. In addition, it is necessary to explain how this part is connected to the improvement that the proposed methodology is intended to make, as presented in the abstract.

(A1) Thank you for your comment. We have thoroughly revised Section 2 to provide a more detailed explanation of the concept of federated learning (FL) and to elaborate on previous studies (see pages 2-3). In particular, we have clearly described the methodological differences between client selection or clustering approaches in traditional FL procedures and the sequential FL procedure proposed in this work. To further clarify our academic contributions, motivation, and how our work addresses the limitations of prior research, we have added a dedicated Background section. In addition, we have revised the Introduction section (see page 2) to better emphasize the key contributions of our work.

 

(C2). It is questionable whether Sections 3 and 4, which describe the proposed scheme, need to be separated, and it would be better to combine the two sections, with a structural and theoretical description of the proposed scheme in Section 3, experiments in Section 4, and additional discussion of the experimental results in Section 5. The experiments simply end with a performance comparison, and there is no discussion of how they are different or similar in relation to real life or existing theories.

(A2) Thank you for your comment. As per your suggestion, we have revised the structure of the manuscript by merging Sections 3 and 4. The newly combined section now includes both the theoretical description of the proposed AFLS scheme and the experimental results (see Section 4). In addition, we have enhanced the discussion of the experimental results by explaining how they compare to existing state-of-the-art methods in terms of performance and practical relevance. 

 

(C3) In conducting the experiment, only the performance of 10 specific IoT devices was evaluated, and the authors seem to need to consider how to generalize this. In Section 6, it seems necessary to describe the limitations of the study, and explain the future research direction to overcome these limitations. In Section 5, it is necessary to add additional explanations about how the designed experimental plan and environment are fairly evaluated with existing FL methodologies and the proposed AFLS. The current experimental results section lacks analysis and discussion (how the obtained results affect and are applied to real-life or general cases).

(A3)  Thank you for your comment. In our experiments, we conducted simulations with various numbers of IoT devices ranging from 5 to 50 (see Figure 3). The results showed similar trends regardless of the number of participating devices. To clarify this point, we have added a corresponding footnote in the revised manuscript (see page 8).

In addition, as per your suggestion, we have included a discussion of the limitations of our study and outlined future research directions in the Conclusion Section (see page 12). Furthermore, we have expanded the discussion in Section 4 to provide a more detailed analysis of the experimental results (see pages 5-12).

 

 

Reviewer 5 Report

Comments and Suggestions for Authors

1.Here the AFLS framework selects FL modes based on the non-IID degree ω, with a fixed accuracy threshold (0.6). How sensitive is AFLS to changes in this threshold, and is the threshold selection heuristic or data-driven?

2.The results obtaied show that FedAVG has the shortest convergence time but fails to meet accuracy in highly non-IID settings, while AFLS balances both. Could the authors elaborate on the conditions under which AFLS switches modes, and whether there is any risk of premature convergence or oscillation in mode selection?

3.The evaluation compares AFLS with FedAVG, FedS, and FedS-D2D. Why were other adaptive or hybrid FL methods (e.g., FedProx, MOON, or personalized FL approaches) not included as baselines to more comprehensively evaluate the relative performance of AFLS?

Author Response

We really appreciate your time and valuable comments. We have carefully revised our paper taking into consideration of your comments and suggestions. The following are detailed responses to your comments (C: Comment, A: Answer). Please note that modifications are marked as “bold” in the revised manuscript.

(C1) Here the AFLS framework selects FL modes based on the non-IID degree ω, with a fixed accuracy threshold (0.6). How sensitive is AFLS to changes in this threshold, and is the threshold selection heuristic or data-driven?

(A1) Thank you for your comment. If a relatively low threshold is set, AFLS is more likely to select the traditional FL procedure to reduce inference time, rather than mitigating accuracy degradation caused by the non-IID problem. In our work, the accuracy threshold is assumed to be determined by the AI service provider according to the requirements of the target AI application. To clarify this point, we have added a description to the revised manuscript (see page 5).

