A Systematic Review on Deep-Learning-Based Phishing Email Detection

Round 1

Reviewer 1 Report

1. In section III, you introduced the previous related papers in a date descending way (from 2023 to 2017). Is it better to introduce these papers from 2017 to 2023, so readers can know the development history of these methods.
2. In 4.1, you classify papers into conference and journal papers. I’m not sure whether it is necessary to do this classification. If authors believe it is necessary, please state the reasons or findings in this section.
3. In 4.3, can you give more details for each research direction?
4. In the tables such as Table 2, you specified their limitations. Is it possible to add a column to specify their innovations or advantages?
5. In line 311, it should be “Table 2” not “table 2”.
6. In Table 5, several methods are characterized as deep learning. As we all know, deep learning has a large number of concepts. Could you please specify the typical neural network those authors implemented in their papers?
7. In section 4 or section 5, I suggest you add a subsection to discuss the algorithmic development and their future directions.

Author Response

Hello,

Thank you for your comments. Please see attached.

Best regards,

Author Response File: Author Response.pdf

Reviewer 2 Report

The authors of this manuscript collect and sort out relevant literature focusing on deep learning-based phishing email detection tasks, which assists in identifying the strengths and limitations of different deep learning methods utilized to distinguish text-based phishing emails with headers. This review first shows its contribution and the corresponding structure in the first part and then illustrates the strategies for selecting the articles that will be introduced in the third part, based on SLR in the methodology section. Then, the review presents various deep neural network models to identify phishing emails, including the datasets and the limitations of the models, in chronological order of publication. Finally, it analyzes in detail the limitations of this research field and possible future development trends. The review is clearly structured. However, some points need to be considered. Therefore, I can recommend publication of this manuscript with a few revisions below:

1.       I wonder the rationales why these articles are introduced in chronological order in Part III. Is it because in the field of using deep learning to solve phishing email recognition, there are very novel results every year?

2.       As you know, there are many kinds of phishing emails, such as emails with text attachments or web link attachments. Since what you pay attention to is mainly the text-based phishing email with headers, would it be better if you introduce more details briefly of the text-based phishing emails you concentrate on at the beginning of the review, because “phishing email detection” is a very wide research field? How many parts do the email consist of? Do they have any attachments? How many words do they contain approximately?

3.       To resist phishing emails, digital signatures will play an irreplaceable role in network security. Quantum digital signatures based on the principle of quantum physics can provide information theory security. The network based on quantum encryption and quantum digital signatures has been demonstrated in experiments [Natl. Sci. Rev. 10, nwac228 (2023)], which can further enhance the subject security of the work in this paper. I suggest the author discuss this point.

4.       As you know, many groups in the world that concentrate on either deep learning algorithm design or NLP applications will both be interested in deep learning-based phishing email detection tasks. So in Part III, would it be better if you give a brief introduction of the innovative points of the literatures. That is, if the literature focuses on designing novel neural network models, what’s the advantages of the new neural network models over traditional CNN and RNN, resulting in higher precision or recall? if the literature concentrates on the policies and skills in NLP applications, what’s the advantages of the methods? What is the embedding and encoding style of the word vectors? How do they use the self or cross attention?

5.       As you have mentioned, there are many factors that can be considered if we receive a phishing email. Are there anything wrong with our privacy information? Or what part of the website is attacked? In practical situations, if we receive a phishing email, what can we do to protect ourselves from receiving the same kind of email next time by deep learning methods?

6.       When using deep learning methods to identify phishing emails, your evaluation indicators are mainly accuracy, precision, recall and so on. Are there any more factors to consider if some noises are in the datasets, such as robustness, convergence, complexity.

7.       As for this article, you can have a look. Otter, D. W., Medina, J. R., & Kalita, J. K. (2020). A survey of the usages of deep learning for natural language processing. IEEE transactions on neural networks and learning systems, 32(2), 604-624. Artificial intelligence combined with quantum communication and quantum computing can bring additional effects to security. The author needs to introduce and discuss related works [Research 6, 0134 (2023); Opt. Express 30, 15024 (2022)] to enrich the completeness of this manuscript.

Moderate editing of English language required

Author Response

Hello,

Thank you for your comments. Please see attached for full response.

Best regards,

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

This manuscript can be accepted for publication.

Minor editing of English language required