An Asynchronous Federated Learning Aggregation Method Based on Adaptive Differential Privacy

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

In this manuscript, the authors summarized and discussed an asynchronous federated learning aggregation method using adaptive differential privacy to balance privacy protection and model accuracy. On this basis, they further investigated a Newton’s cooling law-based privacy budget allocation that dynamically adjusts ε per training round, a dual-weighted aggregation scheme incorporating device freshness and privacy budgets, and the ADP-FL algorithm integrating these mechanisms with Gaussian/Laplace noise. This work provides new insights into the development of privacy-preserving distributed learning systems by achieving higher accuracy and lower communication overhead under identical privacy budgets. I would suggest accepting it after the following minor concerns are addressed：

1. For the summary section, it is suggested to add the main idea of the proposed algorithm. For example: The algorithm name（ADP-FL） was not clearly stated, and the part about "higher accuracy and lower communication costs" did not include any data comparison or data explanation.
2. The literature in this paper is too old as a whole (such as "2. Related Work"). It is suggested to supplement the latest research progress in related fields in the past three years.
3. In the "5.Experiment" section, the datasets used in this paper are all small datasets, and it is recommended to supplement large datasets for performance evaluation experiments of this work and comparative work, such as CIFAR-100.
4. In the "5. Experiment" section, The model architecture (RNN/VGG9/CNN) lacks modern architectures (such as Transformer) , making it difficult to prove the effectiveness of the algorithm on complex models.
5. The data in Table 2 is inconsistent with that shown in Figures 4, 5, and 6.
6. The text in Appendix A and Figure 7is too small to be seen clearly.
7. In the experiments of this article, only the accuracy rates of the algorithm on different datasets were considered. It is suggested to add indicators such as Precision, Recall, and F1 to analyze the impact of the algorithm performance on different datasets.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript presents a novel approach integrating adaptive differential privacy with asynchronous federated learning. The authors propose an adaptive parameter-based differential privacy algorithm combined with weighted aggregation to balance privacy preservation and model accuracy effectively. The contribution aligns well with current challenges in federated learning, particularly addressing trade-offs between data privacy, model accuracy, and efficiency

Overall, the work addresses a relevant and timely topic.  However, certain aspects require further clarification and improvement.

Authors are required to address the following (minor issues):

• The manuscript would benefit from a more detailed theoretical justification for selecting Newton's cooling law to dynamically adjust the privacy budget. Providing supporting references or comparative analysis with alternative methods would further strengthen the rationale and clarify the advantages of this approach.

• Authors are required to clarify the hyperparameter selection strategies (e.g., α coefficient, λ coefficient, privacy budget ϵ, and relaxation parameter δ) would significantly enhance reproducibility and practical applicability.

• The description of experimental settings (e.g., computational setup, software frameworks, and detailed model architectures) could be further elaborated to improve reproducibility.

• Similarly, the discussion regarding real-world deployment scenarios is relatively limited. It would be beneficial to elaborate further on practical considerations, including computational overhead, network latency, communication costs, and evaluations involving real-device implementations.

• Figures and tables could benefit from enhanced clarity with improved legends and consistent formatting.

 

Comments on the Quality of English Language

• Proofreading for consistency in tense, subject-verb agreement, and punctuation would enhance readability.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf