CamoNet: A Target Camouflage Network for Remote Sensing Images Based on Adversarial Attack

Round 1

Reviewer 1 Report

In this paper, the authors propose a novel object camouflage network to enhance the performance of gradient attacks on object detection. Experiments on DOTA and DIOR show that this approach can further reduce the range of disturbance while ensuring the attack effect, making the attack more concealed and imperceptible to the naked eye. This research is meaningful and has certain innovations.

The paper is well written and the logic is clear. A few minor revisions are list below:

1. Some writing errors must be corrected, such as the keyword 'Target camouflage.)' in line 10.

2. In the introduction section (lines 81 to 89), it is recommended to cite important references to make the argument stronger.

3. The formula numbers are messed up, formulas 3 and 4 appear twice.

4. What are the disadvantages of CamoNet?

5. There are too few keywords to cover the content of the paper, it is recommended to add some keywords.

6. It is recommended to cite the following two papers. The following two papers are the latest literature on the application of convolutional neural network and transformer in the field of remote sensing.

-- Kokila, S., Hybrid Behrens-Fisher- and Gray Contrast–Based Feature Point Selection for Building Detection from Satellite Images

-- Ghasemloo,  Estimating the Agricultural Farm Soil Moisture Using Spectral Indices of Landsat 8, and Sentinel-1, and Artificial Neural Network. 

Minor editing of English language required

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper focuses on topic of object detection from remote sensing images by CNN. Generally speaking, this is a high-quality paper, with distinct innovations and reports of promissing detection results. My suggestions to revise the paper are as follows:

1) One or two more key words should be added.

2) Object detection from remote sensing images by deep learning methods is very hot topic in recent years, so, consider to update the references or to cite more newly published papers.

3) How to understand the sentence "CNNs are fragile and can be easily fooled"? Try to explain it more in the introduction section.

4) Some variables are not defined or introduced, for example, S in Eq.(2)

5) Line 251, "We list the AP of each category in the table." Which table? What is AP? Is it average precision as it defined in Line 268? This paragraph is to introduce the original datasets, it is not suitable mention issue of detection precision.

6) Give out the full name of each class in Fig.5 and 6, not just abbreviation, and how many image does the DIOR dataset have?

7) There are three main parameters which should be initialized, "maximum iterations T; warm up iterations W; score threshold t", try to discuss the influences of one or two of the above parameters on the detection performance.

This paper is well-written.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The subject is new and exciting with the potential to be integrated not only into military activities but also into simulation for the environment

Figure 2 fits better in the methodology section as a workflow chart

Row 247: Algorithm or Figure 1

For me there are some problems with the article structure, please be patient and try to do the following:

1.            Move the description of your algorithm from the introduction to the methodology section

2.            Move the description of preview-related works from methodology to the introduction section

3.            Better describe the objective of the study and hypothesis in the Introduction section

4.            For the Results section keep only the experiments results and the algorithms better place them in the Methodology

5.            Introduce a discussion section where to evaluate the results, especially the Experiment results

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

The submitted manuscript is interesting, but can be further improved upon.

Specific comments:

Line 15 - PNSR abbreviation used before it was fully specified
Line 19 - random parentheses after the keywords
Line 31 - not clear what "It" refers to
Line 40 - space missing between "et al." and reference [11]
Line 96 - did you mean "boost" instead of "exert"?
Lines 122 and 248 - a section usually should not immediately start with a new subsection, but rather have some sort of introduction
- Line 214 - "The purpose of target camouflage is to deceive the detector and make the intelligent interpretation system fail" --> did you try to apply any of the image forensic techniques to the generated images to check if they can detect changes to the image? There are certain algorithms that can detect image manipulation.
- Line 224 - what threshold selection methods have you tried, and what results did you obtain? Why did you decide to use mean plus standard deviation?

General comments:

- Figure 1 - DSI abbreviation used before being fully specified
- Try to move figures closer to the text that refers to them
- A large portion of the Introduction section where the authors discuss previous work should be moved to the Related work section
- Related work section of the paper should be expanded, as it is not thorough enough and right now feels like an afterthought
- In the first paragraph of Section 3 the authors mix various English tenses (past and present, e.g. "we will perform" and "we eliminate"). The authors should, for the sake of clarity, use one or the other
- In section 3, "Gradient Attack Phase" and "Postprocessing Phase" should be subsections
- In Figure 4 please add raw input images before the detection results. It is difficult to judge detection results if the input image cannot be seen
- Discussion on the DOTA and DIOR datasets should perhaps be moved to the Related Work section
- Specify which image categories do the DIOR and DOTA datasets contain. In Figures 5 and 6 you use abbreviations but it is not clear what they actually represent

The English language is great for the most part; only minor changes are required as described in the comments.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The authors have made the changes as I suggested.

Author Response

Thank you again. 

Reviewer 3 Report

Review form 2.

The authors improved substantially the paper design and know after an extensive check of grammar language and minor changes in text are ready to be published.

Lines 32 and 50, Both Phrase begin with the same word, please replace one

The place of datasets used belongs of Methodology section before the Algorithm 1,

Please have double look at the paper once that is publish nothing can’t be change, make it flagship for your research. 

Comments for author File: Comments.pdf

Author Response

Thank you again. We have corrected the errors you raised and have a double look at our paper.

Reviewer 4 Report

The authors significantly improved their manuscript and kindly and thoroughly responded and addressed all of the comments I have made previously. I think that their manuscript should be accepted for publication.

Minor edits: in the first sentence of chapter 5, "We" should be written with a lowercase letter.

Author Response

Thank you again! We have corrected the error you raised.