kumoh national institute of technology
Networked Systems Lab.

Energy-Aware Two-tier Model for Edge-assisted Internet of Things (IoT) using CNN
By : Damian
Date : 2020-05-06
Views : 579

IEEE Internet of Things Journal


Dear Prof. Kim,

I am writing to you concerning the above referenced manuscript which you submitted to the IEEE Internet of Things Journal.

Based on the enclosed set of reviews, this manuscript has been rejected for publication. While the decision of the Associate Editor is not positive, you may wish to consider the reviewers' comments. The current decision does not prevent your future submission of a reworked manuscript, which would receive appropriate peer review as a new submission.

If you have any questions regarding the reviews, please contact the Associate Editor: Prof. Khan M. Iftekharuddin kiftekha@odu.edu. Any other inquiries should be directed to the Administrative Assistant: Mariola Piatkiewicz m.piatkiewicz@ieee.org

Thank you for considering IEEE Internet of Things Journal for publication of your work.

Prof. Honggang Wang
Editor-in-Chief, IEEE Internet of Things Journal
hwang1@umassd.edu, wanghbell@yahoo.com

• Associate Editor Comments, if any, are listed below:

Associate Editor: Iftekharuddin, Khan M.
Comments to Author:
The authors propose two-tier edge IoT network using CNN. The reviewers agree that the topic is interesting, however, the work lacks novelty. They also identify several technical issues with this work including lack of validation and evaluation, comparison with state-of-the art, and weaknesses in the results and analysis.

• Reviewer Comments, if any, are listed below:

Reviewer: 1

Comments to the Author
The authors consider a two-tier edge-IoT network, where IoT devices are at the first level and the edge nodes are at the second level.
In the 1st tier, the authors proposed using an algorithm for clustering and selecting the cluster head, which is response for data compression and offload the data to the edge node.
In the 2nd tier, a CNN model is deployed to work on the sensitivity level of the uploaded data.
The authors conducted the experimental results using the ns-3.26 simulator and it is shown that the proposed approach is superior to all the existing ones.

Although the topic considered in this work is interesting and timely, there are some major comments and concerns, which the authors should take care to further improve their work. Please check the list below for more detail.

1. It is stated in the introduction that "the existing clustering and routing algorithms have higher complexity even for a slight improvement in energy efficiency"; however, it is not clear to the reviewer.
The authors should mention clearly why the state of the arts have high complexity and how about when compared with your proposed algorithm in this work.

2. The motivation part is not clear to the reviewer. More importantly, I cannot see any novelty of this work and differences with the state-of-the-arts. The authors are suggested to significantly rewrite the introduction part and emphasize their contributions, and features and advantages of their work. Further, Section II (Related Works) should be written in the way to support the main idea of this work; otherwise, this section is meaningless.

3. What is exactly denoted by criteria in "The M Score is a combination of significant criteria". This part should be made clear; otherwise, the readers may have to guess what is criteria and how to combine them to create a unique metric for performing the clustering head selection strategy.

4. In Section II-B, the sentence "These works also focused on offloading the data processing task to the edge deviceshowever, they were unable to achieve energy efficiency among the IoT nodes" seems to be not correct. Lots of research works on Edge-IoT focus on improving the energy efficiency of end devices. They key point here is enabling the offloading from end IoT devices to the edge nodes, thus saving energy and improving the completion time.

5. The title of Section III is Problem Formulation, which is therefore expected to show some mathematical problem formulations instead of writing some paragraphs on the related works. I am not sure what is the purpose of this section since it seems to be duplicated from Section II and in part from Section I, where the authors present the main focus and contribution of this work.

6. Regarding the proposed approach, the reviewer has the following concerns:
6.1. the concept "Voronoi Cell-based Correlation Cluster Formation" and related terms like Voronoi is new to the reviewer and may be most readers. The authors should show the originality of this concept.

6.2. It is stated that a fully connected CNN model is proposed and deployed at the edge node to compress the data uploaded from the IoT devices. However, the reviewer cannot find the proposed model in the manuscript. Figure 3 shows a neural network model with the output of compressed data, but it is not mentioned at any point in the manuscript and the novelty of the proposed model is not showed and explained clearly.

6.3. In Section I, the authors noted that the existing clustering and cluster head selection algorithms have high complexity and at this point, the reviewer do think that the approach proposed in this work is a light scheme. But I cannot find any analysis for the computational and time complexity of the proposed approach. The authors should double-check this point.

7. All the figures are of low quality. The authors should use vector images instead of bitmap ones. More, some figures are not formally presented, e.g., Figure 4 (Simulation results of the proposed model) can be removed without any effect on the paper quality. Showing all what you have done is not a good way when doing a research work, instead the authors should merely show their novelty and contributions compared with the state of the arts.

8. Regarding the simulation results, there are some concerns:
8.1. There are some abnormal points in the simulation results. For example, in Figure 7, the trend is higher PDR when the number of IoT nodes increase, but the PDR of K-S Statistics with 30 IoT nodes does not follow this trend. Another example can be found in Figure 9.
It seems that the plot is achieve using a single-shot network setting. If it is a case, the authors should check the Monte-Carlo simulation.

8.2. The proposed approach is able to achieve better result than all the existing schemes in terms of throughput, energy efficiency, reliability, delay, and PDR. As far as the reviewer knows, there is a tradeoff between the throughput and energy consumption, same thing observes for other metrics. If you want to achieve something, you must to pay something, all is almost difficult to achieve simultaneously.

8.3. The authors are recommended to compared their proposed approach with the optimal solution as well as non-learning schemes. It would better verify the effectiveness of their proposed algorithm.

