A semi-supervised game model classification method and device based on bispectrum feature representation

Abstract

The application discloses a semi-supervised game model classification method and device based on a bispectrum feature representation, and the method comprises the following steps: acquiring an original received signal and an original received signal sample of a transmitting stage; inputting the original received signal into a trained semi-supervised game model for data signal processing; wherein the trained semi-supervised model is obtained by training the frequency order of the original received signal sample based on a high-order spectrum feature representation; and based on the data signal processing, the original received signal is classified by using a classifier to obtain a classification result. The high-order spectrum feature is used as the frequency order representation of the signal, so that the end-to-end semi-supervised classification is realized, and a higher evaluation score and classification accuracy are shown on a real data set and a simulation data set.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wireless communication, in particular to a semi-supervised game model classification method and device based on bispectrum feature representation. BACKGROUND

[0002] Radio frequency fingerprint identification (RFFI) is widely used in Internet of Things device authentication by measuring the characteristics of received electromagnetic signals to distinguish the source of the transmitter and associate with the relevant platform. The RFFI scheme based on feature extraction requires the selected RFF to have stability and reliability, so the RFFI method based on steady-state signals has more practical significance, mainly including high-order spectrum, Hilbert spectrum, power spectral density, multi-dimensional approximate entropy, wavelet packet, etc. In recent years, the development of deep convolutional neural network (CNN) has made the RFFI field develop rapidly. However, the essence of CNN is supervised learning, and its recognition accuracy depends on learning the feature distribution from a large number of labeled samples. However, in actual RFFI applications, there is a huge asymmetry between the scale of the labels and the number of samples from Internet of Things devices, and complete labeling of data seems too harsh, which limits the improvement of model performance. Therefore, semi-supervised learning can be embedded into the RFFI framework to optimize the training process of the deep learning system. GAN is a game model composed of a generator and a discriminator, which has achieved good results in data enhancement. The discriminator is responsible for binary classification of the output of the generator and the real sample, and the generator adjusts its parameters to deceive the discriminator through the gradient information of back propagation. ACGAN is used to expand the data set, and AlexNet is selected as the classifier to realize the modulation classification of communication signals. The use of GAN for underwater acoustic communication signal enhancement reduces the computational complexity of manual feature extraction. However, the above method only uses GAN for data enhancement and does not truly realize end-to-end classification of modulation signals, and the recognition range is limited to modulation methods, and the identification of individual transmitters has not been involved. SUMMARY

[0003] The present application aims to at least partially solve one of the technical problems in the related art.

[0004] To this end, the present application aims to propose a semi-supervised GAN framework based on bispectrum feature representation, which embeds bispectrum as RFF into the framework, so that the classifier can obtain more characteristic information about the transmitter while directly classifying the original signal. We introduce residual units of different scales in the network to solve network degradation, and extract multi-level representations of the signal and directly complete semi-supervised classification. The data is tested on simulation and real data, and the results show that the performance of the proposed framework is significantly improved.

[0005] Another object of the present application is to provide a semi-supervised game model classification device based on bispectrum feature representation.

[0006] To achieve the above object, in one aspect, the present application provides a semi-supervised game model classification method based on bispectrum feature representation, comprising:

[0007] The original received signal and the original received signal sample of the transmitting stage are acquired, and the original received signal is input into the trained semi-supervised game model for data signal processing, wherein the trained semi-supervised model is trained based on high-order spectrum feature representation for the frequency order of the original received signal sample; based on the data signal processing, the original received signal is classified by using a classifier to obtain a classification result.

[0008] The semi-supervised game model classification method based on bispectrum feature representation according to the embodiments of the present application takes high-order spectrum feature as the frequency order representation of a signal, thereby realizing end-to-end semi-supervised classification.

[0009] To achieve the above object, in another aspect, the present application provides a semi-supervised game model classification device based on bispectrum feature representation, comprising:

[0010] The acquisition module is configured to acquire the original received signal and the original received signal sample of the transmitting stage; the processing module is configured to input the original received signal into the trained semi-supervised game model for data signal processing, wherein the trained semi-supervised model is trained based on high-order spectrum feature representation for the frequency order of the original received signal sample; and the classification module is configured to classify the original received signal by using a classifier based on the data signal processing to obtain a classification result.

