Deep Anomaly Detection Based on Variational Deviation Network
Round 1
Reviewer 1 Report
The research study entitled “Deep anomaly detection based on variational deviation network,” introduces a new anomaly detection model, which is variational autoencoder model on the basis of the deviation network, to cope with the weaknesses and limitations of the existing deep anomaly detection methods. The anomaly detection framework is designed in such a way that it initially learns the normal distribution of normal data via variational self-encoding as a reference score, and in order to learn anomaly scores, it then utilizes another neural network.
In the reviewers’ opinion, this study should be of interest to the technical community of deep learning, looking for deep anomaly detection techniques, and their engineering applications. In addition, this manuscript is well-written and well-organized, and this is easy to follow. Figures, tabled, and supportive discussions are informative. However, there are only few concerns that I would like to ask the authors address them before it can be recommended for publication in the journal of Future Internet, MDPI. A revision should be made by considering the following comments:
- In the abstract, it would be better to highlight and specifically stress what has been done in this research work. The present abstract fails to accurately reflect the novelty of the proposed framework. In addition, please clearly explain the need for the current approach.
- The explanation related to the algorithm 1 and 2 are not enough.
- The authors are suggested to consider the expansion of the introduction so as to more highlight the challenging issues in this research field that motivated the authors to carry out the current study. In addition, please make the introduction section clearer.
- More details and references concerning variational auto-encoder and autoencoder discussed in section 4.3.1
are required. Moreover, the paper shows a lack of recent references. Some equations and formulations in the text need references and/or mathematical demonstrations.
- There are some grammatical mistakes and English typos throughout the manuscript. The authors are required to thoroughly revise the paper and correct the mentioned errors.
- To sum up, in the reviewer’s opinion, this study has a great potential that can open a very interesting line for future research of the authors, however it still has some important works to be published as they mentioned in the last paragraph of the conclusion. Therefore, I encourage the authors to continue working on the subject.
Comments for author File: 
Comments.pdf
Author Response
Please see attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
The manuscript sounds technically poor, I have following concerns should be addressed before any decision. The paper currently need revision.
*The existing literature should be classified and systematically reviewed, instead of being independently introduced one-by-one.
*The abstract is too general and not prepared objectively. It should briefly highlight the paper's novelty as what is the main problem, how has it been resolved and where the novelty lies?
*The 'conclusions' are a key component of the paper. It should complement the 'abstract' and normally used by experts to value the paper's engineering content. In general, it should sum up the most important outcomes of the paper. It should simply provide critical facts and figures achieved in this paper for supporting the claims.
*For better readability, the authors may expand the abbreviations at every first occurrence.
*The author should provide only relevant information related to this paper and reserve more space for the proposed framework.
*However, the author should compare the proposed algorithm with other recent works or provide a discussion. Otherwise, it's hard for the reader to identify the novelty and contribution of this work.
*The descriptions given in this proposed scheme are not sufficient that this manuscript only adopted a variety of existing methods to complete the experiment where there are no strong hypothesis and methodical theoretical arguments. Therefore, the reviewer considers that this paper needs more works.
*Key contribution and novelty has not been detailed in manuscript. Please include it in the introduction section
*What are the limitations of the related works
*Are there any limitations of this carried out study?
*How to select and optimize the user-defined parameters in the proposed model?
*There are quite a few abbreviations are used in the manuscript. It is suggested to use a table to host all the frequently used abbreviations with their descriptions to improve the readability
*Explain the evaluation metrics and justify why those evaluation metrics are used?
*Some sentences are too long to follow, it is suggested that to break them down into short but meaningful ones to make the manuscript readable.
*The title is pretty deceptive and does not address the problem completely.
*Every time a method/formula is used for something, it needs to be justified by either (a) prior work showing the superiority of this method, or (b) by your experiments showing its advantage over prior work methods - comparison is needed, or (c) formal proof of optimality. Please consider more prior works.
*The data is not described. Proper data description should contain the number of data items, number of parameters, distribution analysis of parameters, and of the target parameter itself for classification.
*The related works section is very short and no benefits from it. I suggest increasing the number of studies and add a new discussion there to show the advantage.
*Use Anova test to record the significant difference between performance of the proposed and existing methods.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
This paper proposes a new anomaly detection model, which learns anomaly scores through variational deviation networks.
A small amount of anomaly data and a reference score generated by variational self-encoding are used to force the anomaly score to have a significant deviation from the normal data.
The study has an interesting applied value. However, major concerns about the paper are the following:
- The analysis of related work is minimal and lacks recent studies on deep learning models for anomaly detection including autoencoders, which should be included (see DOI:
10.3390/fi12100167, 10.3390/s22010123, 10.1109/ICMLA.2019.00285, 10.3390/s22031244, 10.1007/978-3-030-88942-5_36)
- The experimental setup is not clear. Are datasets split in training, validation, and testing sets? In which percentage? Which hyperparameter values have been used for each method?
- Experimental Results: Table 1 does not respect page boundaries. It should be broken up in two tables, or rotated. The best results for each dataset could be emphasized in bold.
- In general the redaction of the manuscript appear rushed and could beneift from a professional proofreading iteration.
I recommend the authors to prepare a revised version of the manuscript
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The authors have addressed most of my comments and have highly improved the paper’s quality. Therefore, in my opinion the current manuscript fits the standards to be published, and the presentation of the problem at hand is clear. In addition, this study has interest from an academic viewpoint. Therefore, the current manuscript can be accepted for publication in the journal of Future Internet, MDPI publication.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 2 Report
- Please explain in your captions of figure and title of table, why are these tables or figures necessary in your paper? What are the purposes and what are the message you want to deliver via these figures and tables?
- The current metrics might not be sufficient to judge the performance of the model holistically. Please enhance the result analysis part of your paper.
- What are the limitations of the related works
- Are there any limitations of this carried out study?
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 3 Report
The authors addressed most of my concerns. The paper quality has certainly increased. My only concern is that even though the authors claimed that the literature review was extended, I do not see any added reference. I am afraid that it could be due to the latexdiff tool used to export the new version of the paper, which sometimes does not export the updated references. Unfortunately the version without annotations was not provided. I would be glad to accept the paper if my the literature review is correctly extended as required. For convenience, I report the links to the suggested publications from the previous round of reviews:
https://www.mdpi.com/1999-5903/12/10/167
https://www.mdpi.com/1424-8220/22/1/123
https://ieeexplore.ieee.org/document/8999323
https://www.mdpi.com/1424-8220/22/3/1244
https://link.springer.com/chapter/10.1007/978-3-030-88942-5_36
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Round 3
Reviewer 3 Report
The article is in perfect shape for publication. I recommend it’s acceptance as is.
This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.
