Multi-Attack Detection: General Defense Strategy Based on Neural Networks for CV-QKD
Round 1
Reviewer 1 Report
The paper presents two neural network models for the detection of multiple attacks to CV-QKD systems. An experimental setup is also used to collect training data.
A sound justification of the advatanges in identifying that a combination of multiple attacks is performed instead of a single one could clarify the significance of the work. Also a comparison with the performance of single-attack models in the literature could be useful for the same purpose.
The long discussion on preprocessing at the end of the introduction might be moved in the following sections. Furthermore, since real-time operations are discussed, the Authors could better describe the actual operations, including the speed at which the preprocessing can be carried out in the inference phase.
In general, the experimental setup could be more thoroughly presented, including a discussion about operation speed, and the choice of sample length and type could be justified. The above discussion could improve the significance of the presented results.
In Sec. 2.1, sufficient details should be provided in the main text to present the equations (2) and (3). Moreover, the appendices contain some material that could be in the manuscript and they could be ordered according to their citations.
The comparison between LP-NN and BR-NN could be moved after Sec. 2.3.2, where BR-NN is yet to be discussed.
PM is mis-spelled in Figure 1.
Some acronyms are not defined (e.g., KNN, SVM, ...)
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
In the manuscript entitled “Multi-attack detection: General defense strategy based on neural networks for CV-QKD”, Hongwei Du et al. propose two multi-attack neural network detection models to handle the coexistence of multiple attacks. The comments are given as follows:
- The output of BR-NN model should have a total of Table 2 only describes part of the situations represented by the output. It is suggested that the author explain how to deal with the contradictory situations such as output [1 1 0 0] when testing the model.
- The main reason why the two networks in this article can accurately identify hybrid attacks (such as CA&SA) is that the multi-label learning method is used, while the same method is not used in [29]. Therefore, it is unfair to directly compare the network performance in Section 3.1 with [29].
- Line 82, formula (2), should be changed to .
- There are some spelling and expression mistakes in the manuscript, such as:
- Line 418, "knowns" should be changed to "attacks".
- “Droupout” in Fig. 3 should be changed to “Dropout”.
- “Testning” in Fig. 4 should be changed to “Testing”.
- Line 192, “vector” should be changed to “vectors”
- It is not recommended to use curly braces for both vector and set symbols. Curly braces generally represent sets and are unordered
- Several expressions are colloquial or grammatically incorrect, such as "eight outputs" in line 234 and "output is 64 neurons" in line 264.
[29] Mao Y, Huang W, Zhong H, et al. Detecting quantum attacks: a machine learning based defense strategy for practical continuous-variable quantum key distribution[J]. New Journal of Physics, 2020, 22(8): 083073.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
In this manuscript, the authors present a work that use neural network for detecting attacks in CV QKD system. I can recommend its publication in Photonics after the following comments are cleared:
1. In page 2 line 59, the authors mentioned “once the system receives abnormal data, it can immediately stop the key transmission with Alice without waiting until the key transmission is completed to check whether it is attacked.” However, since the method proposed in this paper is an ad-hoc method, which is not connected to the security analysis. How should the user coos a suitable threshold to stop the system operation?
2. In section 2.1, the authors explained the preparation of the data set. Why do the authors choose this configuration, for example 500 data as the original dataset, and 25 pulses in each group? Does the ability of attach recognition highly relates to the configuration selection?
3. Is the effectiveness of the proposed method affected by the finite-size effect?
4. The unknown attacks mentioned in the manuscript are combinations of the attack (hybrid attack) models studied here right? If so, the authors should make it clear in the main text, especially in section 3.3.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The Authors successfully addressed most of the comments, however some remarks still hold, as detailed below.
* The appendices A, B, and C could be ordered according to their citation in the manuscript text. Furthermore, Eq. A23 is referenced three times in the manuscript, but cannot be understood without the full explanation available only in the Appendice.
* Concerning the experimental setup, the operation speed at which the optical system in Fig. 1 works could be discussed. Moreover the compliance of this setup with the real QKD system receiving attacks could be better detailed.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