[0011] The semi-supervised game model classification device based on bispectrum feature representation according to the embodiments of the present application takes high-order spectrum feature as the frequency order representation of a signal, thereby realizing end-to-end semi-supervised classification.

[0012] Additional aspects and advantages of the present application will be made apparent from the following description, which proceeds with reference to the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0013] The above and / or additional aspects and advantages of the present application will become apparent and be readily understood from the following description, taken in conjunction with the accompanying drawings, in which:

[0014] Figure 1 A flowchart of the semi-supervised game model classification method based on bispectrum feature representation according to the embodiments of the present application;

[0015] Figure 2 A framework diagram of the semi-supervised game model based on bispectrum feature representation according to the embodiments of the present application;

[0016] Figure 3 A structure diagram of a RU according to an embodiment of the present application;

[0017] Figure 4 A trend curve diagram of generator and discriminator loss according to an embodiment of the present application;

[0018] Fig. 5(a) and Fig. 5(b) are respectively a robust performance diagram of HS-SGAN at different periods and training samples according to an embodiment of the present application;

[0019] Figure 6 A performance comparison diagram of three methods on a simulation data set and a real data set according to an embodiment of the present application;

[0020] Figure 7 A structure diagram of a semi-supervised game model classification device based on bispectrum feature representation according to an embodiment of the present application. DETAILED DESCRIPTION

[0021] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0022] In order to enable persons skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by persons skilled in the art without creative labor should belong to the protection scope of the present application.

[0023] The semi-supervised game model classification method and device based on bispectrum feature representation according to the embodiments of the present application will be described below with reference to the accompanying drawings.

[0024] Figure 1 is a flowchart of the semi-supervised game model classification method based on bispectrum feature representation according to an embodiment of the present application.

[0025] As shown in the figure, the method includes but is not limited to the following steps: Figure 1

[0026] S1, obtaining the original received signal and the original received signal sample of the transmitting stage;

[0027] S2, inputting the original received signal into the trained semi-supervised game model for data signal processing; wherein the trained semi-supervised model is trained based on the high-order spectrum feature representation of the frequency order of the original received signal sample;

[0028] ​S3, based on data signal processing, classifying the original received signal using a classifier to obtain a classification result.

[0029] The embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0030] Specifically, the present application proposes a semi-supervised GAN framework based on bispectrum feature representation, which embeds bispectrum as RFF into the framework, so that the classifier obtains more information about the transmitter while directly classifying the original signal. Different scale residual units are introduced in the network to solve network degradation, and multi-level representation of the signal is extracted and semi-supervised classification is directly completed. Data is tested on simulation and real data, and the results show that the performance of the proposed framework is significantly improved. The framework is as shown in Figure 2

[0031] Further, signal reconstruction and modeling are performed. Considering a wireless communication system composed of a transmitter, a channel and a receiver, the received signal R (i) (nT) can be represented as:

[0032] R (i) (nT) = T (i) (s(nT))*h(nT) + w(nT), n = 1, 2,..., N

[0033] In the formula, s(nT) is the modulated baseband signal, T (i) (·) is the nonlinear distortion of the superposition of the i-th radiation source fingerprint, s(nT) is sent to the channel h(nT) after nonlinear amplification, and w(nT) is the additive noise introduced in the transmission process. The essence of RFFI is to distinguish the nonlinear transformation T (i) (·).

[0034] Methodology: The characteristics of different radiation source individuals are hidden in GAN, which is an unsupervised DL model composed of a generator and a discriminator. The discriminator is responsible for binary classification of the output of the generator and the real sample, and the generator adjusts its parameters through the gradient information of back propagation to blind the discriminator. The essence of network optimization is the game between G and D, and its loss function can be represented as:

[0035]

