A Novel Federated Learning Framework Based on Conditional Generative Adversarial Networks for Privacy Preserving in 6G

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The paper proposes a federated learning (FL) framework based on conditional generative adversarial networks (CGAN). The topic addressed is timely and interesting, given the growing concerns about privacy and the momentum of federated learning. The paper is well written, properly presented, and the results seem to improve the performance of existing solutions, becoming an interesting article to the audience of this journal.

 

However, I have some comments that should be addressed before the paper can be accepted for publication:

 

- The relation of the current paper to 6G networks is not elaborated at all in the manuscript.

 

- The revision of the background in section 2 lacks a detailed review of the related works, since only 5 works which apply GAN in FL are described in section 2.3. It would be better to highlight what are the shortcomings of existing approaches and how the proposed solution addresses them.

 

- Figure 1, which represents the overall proposed framework, is not enough to visualize the different operations involved in the proposed framework. It would be better to have an additional diagram depicting the different operations involved in the stages of the framework?

 

- It is not clear to this reviewer, to which extent the methods described in section 3 are directly used from existing CGAN approaches or a contribution of the current article. This should be clearly indicated and referenced in the manuscript.

 

- The results show that the current solution outperforms baseline implementations in every single metric. Does the current proposal has any drawback aside from increased computational complexity?

 

- The improvement in privacy preserving is not clear to this reviewer's understanding. Moreover, why the PSRN is only studied for the proposed approach and FedGen, but not the other alternatives?

 

- The main findings and conclusions should be highlighted in the conclusion section.

 

- The manuscript should be revised for typos, mistakes (e.g., KL in line 241 should be KD, I guess) and grammatical errors.

 

- Figures and graphs have very low resolution and detail. The quality of these images should be considerably improved.

Comments on the Quality of English Language

The overall quality of the English level is fine, just some minor typos, mistakes and grammatical errors need to be revised throughout the manuscript.

Author Response

1. The relation of the current paper to 6G networks is not elaborated at all in the manuscript.

Response: We have made additions to the abstract, introduction and conclusion sections of the paper.

2. The revision of the background in section 2 lacks a detailed review of the related works, since only 5 works which apply GAN in FL are described in section 2.3. It would be better to highlight what are the shortcomings of existing approaches and how the proposed solution addresses them.

Response: Currently, there are relatively few articles that apply CGAN in federated learning and address similar issues. Mentioned related work is limited due to this. The shortcomings of existing methods have been emphasized again in Section 2.3. How the proposed solution addresses these drawbacks is explained at the end of Section 1 (Introduction), in Section 3.1 (Overview), and in Section 6 (Conclusion).

3.Figure 1, which represents the overall proposed framework, is not enough to visualize the different operations involved in the proposed framework. It would be better to have an additional diagram depicting the different operations involved in the stages of the framework?

Response: In Figure 1, we have visualized all the operations involved in the framework. The specific meanings of each part in the figure are described in detail in Sections 3 (Methods), 3.1 (Overview), 3.2.1 (EC-N Update), 3.2.2 (GD-N Update), and 3.2.3 (Server Update).

4. It is not clear to this reviewer, to which extent the methods described in section 3 are directly used from existing CGAN approaches or a contribution of the current article. This should be clearly indicated and referenced in the manuscript.

Response: The CGAN used in this paper is an existing one, but the specific model structure and parameters are defined by us. This has been explicitly stated in line 187 of Section 3 (Methods).

5. The results show that the current solution outperforms baseline implementations in every single metric. Does the current proposal have any drawback aside from increased computational complexity?

Response: There are various challenges in federated learning. This paper primarily focuses on issues related to insufficient data volume, uneven data distribution, and privacy protection, proposing corresponding solutions. As for other challenges in federated learning, such as client dropouts and client trust issues, the impact of the proposed approach remains uncertain and will be the subject of our future research efforts. We have also provided an explanation of this in the 6. Conclusion section.

6. The improvement in privacy preserving is not clear to this reviewer's understanding. Moreover, why the PSRN is only studied for the proposed approach and FedGen, but not the other alternatives?