8.4. The experimental results should be validated with more numbers of IoT devices. This would be useful to apply the scheme proposed in this work for massive connectivity in IoT networks.

9. Minor comments
- Although the manuscript is almost accessible, the authors are still recommended to proofread their work since there are some typo and grammatical errors.
- The references should be formatted in a consistent way.

Reviewer: 2

Comments to the Author
The paper discusses an approach using 5 methods namely, cell-based cluster formation, multipart CH selection, redundant data elimination, optimal routing and ML-based data compression, to achieve energy efficiency in an IoT network. This method have been compared with 3 state-of-the-art methods and shown to have better performance in terms of energy, throughput and reliability.
There are a few points that the authors need to address in the paper.
Sensitivity level of the data speaks of a sensitivity range but it turns out you were making a binary choice of whether the data is sensitive or not.
Data Validation by Threshold Value DVTV): How is this threshold determined at the edge node and how does it relate to the sensitivity of the data?

Thus, the non-sensitive data needs to be compressed which results in better performance. – Better performance of what?

The F2CNN is fast and has the following layers:
And in the hidden layers there is the proposed F2CNN-LZW. It is not clear which architecture you are explaining in this

The LZW technique uses the code-book for data compressionWhat is the codebook being referred to here?

For data compression, a novel F2CNN-LZW lossless compression technique is presentedPlease explain what is the novelty in this technique that justifies the novelty claim. Also justify why novel is claimed for the GR-SimRank scheme and two-tier architecture.

Equation 18 and 19 does not make much sense. What is O in (18) and why do you have two different equations for determining the output. Is (19) really an equation to determine the output or it is the output (compressed data) itself?

What does fast mean for the F2CNN model and how is this measured and achieved?

How was the F2CNN-LZW network trained and with which datasets? Did you consider the complexity of the training and deployment? Part of the problem statement of the paper is that achieving energy efficiency is restricted by complex compression method. Neural networks introduce their complexity, and since this is employed for data compression, it would be useful to see, in your analysis and results, how this particular network has mitigated the issue of complex compression. There is a claim that with this method all data packets are processed in parallel and compressed rapidly but this is not substantiated in the result section. Also, how does this compression technique contribute to improving energy efficiency?

Fig. 1: If you can make it clearer with the arrows or numbering the stages of the algorithms if they are happening sequentially. For F2CNN-LZW, shouldnt the output be non-sensitive compressed data? But the arrow instead points to the sensitive data.

Table 3: What is the difference between data reduction and compression. You should make it clear what the definition of these two is, as they are sometimes used interchangeably. Compression is sometimes referred to as a data reduction technique.

Comparative analysis (Energy consumption):

You claim that the major reason for lower energy consumption is because the model uses edge computing to offload the data processing from the IoT nodes.
What is not clear is the contribution of each of the 5 methods in lowering energy consumption. What specific data processing tasks do the edge offload from the IoT nodes?

It is not very clear how the energy consumption in this paper is measured and how this fairly compares to the three state-of-the-art methods. Did you measure only energy consumption at the IoT nodes alone or among all the nodes in the entire network? If energy consumption in IoT nodes alone are measured, what is the essence of the data compression technique and its contribution to the overall objective of improving energy efficiency?

As the 5 methods are claimed to contribute to energy efficiency, a discussion of how each of these methods have individually contributed to improving the energy efficiency of the system is missing. This will further justify their use in addition to the already shown aggregated contribution to the system.

Reviewer: 3

Comments to the Author
The paper addresses a potentially interesting topic, well suited to the expected readers' audience, even if of not outstanding technical originality. While energy management in edge-enable-IoT is still a largely unexplored area, the presented solution is mainly a variant of the more traditional problem of route design, and compression method.
Anyway, I have several major concerns with the submitted paper:
1) Most relevant, the technical content and contribution of the paper are too limited for a high-quality IEEE Journal. The paper does not present at all the design, implementation, deployment, optimization, and evaluation work made on a real prototype of the proposed solution. The targeted research area is already relatively mature; it is not sufficient to propose an algorithm (which is a variation of existing solutions in the literature).
2) No validation and assessment of the reported simulation results with measurements from in-the-field deployment experiences, of course in simplified execution environments and topologies
3) No lessons learned from direct implementation and experimentation of a real prototype
4) The results have no comparison of the complexity analysis of the proposed algorithm with existing approaches. Since computation overhead is an important metric to evaluate the algorithm, the authors should add such a result.
5) While energy is a crucial metric indicators in edge-enable-IoT environment. However, the author did not mention what kind of energy model used in this work. The Authors should have clarified better and more explicitly how energy considerations apply only on the subset IoT node in the edge-enable-IoT environment. Given that the consideration energy is one of the few significant aspects of technical originality of the proposal, this weakness is not negligible
6) Be specific in the Abstract (1st half) in the problem of literature that has been addressed in this work. Make sure that all paragraphs length does not exceed 25 lines. In the 4th paragraph of the introduction section, it needs relevant recent literature for novelty verification with the recent 3 years existing works in the domain.

Reviewer: 4

Comments to the Author
1) The motivation of this paper should be discussed in detail.
2) The result section is very poor. Please revise the section with some new results.
3) The contributions of this paper are somehow novel, please explain it in detail.
4) The related work section should be updated.
5) There are some grammatical mistakes and typos, please revise the paper.

Note: If any of the reviews include an additional attached file, it will not be attached to this message. To access it, log in your Author Center and click on "Decision (View Letter)" -- this brings up this message, and at the bottom is the attached file. Click on that file name and you can see the attached comments from reviewers.

Date Sent:


© Clarivate Analytics | © ScholarOne, Inc., 2020. All Rights Reserved.