[0036] In the formula, V(G, D) represents the function to be optimized, E(·) represents the expectation operator, x ~ P data (x) is a real sample, and z ~ P z ​(z) is input noise. But the original GAN can only be unsupervised expansion to the real sample, and cannot effectively use the correlation information of the labeled data and unlabeled data. Inspired by this, we construct a semi-supervised self-classification GAN based on high-order spectral representation, called DR-GAN. Compared with the original GAN, in DR-SGAN, the discriminator is divided into supervised and unsupervised parts. D receives unlabeled samples and G(z) uses the Sigmoid function of the output layer to complete the true and false two-classification, and the output is represented as:

[0037]

[0038] The labeled sample uses the Softmax function of the output layer to complete the supervised classification. The cross-entropy function is deployed as the loss function, which can be represented as:

[0039]

[0040] In the formula, p i represents the predicted value of the i-th label, and T i represents the real label of the sample. In the supervised classification problem, the output of the generator is regarded as the K+1th class. In particular, the discriminant model of the supervised part and the unsupervised part shares the weights and parameters except the output layer, and the supervised discriminator is updated by minimizing the cross-entropy loss, while the unsupervised discriminant model is optimized by using the adversarial training.

[0041] Further, the radio frequency fingerprint extraction based on high-order spectrum. The recognition accuracy of the existing RFFI method depends largely on the effectiveness of the signal feature extraction, and the high-order spectrum is considered as an effective tool for signal analysis, which can directly show the frequency, phase coupling phenomenon between each harmonic component, and has immunity to Gaussian color noise. The invention introduces the reduced high-order spectrum into the MDGAN to help the classifier perceive more emitter information. Bispectrum is the lowest order high-order spectrum, which is usually defined as the 2nd order Fourier transform of the 3rd order cumulant:

[0042]

[0043] In the formula, cum3 (τ1, τ2) = E (x H (nT) x (nT+τ1) x (nT+τ2)), E (·) represents the expectation operator. Considering the symmetry of ω1 and ω2, the bispectrum slice of the diagonal dimension is extracted as the radio frequency fingerprint, that is, The invention hopes to connect to the same signal R (i) (nT) among them, improve the feature separation degree of the emitter.

[0044] Further, the network structure is described. The application uses CNN to construct the generator and discriminator respectively, and expects that the deep network can map the original features to a higher dimensional space, but the stacked convolutional layers often lead to the degradation of network performance. Therefore, the application introduces residual blocks into the deep discriminant network to solve the gradient disappearance and dispersion, and improve the classification efficiency. The structures of the generator and the discriminator are given in the table. The input of the generator is 1*100 Gaussian white noise, and the output is 1*1500 signal vector after multi-layer convolution. The addition of batch normalization layer and pooling layer can prevent gradient disappearance. In the discriminator, we connect multiple residual blocks in series, and each residual block is composed of dilated convolution of different scales. The structures of the generator and the discriminator are shown in Table 1 and Table 2 respectively:

[0045] Table 1

[0046]

[0047] Table 2

[0048]

[0049] In order to solve the problem of gradient dispersion caused by stacked convolutional layers, the application introduces residual units (Residual Unit, RU) in D. The application connects multiple RUs in series, and each RU is composed of convolutional layers of different scales. The purpose is to extract signal features of different levels. The size of the convolution kernel is 1*7, 1*5 and 1*3, and the number of convolution kernels in RU1-RU4 is 128, 64, 32 and 16 respectively. The step is 1*1, and the structure of RU is shown in Figure 3 .

[0050] Further, the loss function is optimized. The training of the DL model is actually the solving process of minimizing the loss function. The loss function of the discriminator includes supervised and unsupervised parts, and its expression is:

[0051]

[0052] The application prevents the generator from overtraining on the current discriminator by requiring the generator to output statistical information matching the real data, in order to improve the stability of GAN and avoid mode collapse. The loss function of the generator is redefined as the expected value of the generator matching the intermediate layer features of the discriminator, so that the generated samples can fit the real data distribution as much as possible, which can be expressed as:

[0053]

[0054] In the formula, F(·) represents the feature output of the intermediate layer of the discriminant network, represents 2-norm. The label smoothing method is introduced in the training process to reduce the discrimination distance between the model's probability of belonging to the full probability class and the probability of belonging to the 0 probability class, solve the problem of wrong labeling and improve the generalization of the network. The present application replaces label 1 in the unsupervised discriminator with 0.9 and replaces 0 with 0.1.

[0055] Figure 4 The losses of the generator (G) and the discriminator (D) trained for 600 iterations are shown. After strong fluctuations, the loss of the discriminator is stable, while the loss of the generator fluctuates slowly and converges quickly.

[0056] Dataset setup: The performance of simulated and real datasets is discussed. Taylor polynomials are used to simulate nonlinear distortion in simulated data. Real data is collected on six software-defined radio platforms based on USRP-2922s. The experiment uses single-carrier (500MHz) binary phase shift keying (BPSK) modulation. In the simulated and real data sets, the number of training samples and test samples is 2500. Gaussian white noise is added to the received signal to simulate the decrease in signal-to-noise ratio (SNR). The present application uses T / R distance to verify the robustness of USRP-2922 identification.

[0057] Results and discussion: Table 3 analyzes the effect of the number of convolution layers in G and the number of RU layers in D on identification performance. The signal-to-noise ratio is 14dB. The accuracy of the evaluation index is defined as the matching ratio of the specific label to the entire label. The combination of four convolution layers and four RUs achieves the highest accuracy.

[0058] Table 3: Average identification accuracy of simulated data with different parameters

[0059]

[0060] Figure 5a The identification accuracy of the simulated data set changes with the number of iterations, Figure 5b The identification accuracy of the simulated data set changes with the number of labeled samples.

[0061] Figure 5a Figures 5a and 5b respectively depict the curves of identification accuracy versus the number of iterations and the number of labeled samples at three different SNRs. In Figure 5a , the identification accuracy increases steadily with the increase of the number of iterations, which proves that the model has strong stability. When the signal-to-noise ratio is > 6dB, the accuracy is > 62%. As Figure 5b shown, when the number of labeled samples exceeds 1000, the identification accuracy is relatively stable. HS-SGAN can achieve high accuracy using only about 40% of the labeled samples in the training set. This is because the present application extracts rich distribution information from unlabeled data, which reduces the dependence of HS-SGAN on labeling.

[0062] The present application uses the most advanced supervised method as a baseline to illustrate the superiority of HS-SGAN. The results are shown in Figure 6 The number of labeled samples used to train HS-SGAN is 2000. Hilbert-Huang transform (HHT) time entropy is used as a feature, and k-nearest neighbor (KNN) algorithm is used for classification. Hilbert spectrum is used as RFF, and deep residual network is used as a classifier. When the signal-to-noise ratio is <8dB, the Acc of HS-SGAN is more advantageous, although it requires less label information. HS-SGAN shows better flexibility in increasing the T / R distance in real data. This means that HS-SGAN can use less label information to show higher recognition accuracy than supervised schemes, and the embedded bispectrum RFF can better reflect the feature information of the transmitter.

[0063] In summary, in the present application, a SSL classification framework of RFFI is constructed using GAN. The framework embeds high-order spectral representation into the original signal, improves the separation of the transmitter, redesigns the loss function, improves the stability of training, and realizes the end-to-end classification of the transmitter. The effectiveness of HS-SGAN is verified in simulated data and real scenarios, and better recognition effect than supervised learning can be obtained even in the case of limited labeled samples.

[0064] According to the semi-supervised game model classification method based on bispectrum feature representation according to the embodiments of the present application, the high-order spectral features are used as the frequency order representation of the signal, so that the end-to-end semi-supervised classification is realized.

[0065] In order to realize the above-mentioned embodiments, as Figure 7 shown in the present embodiment, a semi-supervised game model classification device 10 based on bispectrum feature representation is also provided, which comprises an acquisition module 100, a processing module 200 and a classification module 300.

[0066] The acquisition module 100 acquires the original received signal and the original received signal sample of the transmitter stage;

[0067] The processing module 200 is used for inputting the original received signal into the trained semi-supervised game model for data signal processing; wherein the trained semi-supervised model is trained based on the high-order spectral feature representation as the frequency order of the original received signal sample;

[0068] The classification module 300 is used for classifying the original received signal based on data signal processing to obtain a classification result.

[0069] According to the semi-supervised game model classification device based on bispectrum feature representation according to the embodiments of the present application, the high-order spectral features are used as the frequency order representation of the signal, so that the end-to-end semi-supervised classification is realized.