 

(C2) The results obtained show that FedAVG has the shortest convergence time but fails to meet accuracy in highly non-IID settings, while AFLS balances both. Could the authors elaborate on the conditions under which AFLS switches modes, and whether there is any risk of premature convergence or oscillation in mode selection?

(A2) Thank you for your comment. AFLS selects the FL procedure based on the degree of non-IIDness, and this selection is performed before the start of training. Once selected, the FL procedure remains fixed throughout the entire learning process. As per your comment, if the data characteristics at IoT devices change over time, i.e., the degree of non-IIDness dynamically varies, the initially selected FL procedure may no longer be optimal. We acknowledge this as a limitation of our current work and have added a corresponding discussion, along with potential directions for future research, in the Conclusion section (see page 12).

 

(C3) The evaluation compares AFLS with FedAVG, FedS, and FedS-D2D. Why were other adaptive or hybrid FL methods (e.g., FedProx, MOON, or personalized FL approaches) not included as baselines to more comprehensively evaluate the relative performance of AFLS?

(A3) Thank you for your comment. In our simulations, we selected baseline schemes that share the same underlying principle as AFLS to fairly evaluate its decision-making between two fundamental FL procedures (i.e., traditional and sequential FL). Methods (e.g., FedProx, MOON, and other personalized FL approaches) pursue different objectives (e.g., personalization, regularization, or contrastive learning), which makes a fair and direct comparison with AFLS difficult in our current evaluation framework. Nevertheless, we agree with your suggestion that comparing AFLS with more advanced FL schemes would provide a broader understanding of its performance. We have acknowledged this point in a footnote in Section 4 (see page 8).

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

I have no more concerns.

Author Response

We really appreciate your time and valuable comments. We have carefully revised our paper taking into consideration of your comments and suggestions. The following are detailed responses to your comments (C: Comment, A: Answer). Please note that modifications are marked as “bold” in the revised manuscript.

(C1) I have no more concerns.

(A1) Thank you for your valuable comment and effort during the review process.

Reviewer 3 Report

Comments and Suggestions for Authors

all my round 1 comments are addressed in this revised version- recommend acceptance 

Author Response

We really appreciate your time and valuable comments. We have carefully revised our paper taking into consideration of your comments and suggestions. The following are detailed responses to your comments (C: Comment, A: Answer). Please note that modifications are marked as “bold” in the revised manuscript.

 

(C1) all my round 1 comments are addressed in this revised version- recommend acceptance  

(A1) Thank you for your valuable comment and effort during the review process.

Reviewer 4 Report

Comments and Suggestions for Authors

This study proposes an adaptive FL procedure selection (AFLS) scheme that selects an appropriate FL procedure between the conventional FL procedure or the sequential FL procedure depending on the degree of non-IID between IoT devices to achieve both adequate learning accuracy and low convergence time. This overcomes the limitation of the sequential FL scheme that performs deep model learning in a serialized manner due to the lack of parallelism in the past, which leads to long convergence time. In addition, the proposed methodology introduces a device-to-device (D2D) based sequential FL scheme to further reduce the convergence time of the sequential FL procedure. It is shown that the proposed AFLS can reduce the convergence time by up to 16% compared to the sequential FL scheme, and improve the learning accuracy by up to 6–26% compared to the conventional FL scheme. The authors have carefully considered the previous comments and revised the manuscript, and the reviewer believes that the appropriately revised paper has been submitted.

Author Response

We really appreciate your time and valuable comments. We have carefully revised our paper taking into consideration of your comments and suggestions. The following are detailed responses to your comments (C: Comment, A: Answer). Please note that modifications are marked as “bold” in the revised manuscript.

 

(C1) The authors have carefully considered the previous comments and revised the manuscript, and the reviewer believes that the appropriately revised paper has been submitted.

(A1) Thank you for your valuable comment and effort during the review process.