AWDP-FL: An Adaptive Differential Privacy Federated Learning Framework
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThis paper proposes a new federated learning framework, Adaptive Weighted Differential Privacy Federated Learning (AWDP-FL), which aims to solve the balance between data privacy protection and model performance. The framework controls the gradient size by dynamically adjusting the cropping threshold, so as to achieve effective privacy protection. After gradient clipping, an adaptive mechanism is used to update the gradient, which can improve the performance of the model and ensure data privacy. Dynamic Gaussian noise is added when uploading model parameters to enhance privacy protection while maintaining the accuracy of the model. The server-side updates the global model by aggregating the perturbation parameters of all participants, ensuring that the high performance of the model is maintained while protecting privacy. Through experiments, the AWDP-FL framework can achieve good model performance while strictly protecting privacy, which provides strong technical support for application scenarios that need to pay close attention to data security.You also need to pay attention to the format of formula and chart names.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for Authors+ The abstract effectively outlines the research's objectives, methods, and key findings. The overall writing logic is smooth and easy to read.
+ The three pseudocodes given in the paper can fully show the process and ideas of the method, which is very helpful for the reader to understand.
+ The author has made sufficient theoretical foreshadowing before the experimental demonstration, and the structure of the article is relatively complete.
+ The experimental results demonstrate improvements of AWDP-FL over current federated learning methods that apply differential privacy.
- The references are not fully listed in the article. For example, I did not see where [23] and [24] were cited. The author needs to further check the reference information to ensure that it is complete and correct. In addition, it’s better to pay more attention to the latest thesis in these two years.
- The framework of the article before the formal introduction of the method needs to be further optimized, because the current introduction and previous work are not clear. In the “Introduction” and “Relevant Technical Theories” chapter, a considerable part of the content can be considered to be rewritten as Related Work and Preliminary.
- At the end of the Introduction, the first two points of the article contribution point stated by the author does not reflect the degree of differentiation, and the statement can be further strengthened.
- In the experimental part, the authors can consider adding a comparison of communication overhead, time overhead, and other dimensions. What's more, since the paper work involves differential privacy, different noise intensities may also affect the experimental results. The comparison of multiple dimensions is more beneficial to help the readers understand the comprehensive performance of the author's method.
- In 3.1(4), the authors mention that the server only randomly selects a subset of clients at a time to broadcast model parameters. So what's the ratio? Is it possible to set it as a hyperparameter in the experiment section?
- Note the network structure mentioned by the author in Section 4.1, and one of the highlights of the work in this paper is the gradient clipping optimization at the neural network layer level. So, will different neural network layers also have an impact on adaptive clipping optimization? If so, it also needs to be shown as a dimension of experimental comparison.
- In Section 4.2.1, the author has defined three values as a measure of the privacy budget. This is reasonable to some extent, but a more detailed explanation may be required. In fact, in many cases, privacy budgets are related to the amount of data, the number of training rounds, and can be calculated automatically. The ideas that the authors can formulate include the following: (a) AWDP-FL can achieve the same accuracy with less privacy budget. Or (b) the actual privacy cost paid by AWDP-FL in any case did not exceed the privacy budget.
- It would be better to prove github link of code to help readers reproduce and evaluate model.
Comments on the Quality of English Language
In some details, the manuscritp needs to be further optimized and improved.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsRemarks:
1) The abstract is still quite voluminous and it could have been much shorter or, at least, the number of repetitions could be reduced. It attempts to provide detailed technical information at some points that may seem excessive to anyone with little or no basic knowledge of the domain. Probably, bringing it a little down from the realms of the abstract and paying more attention to the authors’ main findings and their most important contributions, could provide even more value.
2) The introduction is long with the authors presenting many ideas and work that are relevant before presenting the problem under study. It could use a better organization of ideas in which the problem is presented in the beginning, then the difficulties followed by a short note on the proposed solution.
3) Despite, the manuscript including a discussion of related works, it can be further elaborated on how this study relates to previous research. For example, by creating a table that compares the presented AWDP-FL framework with other models on basic differences and benefits and drawbacks table would help to better understand the contribution.
4) The authors use a lot of technical terms in the manuscript and often do not expand the meaning of these terms to the nonspecialist. It is typical for academic writing to explain or at least give the definition of such terms as adaptive gradient clipping, Gaussian noise, etc because such a paper can be read not only by those who are interested in the field.
5) In the experimental analysis, the authors presented a general approach of the experiments conducted, but there are no specifics with regard to how the entire implementation was done, which datasets were chosen, why, and how the performance metrics were determined. To make their findings easier to replicate by other scientists the authors should have given more elaborative experimental procedures.
6) While the manuscript does provide a comparison with the existing methods, the same can be done in a more critical manner in the discussion part of the paper. For example, we omitted a comparison of the AWDP-FL framework to other techniques; information on cases when the proposed approach could be less effective; and an analysis of the limitations of the proposed technique in comparison with other methods.