[0070] It should be noted that the foregoing description of the embodiment of the semi-supervised game model classification method based on the bispectrum feature representation is also applicable to the embodiment of the semi-supervised game model classification device based on the bispectrum feature representation, and details are not repeated here.

[0071] In addition, the terms "first", "second", "third", etc. are used only for descriptive purposes and should not be construed as indicating or implying relative importance or an indicated number of technical features. Therefore, the features defined with "first", "second", etc. can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "a plurality of" is at least two, for example, two, three, etc., unless otherwise explicitly specified.

[0072] In the description of the present application, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present application, the illustrative description of the above terms is not necessarily directed to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. In addition, those skilled in the art can combine and combine different embodiments or examples described in the present application and the features of different embodiments or examples without contradiction.

[0073] Although the embodiments of the present application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application.

Claims

1. A semi-supervised game model classification method based on bispectral feature representation, characterized in that, Includes the following steps: Acquire the raw received signal and raw received signal sample from the transmitter stage; The original received signal is input into a trained semi-supervised game model for data signal processing; wherein, the trained semi-supervised model is obtained by training based on the high-order spectral features represented as the frequency order of the original received signal samples; Based on the data signal processing, a classifier is used to classify the original received signal to obtain a classification result; The method further includes: The semi-supervised game model based on bispectral feature representation embeds the bispectrum as an RFF into the framework, enabling the classifier to obtain characteristic information about the emitter while directly classifying the original signal. Residual units of different scales are introduced into the network to solve network degradation, and multi-level representations of the signal are extracted to directly complete the semi-supervised classification. The original received signal is represented as follows: in, R (i) ( nT The original received signal is shown in the image. The baseband signal has been modulated. For the first i Nonlinear distortion of radio frequency fingerprints caused by superimposed radiation sources After nonlinear amplification, it is transmitted to the channel. , Additive noise introduced during transmission; The characteristics of different radiation sources are hidden in GANs, which are unsupervised deep learning models consisting of a generator and a discriminator. The discriminator is responsible for binary classification of the generator's output and real samples. The generator adjusts its parameters using gradient information from backpropagation to deceive the discriminator. The essence of network optimization is a game between G and D, and the loss function is expressed as: In the formula This represents the function to be optimized. Represents the expectation operator. For real samples, Input noise; In the semi-supervised game model, the discriminator is divided into a supervised part and an unsupervised part. The discriminator receives unlabeled samples from the original received signal samples, and the generator uses the Sigmoid function of the output layer to perform true / false binary classification. The output is represented as: Supervised classification is performed on the labeled samples in the original received signal samples using the softmax function of the output layer, and the cross-entropy function is deployed as the loss function, expressed as: in, Indicates the first i The predicted value for each label, This represents the true label of the original received signal sample.

2. The method according to claim 1, characterized in that, The higher-order spectral features are defined as the second-order Fouier transform of the third-order cumulants: in, , Represents the expectation operator.

3. The method according to claim 2, characterized in that, Set the input of the generator to 1. 100 Gaussian white noise, after multiple convolutions, outputs 1. A signal vector of 1500 is used to concatenate multiple residual blocks in the discriminator; wherein each of the multiple residual blocks is composed of dilated convolutions of different scales.

4. The method according to claim 3, characterized in that, The loss function of the discriminator includes a supervised part and an unsupervised part, and its expression is: 。 5. The method according to claim 4, characterized in that, The loss function of the generator is defined as the expected value of the generator matching the intermediate layer features of the discriminator, expressed as: in, This represents the feature output of the intermediate layer of the discriminant network. It represents the 2-norm.

6. A semi-supervised game model classification device based on bispectral feature representation using the method described in claim 1, characterized in that, include: The acquisition module is used to acquire the raw received signal and raw received signal samples from the transmitter stage; The processing module is used to input the original received signal into the trained semi-supervised game model for data signal processing; wherein, the trained semi-supervised model is trained based on the high-order spectral features represented as the frequency order of the original received signal samples; The classification module is used to classify the original received signal using a classifier based on the data signal processing to obtain a classification result.

Images