Cluster-Based Secure Aggregation for Federated Learning

Round 1

Reviewer 1 Report

In this paper, we proposed a new cluster-based secure aggregation (CSA) strategy for federated learning using heterogeneous devices. The proposed CSA guarantees 91% learning accuracy while reducing the total learning time in situations where the size of training data, the computing power of each node is different, and the response time of  each node is also different. The CSA clusters nodes with similar response times and performs cluster-by-cluster aggregation of local updates on nodes. So the scope of the paper is very interesting. The novelty is clear and well described. So generally my opinion about the paper is positive.

But I recommend the following issues to be extended before the final decision:
-> The abstract is long. But it does not contain the main results presented in a quantitative way. So please revise. 

-> the literature review should be revised. In the present form, many papers are cited as multi-citation without deep discussion. Please revise eg. "Masking techniques to protect each node’s local result end [5-6], methods to effi- 50 ciently handle dropout nodes and reduce communication overhead [19-20], and studies 51 to detect Byzantine users [21-22] have been conducted" Also please see that papers 7-18 are missing in line 50. I see that some of them are discussed in section 2 but please revise to be corret.

-> also I recommend adding at least 2-3 papers from 2023 year to assure that a literature review is present.

-> Between section 3. A Cluster-based Secure Aggregation Model and subsection 3.1 Background and Configuration please add a general introduction to section 3.

-> Generally please discuss deeper the features of cluster-based secure aggregation that make a suitable method to solve the investigated research problem. 

-> Section  6 should be extended with the limitation of the proposed approach. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

In the list of contributions claimed by the author to be checked, few things are classical methods, not the author's own methodology

For edge, note how the roots are assigned to the cell

What is the need for rounding the parameter to an integer? Arthur's requested to justify

the rounding of may results in loss of learning because the rate values represent the amount of learning on a particular  feature

The communication overhead Hindi proposed methodology is not defined. the proposed method actually increases the communication overhead, which is the trade of the proposed method, which is not reported

The simulation platform and tool used for the simulation need to be added

What is the data source for every node? Need to be added. to exploit the weightage of aggregation, the information about different amount of data at different node need to be addressed in simulation shut up

What is the target application for which the model are trained? Whether those application targets are met or not needs to be addressed

Running time performance is not universal. It is based on the computational platform utilized. without the information of computational platform or which runtime performance is measured, the results are incomplete are not useful

 

Figure 7 and 8 results are helpful only if the computational platform and simulation  environment is defined

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Paper can  be accepted now.