Response: To enhance privacy protection, we have implemented an improved measure to guard against potential gradient leakage attacks in federated learning, thus mitigating the risk of privacy breaches. This improvement involves the use of data extractors to minimize the direct exposure of raw data to the classifier by performing feature extraction before transmitting the data to the classifier. Furthermore, clients only upload model parameters of the classifier and generator to the server. This means that even in the event of a gradient leakage attack, attackers can only reconstruct the data input to the classifier based on the classifier parameters, as they cannot access the parameters of the extractor, thus preventing the reconstruction of the original data. This method effectively safeguards data privacy.

In our research, we primarily focused on FedGen and our proposed method, conducting a comparison based on PSNR values. This is because only these two methods include generators that generate new data, allowing us to assess their performance by calculating the PSNR values between the original data and the generated data. In contrast, other methods did not implement privacy protection measures, which could result in complete data leakage in the event of a gradient leakage attack, leading to inferior privacy protection.

We have also provided additional explanations for this section in the analysis of PSNR values in the 5.2 (Experimental Results of Deep Residual Network).

7. The main findings and conclusions should be highlighted in the conclusion section.

Response: Certainly, in the conclusion section, we have reiterated the key findings and conclusions.

8. The manuscript should be revised for typos, mistakes (e.g., KL in line 241 should be KD, I guess) and grammatical errors.

Response: I have carefully reviewed and corrected the typos and grammar errors in the manuscript. On line 241, we do indeed mean KL divergence, which is a method for calculating loss. During model training, we employ the Knowledge Distillation (KD) technique, but when calculating loss, we use KL divergence.

9. Figures and graphs have very low resolution and detail. The quality of these images should be considerably improved.

Response: Certainly, we have increased the resolution of the framework diagram from 330 dpi to 1000 dpi.

Reviewer 2 Report

Comments and Suggestions for Authors

The paper presents a comprehensive investigation of the performance of NFL-CGAN in the context of federated learning. The study demonstrates notable improvements in model accuracy and robustness across various datasets, including FMNIST, CIFAR10, Office, and Digit5. The findings indicate that NFL-CGAN outperforms benchmark algorithms in both IID and non-IID scenarios, showcasing its potential for enhancing personalized local network performance on individual client datasets. The NFL-CGAN model consistently showcases outstanding performance in comparison to benchmark algorithms, particularly in scenarios involving non-IID data distributions. The paper effectively addresses concerns surrounding data privacy, highlighting the potential vulnerabilities of shared discriminator models within GANs and the impact of malicious behavior on model training and data privacy. The study effectively presents its findings with clear and informative visual aids, including RTA results and training loss curves, providing readers with a comprehensive understanding of the model's performance and convergence.

1. Insight into NFL-CGAN Architecture: It would be beneficial to have a more detailed description of the architecture and methodology specific to NFL-CGAN, shedding light on its unique mechanisms for improving federated learning performance.

2. Further Discussion on Privacy Preservation: Consider expanding on the specific strategies or mechanisms implemented within NFL-CGAN to address data privacy concerns, including any techniques used to mitigate potential privacy breaches and adversarial attacks discussed in the text.

3. Potential for Real-World Implementation: It would be valuable to discuss the practical applications and implications of the study's findings, particularly in real-world scenarios involving sensitive or private data. How might the performance of NFL-CGAN translate to practical use cases in industries with stringent data privacy requirements?

4. How might the findings on NFL-CGAN's performance impact the practical implementation of federated learning in industries requiring stringent data privacy measures?

5. The study's discussion around shared discriminator models and potential privacy breaches raises crucial questions about safeguarding sensitive data in federated learning models. What additional measures can be integrated to address these vulnerabilities effectively?

6. The rapid convergence and stability of NFL-CGAN as evidenced by the training loss curves prompt reflection on its potential for accelerating learning in federated settings. What implications does this hold for improving model convergence in real-world distributed learning scenarios?

7. Considering the outstanding performance of NFL-CGAN across both IID and non-IID scenarios, what nuanced considerations should be taken into account when applying this model to diverse data distributions in practical applications?

8. The observed discrepancies in RTA results among different models underscore the importance of understanding the underlying factors contributing to these variations. What insights can be gained from delving deeper into the local and global optimization aspects of federated learning models?

9. What nuanced considerations should be taken into account when applying NFL-CGAN to diverse data distributions in practical applications, given its outstanding performance across both IID and non-IID scenarios?

10. The rapid convergence and stability of NFL-CGAN as evidenced by the training loss curves prompt reflection on its potential for improving model convergence in real-world distributed learning scenarios. What implications does this hold for accelerating learning in such settings?

 

 

 

 

 

 

 

Comments on the Quality of English Language

Moderate revisions 

Author Response

1. Insight into NFL-CGAN Architecture: It would be beneficial to have a more detailed description of the architecture and methodology specific to NFL-CGAN, shedding light on its unique mechanisms for improving federated learning performance.

Response: We have introduced the NFL-CGAN architecture in 1. Introduction, 3. Methods, and 6. Conclusion. Now, in 3. Methods, we provide additional descriptions.

2. Further Discussion on Privacy Preservation: Consider expanding on the specific strategies or mechanisms implemented within NFL-CGAN to address data privacy concerns, including any techniques used to mitigate potential privacy breaches and adversarial attacks discussed in the text.

Response: To further enhance data privacy protection in the NFL-CGAN model, the following strategies and techniques can be employed: Firstly, differential privacy technology allows controlled noise to be introduced during model training and inference, ensuring the privacy of individual data while maintaining model performance. Secondly, adversarial training can be used to enhance the model's resistance to adversarial attacks, such as Generative Adversarial Network (GAN) attacks. Differential privacy strategies can reduce the reliance on client data, thus lowering the risk of privacy leakage. Lastly, layered encryption technology can provide robust encryption protection during the transmission of model parameters, ensuring that data is less susceptible to attacks during communication. These comprehensive measures contribute to elevating the level of data privacy protection in the NFL-CGAN model for privacy-sensitive applications while maintaining good performance.

3. Potential for Real-World Implementation: It would be valuable to discuss the practical applications and implications of the study's findings, particularly in real-world scenarios involving sensitive or private data. How might the performance of NFL-CGAN translate to practical use cases in industries with stringent data privacy requirements?

Response: NFL-CGAN, as a federated learning strategy, holds immense potential in the real world. It not only enables collaboration among multiple clients to enhance classification performance but also ensures effective protection of sensitive or private data. This approach finds widespread applications in fields such as healthcare, finance, education, government, and market research. It assists in compliance-driven data analysis, improves services and decision-making, all while meeting data privacy regulations and ethical requirements, providing robust support for data-driven decisions.

In the healthcare sector, the practical application of NFL-CGAN includes enhancing medical image analysis and case diagnoses while ensuring the privacy of patient data. Healthcare institutions can utilize NFL-CGAN for collaborative analysis of medical image data without the need to directly share sensitive patient images. Additionally, attackers are unable to access the original patient data, thus ensuring strong privacy protection. Through collaborative efforts, medical experts can improve the accuracy of disease diagnoses while adhering to regulatory requirements, thereby driving significant advancements in research and service quality within the healthcare industry.

We have also supplemented relevant content in the 6. Conclusion section.

 

4. How might the findings on NFL-CGAN's performance impact the practical implementation of federated learning in industries requiring stringent data privacy measures?

Response: The performance research results of NFL-CGAN have had a positive impact on the application of federated learning in fields where data privacy is a significant concern. These findings have increased the credibility and adoption of federated learning, making it more popular in industries such as healthcare, finance, and education. Businesses and institutions are more willing to adopt this approach because it not only improves performance but also ensures data privacy compliance, meeting industry regulations and legal requirements. Furthermore, these research results have also spurred more innovation and research in the field of privacy protection, promoting the development of new privacy protection technologies and further expanding the scope of federated learning applications. In summary, the performance research of NFL-CGAN provides viable solutions for data collaboration and privacy protection, contributing to increased industry efficiency and innovation while ensuring effective data privacy protection.

5. The study's discussion around shared discriminator models and potential privacy breaches raises crucial questions about safeguarding sensitive data in federated learning models. What additional measures can be integrated to address these vulnerabilities effectively?

Response: This issue is extensively discussed in the related work section 2.1. In summary, to effectively address potential privacy leakage concerns within federated learning models and vulnerabilities in shared discriminator models, the following additional measures can be adopted: Firstly, differential privacy techniques can be used to inject randomized noise, safeguarding client data, and preventing the leakage of sensitive information. Secondly, secure multi-party computation technology enables multiple parties to perform computations without exposing private inputs, thus preserving the privacy of client data. Furthermore, employing encryption techniques to encrypt client data ensures protection during transmission and storage. However, these methods may face challenges such as high communication or computational costs and hardware dependencies.

6. The rapid convergence and stability of NFL-CGAN as evidenced by the training loss curves prompt reflection on its potential for accelerating learning in federated settings. What implications does this hold for improving model convergence in real-world distributed learning scenarios?

Response: NFL-CGAN demonstrates fast convergence and stability in a federated learning environment, which is important for improving multiple learning scenarios. It means that in a variety of real-world applications, models can learn more quickly from data in multiple places, accelerating decision-making and problem-solving. In addition, reducing the number of communications and costs makes federated learning more suitable for large-scale and efficient data collaboration, broadening its application fields. Therefore, it is of great significance for the performance of NFL-CGAN to improve the efficiency and practicality of national learning. 

This content has also been added to the relevant experimental analysis of 5.2.

7. Considering the outstanding performance of NFL-CGAN across both IID and non-IID scenarios, what nuanced considerations should be taken into account when applying this model to diverse data distributions in practical applications?

Response: NFL-CGAN demonstrates rapid convergence and stability in the federated learning environment, which is crucial for improving distributed learning scenarios. It signifies that in various real-world applications, models can learn faster from data distributed across multiple locations, accelerating decision-making and problem-solving processes. Additionally, by reducing the frequency and costs of communication, it makes federated learning more suitable for large-scale and efficient data collaborations, broadening its scope of applications. Therefore, the performance of NFL-CGAN is important in enhancing the efficiency and practicality of distributed learning.

8. The observed discrepancies in RTA results among different models underscore the importance of understanding the underlying factors contributing to these variations. What insights can be gained from delving deeper into the local and global optimization aspects of federated learning models?

Response: Delving deeper into the local and global optimization aspects of federated learning models has multiple values. Firstly, understanding local optimal solutions means that we need to consider the local performance peaks of the model on each client. The data distribution and quantity of different clients may lead to different local optimal solutions, and these differences are one of the reasons for the different RTA results. Secondly, the global optimization strategy involves how to effectively integrate model parameters from different clients. In federated learning, the aggregation method and update rules of model parameters are crucial to obtain better global performance. In addition, research on communication efficiency helps us better understand the impact of different communication strategies and frequencies on the global convergence speed and performance of the model. Research on robustness involves the performance stability of the model under different clients and data distributions, which is crucial to maintaining the reliability and adaptability of the model. Finally, privacy and security issues need to be specially considered during the global optimization process to ensure that data privacy is fully protected, which is particularly important in real-world applications. These research directions will help improve the performance and reliability of federated learning to better address the challenges of data collaboration and privacy protection.

9. What nuanced considerations should be taken into account when applying NFL-CGAN to diverse data distributions in practical applications, given its outstanding performance across both IID and non-IID scenarios?

Response: NFL-CGAN demonstrates rapid convergence and stability in the federated learning environment, which is crucial for improving distributed learning scenarios. It signifies that in various real-world applications, models can learn faster from data distributed across multiple locations, accelerating decision-making and problem-solving processes. Additionally, by reducing the frequency and costs of communication, it makes federated learning more suitable for large-scale and efficient data collaborations, broadening its scope of applications. Therefore, the performance of NFL-CGAN is important in enhancing the efficiency and practicality of distributed learning.

10. The rapid convergence and stability of NFL-CGAN as evidenced by the training loss curves prompt reflection on its potential for improving model convergence in real-world distributed learning scenarios. What implications does this hold for accelerating learning in such settings?

Response: The fast convergence and stability of NFL-CGAN are significant in real-world accelerated learning scenarios. Fast convergence implies that the model can learn more quickly from multiple data sources, accelerating the process of decision-making and problem-solving. This is crucial in applications that require timely responses and rapid adaptation to constantly changing environments, such as financial market analysis or intelligent mobile applications. Furthermore, fast convergence can also reduce communication overhead, making federated learning more efficient in large-scale data collaborations. Therefore, the performance improvement of NFL-CGAN enhances the efficiency and practicality of learning, providing greater potential for various real-world applications.

Reviewer 3 Report

Comments and Suggestions for Authors

Summary of the Work

In this work, the authors proposed the Federated Learning framework (NFL) based on Conditional Generative Adversarial Networks which enhances data privacy protection and model classification performance. NFL-CGAN divides the local network of each client into private and public modules. The private module contains an extractor and a discriminator to protect privacy by retaining them locally.

Main Results Obtained

This approach

i) offers enhanced privacy protection against Deep Leakage from Gradients (DLG) attacks;

ii) outperforms traditional FL methods in model efficacy.

iii) Integrating Conditional Generative Adversarial Networks into Federated Learning protects client data privacy while maintaining good classification performance of the client models.

iv) improve the superior performance in maintaining privacy compared to FL baseline methods significantly

 

General Considerations

- The authors mentioned the presence of section 6 which should report the conclusions and summarize the main results of this study. However, they mentioned very (too) fleetingly what the concrete prospects of their future work are. They just mentioned the existence of the problem of data heterogeneity without, however, explaining how they intend to address this fundamental problem.

- English should be double-checked; some typos were found.

- The authors did not mention the potential limitations of using NFL based on cGANs for privacy preservation in 6G.

- The authors claimed that providing a method where a Federated Learning Framework is based on Conditional Generative Adversarial Networks for Privacy Preserving in 6G may avoid information disclosure, due to the exchange of shared gradients. To do this correctly, they need to provide evidence and analysis to support their claim. However, some key aspects have not been demonstrated or discussed exhaustively (see the suggestions below)..

- The list of references should be completed based on the answers to the previous questions.

The suggestions that follow are intended to help fill some (not all) gaps in this work.

 

Suggestions

1) It is recommended to expand the section Conclusions where it is possible to find not only the main conclusions but also the real prospects of this work, mentioning, albeit briefly, issues to be addressed and hinting at how to solve them.

2) The authors mentioned two issues in the NFL i.e.,

i) the DGL attacks;

ii) the limited training data available to clients and data imbalance issues. However, there are very important limitations of the NFL. For instance,

2a) FL assumes that the data across local devices are independently and identically distributed (IID). However, in real-world scenarios, the data may not be uniformly distributed across devices, leading to challenges in training a model that generalizes well.

2b) Devices in a federated learning setting may have different hardware capabilities, network speeds, and computing power. This heterogeneity can make it challenging to design a model that performs well across all devices. The author just mentioned in section 6. Conclusions that “in the future, their work will focus on addressing data heterogeneity issues to further improve model performance”.

The authors are asked to discuss more exhaustively the above issues by clarifying how NFL-CGAN may help alleviate these drawbacks.

3) In a federated learning setup, the overall training progress is often determined by the slowest or least capable device (straggler). This can impact the efficiency of the learning process, particularly when dealing with a large number of devices. Why doesn't NFL-CGAN have this drawback?

4) Using cGANs for privacy-preserving purposes in 6G or any other context may offer some advantages, but it is important to stress that no single technique can fully address all of the limitations of Federated Learning (FL). The authors are asked to recognize key areas where challenges persist. For instance, straggler issues are more related to the efficiency of communication and coordination in federated learning, and cGANs alone may not provide a comprehensive solution.

5) The authors stated that NFL-CGAN outperforms traditional FL baseline methods in data classification, showing its higher effectiveness. However, we may object that the effectiveness of cGANs in handling device heterogeneity depends on the complexity of the underlying device variations. Real-world heterogeneity may pose challenges that cannot be entirely addressed by synthetic data generation. The authors are invited to dispel this possible objection.

6) We come now to a crucial point. While cGANs have been used for privacy-preserving synthetic data generation, and Federated Learning aims to train models across decentralized devices without exchanging raw data, some challenges and considerations need to be taken into account. The main issues are:

6a) Training both cGANs and federated learning models can be computationally demanding. Ensuring stable training processes and avoiding issues like mode collapse in GAN training or convergence issues in federated learning is essential.

6b) While, as shown by the authors, the cGANs can reduce the need for transmitting raw data, there is still communication overhead in exchanging model updates between the central server and the local devices. Efficient communication protocols are necessary to minimize this overhead.

The authors are invited to precisely answer the above points 6a) and 6b) by making use of concrete experiments, for example through the experiments described and discussed in Sections 4. and 5. respectively.

 

Conclusions

The manuscript is interesting and timely. However, several points need clarification. I strongly advise authors to take the above suggestions into account. Comprehensive answers to the questions raised above will increase the reader's interest.

Comments on the Quality of English Language

English should be double-checked; some typos were found.

Author Response

Summary of the Work

In this work, the authors proposed the Federated Learning framework (NFL) based on Conditional Generative Adversarial Networks which enhances data privacy protection and model classification performance. NFL-CGAN divides the local network of each client into private and public modules. The private module contains an extractor and a discriminator to protect privacy by retaining them locally.

Main Results Obtained

This approach

1. i) offers enhanced privacy protection against Deep Leakage from Gradients(DLG) attacks;
2. ii) outperforms traditional FL methods in model efficacy.

iii) Integrating Conditional Generative Adversarial Networks into Federated Learning protects client data privacy while maintaining good classification performance of the client models.

1. iv) improve the superior performance in maintaining privacy compared to FL baseline methods significantly

 

General Considerations

- The authors mentioned the presence of section 6 which should report the conclusions and summarize the main results of this study. However, they mentioned very (too) fleetingly what the concrete prospects of their future work are. They just mentioned the existence of the problem of data heterogeneity without, however, explaining how they intend to address this fundamental problem.

Response： Thank you for your suggestions. The corresponding changes have been made in the 6. Conclusion section in response to your feedback.

 

- English should be double-checked; some typos were found.

Response：I have carefully reviewed and corrected typos once again.

- The authors did not mention the potential limitations of using NFL based on cGANs for privacy preservation in 6G.

Response：It has been added to the conclusion in Section 6.

- The authors claimed that providing a method where a Federated Learning Framework is based on Conditional Generative Adversarial Networks for Privacy Preserving in 6G may avoid information disclosure, due to the exchange of shared gradients. To do this correctly, they need to provide evidence and analysis to support their claim. However, some key aspects have not been demonstrated or discussed exhaustively (see the suggestions below)..

- The list of references should be completed based on the answers to the previous questions.

The suggestions that follow are intended to help fill some (not all) gaps in this work.

Suggestions

• It is recommended to expand the section Conclusions where it is possible to find not only the main conclusions but also the real prospects of this work, mentioning, albeit briefly, issues to be addressed and hinting at how to solve them.

Response：Alright, it has been modified in Section 6.

2) The authors mentioned two issues in the NFL i.e.,

1. i) the DGL attacks;
2. ii) the limited training data available to clients and data imbalance issues. However, there are very important limitations of the NFL. For instance,

2a) FL assumes that the data across local devices are independently and identically distributed (IID). However, in real-world scenarios, the data may not be uniformly distributed across devices, leading to challenges in training a model that generalizes well.

2b) Devices in a federated learning setting may have different hardware capabilities, network speeds, and computing power. This heterogeneity can make it challenging to design a model that performs well across all devices. The author just mentioned in section 6. Conclusions that “in the future, their work will focus on addressing data heterogeneity issues to further improve model performance”.

The authors are asked to discuss more exhaustively the above issues by clarifying how NFL-CGAN may help alleviate these drawbacks.

Response：Regarding how to address DLG attacks: the presence of an extractor helps mitigate the risk of direct exposure of data to the classifier by performing feature extraction before transmitting the data to the classifier. Additionally, clients only upload the model parameters of the classifier and generator to the server, which means that even in the event of gradient leakage attacks, attackers can only reconstruct the input data to the classifier based on classifier parameters because they cannot access the extractor's parameters, thereby effectively preventing privacy leakage effects caused by differential label gradient (DLG) attacks and enhancing privacy protection performance.

Response：Regarding the issue of limited and unevenly distributed training data: NFL-CGAN can address data imbalance problems by training generative adversarial networks on local devices. Generative adversarial networks can generate synthetic data, thereby increasing the number of samples and mitigating data imbalance issues. Furthermore, the design of private and public modules in NFL-CGAN allows for the aggregation of client-shared knowledge on the server while preserving client privacy, thereby improving model performance.

Response：Concerning device heterogeneity: The model structure we have designed is relatively simple and lightweight, with relatively low requirements on hardware and computational capabilities for different devices. Of course, in future research, we will make further efforts to optimize the lightweight components of the model to ensure it performs well in various environments. In summary, while these challenges exist, our work aims to alleviate their impact through innovative technologies and methods, further enhancing the performance and applicability of federated learning.

• In a federated learning setup, the overall training progress is often determined by the slowest or least capable device (straggler). This can impact the efficiency of the learning process, particularly when dealing with a large number of devices. Why doesn't NFL-CGAN have this drawback?

Response：This article primarily addresses the issues of uneven data distribution and privacy protection. The model structure we have designed is relatively simple and lightweight, with relatively low requirements on the devices. Therefore, NFL-CGAN does not have this drawback.

 

• Using cGANs for privacy-preserving purposes in 6G or any other context may offer some advantages, but it is important to stress that no single technique can fully address all of the limitations of Federated Learning (FL). The authors are asked to recognize key areas where challenges persist. For instance, straggler issues are more related to the efficiency of communication and coordination in federated learning, and cGANs alone may not provide a comprehensive solution.

Response：Federated learning faces various challenges. This article primarily focuses on issues like insufficient data volume, uneven data distribution, and privacy protection, offering corresponding solutions. As for other challenges in federated learning, such as client dropouts and client trust issues, the impact of the proposed approach remains unclear, and addressing these challenges will be part of our future research plans. We have provided additional clarification on this matter in the Conclusion section.

• The authors stated that NFL-CGAN outperforms traditional FL baseline methods in data classification, showing its higher effectiveness. However, we may object that the effectiveness of cGANs in handling device heterogeneity depends on the complexity of the underlying device variations. Real-world heterogeneity may pose challenges that cannot be entirely addressed by synthetic data generation. The authors are invited to dispel this possible objection.

Response：Regarding the issue mentioned in 5), when dealing with the diversity of real-world devices, we fully recognize that relying solely on synthetic data generation may be challenging to capture complex data features comprehensively. Therefore, our approach is not limited to synthetic data generation alone; instead, it combines various strategies such as differential privacy techniques and adversarial training to enhance the adaptability of the model to device heterogeneity. Specifically, our model is divided into private and public components on each client, allowing the model to adaptively learn based on local data characteristics without solely relying on a global model. Moreover, the introduction of differential privacy techniques adds noise during the training process, protecting data privacy while improving the model's robustness. Additionally, the application of adversarial training effectively enhances the model's resistance to interference. Although synthetic data generation is an essential component of our strategy, the comprehensive use of these diverse strategies aims to better address the issue of device diversity in the real world, thus improving the model's effectiveness and applicability. Of course, we are also aware that despite significant improvements in model performance, ongoing efforts are needed in future research to further refine and optimize the model to better meet various requirements in practical applications.

6) We come now to a crucial point. While cGANs have been used for privacy-preserving synthetic data generation, and Federated Learning aims to train models across decentralized devices without exchanging raw data, some challenges and considerations need to be taken into account. The main issues are:

6a) Training both cGANs and federated learning models can be computationally demanding. Ensuring stable training processes and avoiding issues like mode collapse in GAN training or convergence issues in federated learning is essential.

6b) While, as shown by the authors, the cGANs can reduce the need for transmitting raw data, there is still communication overhead in exchanging model updates between the central server and the local devices. Efficient communication protocols are necessary to minimize this overhead.

The authors are invited to precisely answer the above points 6a) and 6b) by making use of concrete experiments, for example through the experiments described and discussed in Sections 4. and 5. respectively.

Response：

Regarding 6a), the experimental results in our section show that our model has achieved quite satisfactory training results even before 20 rounds of communication. Since our model structure is relatively simple, the computational burden during the training process is moderate, resulting in a relatively stable model training process.

Regarding 6b), when we upload the model, we only upload the model parameters of the classifier and generator, while the parameters of the extractor and discriminator are kept locally. This not only protects data privacy but also to some extent reduces communication overhead. If minimizing communication overhead is a priority, effective communication protocols are indeed crucial and can be considered as part of our future research.

Conclusions

The manuscript is interesting and timely. However, several points need clarification. I strongly advise authors to take the above suggestions into account. Comprehensive answers to the questions raised above will increase the reader's interest.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

All my comments in the previous review have been addressed.

I would only suggest to make sure all acronyms are properly expanded in the first usage (e.g., "Kullback–Leibler (KL) divergence").

I do not have further comments.

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have addressed all my comments. Therefore, I recommend accepting the paper for publication in this esteemed journal.

Comments on the Quality of English Language

Minor editing 

Reviewer 3 Report

Comments and Suggestions for Authors

I appreciate the efforts made by the authors in answering the questions raised in my previous report. I think the model is still in its preliminary stage and requires further developments to improve its performance (e.g., it has to be able to face the data heterogeneity issues, etc. - see my previous report). Anyhow, I think that the present version of the manuscript deserves to be published.