Local Differential Privacy Based Membership-Privacy-Preserving Federated Learning for Deep-Learning-Driven Remote Sensing

Round 1

Reviewer 1 Report

Authors propose LDP-Fed technique embedding Local Differential Privacy (LDP) in Federated Learning (FL), to defend against white-box MIA.

The technique is very interesting. Results are convincing too. But I guess that the description of the method requires a major revision.

In fact, authors describe the method using the pseudo code in Algorithm 1 and Algorithm 2.

-Authors missed to separate "Server executes" part from "Client executes" part. In particular, "Client executes" is missed in Algorithm 1.

-Who decides the epsilon value? Is it cloud based or is it decided at server level? It is not clear in the Algorithm 1;

-What is x in Algorithm 2?

-It is not clear the matching between rows in Algorithm 2 and subsections. In my opinion authors should include any computation (also l(omega) and r(omega) in the Algorithm 2 section and then explain what they are (and technical explanation) in the paragraphs.

-I do not understand the meaning of [-C,l(omega)) union (r(omega),C] in row 5 in Algorithm 2... too many round brackets?

-row 306: what is Section V-D ?

 

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

1. As a paper on remote sensing image classification, I have not seen any remote sensing images or the results of image classification.

2. The paper introduces a large number of principles and methods, but it seems to have little to do with remote sensing image classification.

3. In 2 Materials and Methods, there is no introduction of experimental data, and it is not known what the experimental data in this paper is, nor what the learning samples are.

The English language still needs improvement. A paper usually uses the third person, but the paper uses the first person. For example, the author wrote in the abstract:“We conducted comprehensive experiments to evaluate the framework’s effectiveness on one remote sensing image dataset and three machine learning benchmark datasets.”

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

This article has mainly considered embedding local differential privacy (LDP) into the FL and proposed LDP-Fed to defend against white-box MIA. This paper was well organized, there are some concerns should be addressed before it can be accepted in this journal.

1.      For the experimental dataset, only one group is the remote sensing image dataset, but the title of this article is “... for Deep Learning Driven Remote Sensing Image Classification,” I feel it's not appropriate and the authors should add more remote sensing image datasets for testing.

2.      The author can better demonstrate the method process using the image data, and it is best to present typical images with experimental data.

3.      In terms of experimental results, there is a lack of horizontal comparison, only comparing oneself with oneself. The author should add to prove the effectiveness of the method.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Authors answered to my previous concerns.

Now it is worth of publication.

Reviewer 3 Report

I have no more concerns.