Open-Set Specific Emitter Identification Based on Prototypical Networks and Extreme Value Theory
Round 1
Reviewer 1 Report
- The presented work is very interesting. The main results should be more exposed in the introduction.
- The optimization function or criteria should be well explained
- The fig. 9 is lacking more explanation.
- Authors can refers possible applications for RFID
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
In order to address the problem of particular emitter identification in open-set situations and further increase the recognition accuracy and resilience, this research offers an open-set recognition model based on prototypical networks and extreme value theory. In order to recognize I/Q signals, a one-dimensional convolutional neural network is first created. A squeeze and excitation block with an attention mechanism is then added to the network in order to effectively enhance the weights of the feature channels. In the meanwhile, channel shuffling and group convolution help to enhance recognition. The network is then trained using a joint loss function based on prototype learning to finish decoupling intra-class signals from inter-class signals and aggregating them in feature space. The paper needs some revisions:
-literature review is too weak; the introduction of references should be more specific
-add a paragraph in introduction to highlight what is new in this paper
-the structure of CNN, layers, and also its optimization are unclear
-add some graphical evidence for stability such as Lyapunov and its derivative
-the Robustness need some graphical evidence;
-add some remarks on how your method can be improved using type-2 and type-3 fuzzy systems such as:Modeling renewable energy systems by a self-evolving nonlinear consequent part recurrent type-2 fuzzy system for power prediction;
-add some comparisons with related methods
-add a pseudo code for the main algorithm, to be more clear
-some basic equations need references
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
This article presents a novel protocol for the open set recognition model based on prototypical networks and extreme value theory. This system is relatively simple and accurate and its operating principles are well demonstrated. Its use may be of interest to a great number of applications among researchers studying convolutional neural networks (CNN). From a technical point of view, this article is highly interesting for the CNN and the Weibull models on recognition performance and robustness. Besides, the obtained results and backgrounds are quite satisfactory. The manuscript is well-organized, and the subject is well-studied. Finally, the paper is well written.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report
The authors employ prototypical networks and extreme value theory in order to recognise devices that do not exist in the database (applications like automatic adding new user or preventing unauthorised access to a particular network). The method is based on, somehow with some additional proccess in the network, separation of intra-class signals as well as aggregation of inter-class signals in constructing feature space.
The proposed method seems to have novelty at some level. Howerver, two major concerns: first of all, the device number is quite limited which makes performance of the method questionable. On the other hand; performance of the method is quite parameter-dependent which makes the method less applicable, or exploitable to other applications (see Lamda effect in fig.6 and r in fig.8).
I have the following additional concerns regarding the method as well as the organisation of the paper.
1. Eqn(1) is either incomplete or unclear. It is not an equation, and does not reflect what it refes to represents! Please, give a reference and correct it. Moreover, p(y|x) is rather probability of y given that x, where there needs to be specific value for y?! (this is repeated right after eqn.8 as well)
2. Fig.1(a) and (b) do not give anything new. Weilbull distribution is quite common. Also prob. density-> prob. density function. "fx2 () represents the probability of the extreme value appearing in −ï‚¥x [ , ] " is not meaningful. In Fig.2 "Weilbull CDf probability" is meaningful as well (must be Weilbull CDF). Please, revise the concepts regarding Weilbull distribution.
3. Eqn(6) as well as the following text is not clear. Give a reference or justify how and why you revise the probability. This could be described rather hypotehsis testing. It is not clear as is. It is highly important to justify the method (somehow correct logicallay but must be justified to be applicable to wide range of practices or similar problems!).
4. Give more details about the devices in sect. 4.1. Segmentation of the records needs to be justified. Why do you choose 9 sub-segments? How realistic is the Gaussian noise you used, why dont you use channel noise?
5. The performance of the method should more elaborated by comparing similar methods (fig.10-12 needs to be discussed along with similar RFF/SEI papers, how well your methods work, or why not)
Finally, and more importantly, the reference is very limited and ignored some important, in particular MDPI journals, contributions which makes the paper non-comparable with those targeting similar RF fingerprinting or specific emitter identification methods as well as make positioning the paper in the literature. Please, take a look the following papers, and discuss them in the intro section.
1.An Identity Authentication Method of a MIoT Device Based on Radio Frequency (RF) Fingerprint Technology , Sensors.
2. Radio Frequency Fingerprint-Based Intelligent Mobile Edge Computing for Internet of Things Authentication, Sensors.
3. On the Performance of Energy Criterion Method in Wi-Fi Transient Signal Detection, Electronics.
4. Device fingerprinting using deep convolutional neural networks, Int. Journal of Comm. Networks and Distr. Systems.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 4 Report
The authors have addressed almost all my concerns, and the paper is highly improved.
One point could still be more elaborated based on authors' own view: the comparison in sect.4.3 may still be somehow among component level variations of EVT. There may be other methods with which the proposed method can be compared, or evaluated. If it is not possible, it is better to look at the papers considering RF fingerprinnting with similar emitting (zigbee) and signal capturing (USRP), and comment quantitatively on what accuracies are achieved with similar emitting devices.
Please, consider this.