7) In the manuscript, some experimental results are presented through figures and tables; while for some of them, the summaries provided in the text seem rather concise or sometimes, lacking sufficient analytical remarks. All tables and figures should be indicated by texts describing what is in the figure or table and why such things matter. Some points specifically, which could be elaborated more in terms of discussion on the trends observed are: Some points specifically, which could be elaborated more in terms of discussion on the trends observed are:
8) Fortunately, the presentation of the results is well captured by the result section but the conclusion may need more rewording to give the practical implications of the results section. However, there is an additional opportunity that can be offered under the future work section, such as the new specific ideas dealing with further analysis, for example, the utilization of the other databases, the enhancement of the given method, etc.
9) It should be also pointed out that, perhaps, the manuscript may be improved by referencing recent articles and studies on the subject area. There may be a temptation to refer to some of the areas that have been discussed in the federated learning and differential privacy less often; this may cause the paper to be less technical.
Comments on the Quality of English Language
There are examples of areas where enhancements for language use could be done to increase comprehensibility. There is a strong rationale to take time and proofread the work to check on the grammar, the construction of sentences as well as how the different styles have been used.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 4 Report
Comments and Suggestions for AuthorsThe paper addresses the critical issue of optimizing federated learning model performance while ensuring data privacy, proposing an innovative framework. After a thorough review, I offer the following comments and suggestions:
1.Originality and Contribution:
The manuscript presents a novel framework for federated learning that incorporates adaptive differential privacy, which is a significant contribution to the field of data privacy. The authors clearly articulate the advantages of their framework over existing techniques, which is commendable.
2.Theoretical Analysis and Methodology:
The authors provide a detailed description of the AWDP-FL framework and include pseudocode for the algorithms, aiding the reader's understanding of the proposed method's implementation. However, it is recommended that the authors further elaborate on the rationale behind the selection of key parameters and their impact, such as the choice of momentum update parameters τt1, τt2, ρ.
3.Experimental Design and Results:
The experiments utilize three public datasets (MNIST, Fashion-MNIST, and CIFAR-10), providing a solid foundation for validating the effectiveness of the proposed method. The authors also compare model performance across different privacy budgets, which is comprehensive. However, the authors are encouraged to include more datasets and network architectures in the experiments to further verify the generalizability of their method.
The results indicate that AWDP-FL outperforms existing methods across multiple metrics, but the authors are advised to explore the method's performance under different data distributions and scales.
4.Privacy Protection Analysis:
The paper provides a detailed analysis of privacy protection, including the client parameter upload phase and the server parameter distribution phase. However, the authors are encouraged to further discuss the types of privacy attacks that might be encountered in real-world deployments and how the AWDP-FL framework can defend against these attacks.
5. Discussion and Future Work:
The authors propose directions for future work in the discussion section, including more precise methods for measuring privacy loss and differential privacy strategies in heterogeneous federated learning scenarios. This is a positive direction, and the authors are advised to further clarify the potential impact and expected goals of these future works.
6.Writing and Organization:
The manuscript is well-organized and logically structured. However, the authors are advised to check for grammatical and spelling errors throughout the paper to ensure its professionalism.
It is recommended that the authors provide more background information on the limitations of existing methods in the introduction to better highlight the innovation of their proposed method.
7.Figures and Visualization:
The figures and visualizations in the paper help to illustrate the experimental results, but the authors are advised to check the clarity of the figures and the accuracy of the labels to ensure they are clearly understood by reaaders in the final publication.
8. Ensure that all relevant and key literature in the field is cited to demonstrate the depth and breadth of the research. Please check the consistency and accuracy of citations to ensure that all related work has been properly acknowledged and referenced as follows: The DOI number as: 10.1109/TCCN.2022.3164880 10.1109/TCSII.2022.3152522, 10.1109/TAES.2023.3266409, 10.1109/TAES.2024.3387447, 10.1109/TCSII.2022.3152522, 10.1016/j.dsp.2021.102994, 10.1016/j.sigpro.2022.108673, 10.1016/j.eswa.2024.124151 .
Comments on the Quality of English LanguageModerate editing of English language required.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Round 2
Reviewer 3 Report
Comments and Suggestions for AuthorsThe authors have made significant improvements, specifically in enhancing the clarity and depth of their explanation. The adaptive mechanisms have been explained in a more detailed manner. Furthermore, the results have been expanded.
Final remarks:
1. Analysis of the proposed algorithm complexity. Kindly, explain potential computational costs or overhead.
2. While the privacy aspects are well-discussed, it would be beneficial to expand the security analysis. Explain how robust the proposed framework against advanced privacy attacks is.
Comments on the Quality of English LanguageMinor.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 4 Report
Comments and Suggestions for AuthorsNo more comments.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf