Radio frequency fingerprint identification method based on conditional generative adversarial network and multi-channel attention network

Abstract

The invention discloses a radio frequency fingerprint identification method based on a conditional generative adversarial network and a multi-channel attention network. In the prior art, most radio frequency fingerprint identification models are interfered under a multipath channel condition, and the identification performance is reduced. In order to solve the problem, the invention provides a channel equalization method based on a conditional generative adversarial network, the channel estimation effect is improved by generating an extended leader sequence, and feature extraction is carried out in cooperation with a multi-channel attention network. In the feature extraction stage, four different signal representations including time domain signals, short-time Fourier transform, carrier frequency offset and fast Fourier transform features are used as input of the multi-channel attention network, channel response is used as a modulation factor, and parameters of the attention network are finely adjusted.

Description

Technical Field

[0001] The present invention belongs to the technical field of radio frequency fingerprint recognition, and in particular relates to a radio frequency fingerprint recognition method based on a conditional generative adversarial network and a multi-channel attention network. Background Art

[0002] With the rapid development of wireless communication and Internet of Things (IoT) technologies, a large number of IoT devices are connected to the network, transmitting various status and control information, which puts higher requirements on the security of the devices. How to effectively authenticate IoT devices in the network has become an urgent problem to be solved. Traditional device authentication methods mainly rely on network layer identifiers, such as IP addresses and MAC addresses. However, these methods have natural weaknesses in dealing with forgery and tampering. Once the identifier is leaked, other devices in the network may be attacked and invaded. Therefore, researchers have gradually turned to physical layer-based authentication technology to improve the security and reliability of device authentication.

[0003] Radio Frequency Fingerprint (RFF) identification technology is one of the important methods for physical layer device identification. Unlike identifiers such as IP and MAC, RF fingerprints are based on tiny hardware differences in the device's RF transmitter module during the manufacturing process. These differences will appear as unique signal characteristics during signal transmission and are difficult to copy or forge. These characteristics form a unique RF fingerprint for each device, providing a highly reliable means of identity authentication.

[0004] However, RF fingerprint recognition still faces some challenges in practical applications. Since wireless signals are affected by multipath channels in complex environments, that is, the signals are reflected, refracted and scattered by obstacles during propagation, the characteristics of the received signals change. These changes may mask the hardware characteristics of the device, affecting the accuracy of RF fingerprint recognition. In addition, traditional RFF recognition methods mostly rely on the original signal characteristics in the time domain, and do not make full use of the frequency domain and other information. This single feature extraction method is difficult to maintain a stable recognition effect under variable channel conditions. Based on the above problems, the present invention proposes a RF fingerprint recognition algorithm based on a conditional generative adversarial network and a multi-channel attention network, which uses a conditional generative adversarial network to generate and expand the received signal sequence, and performs weighted fusion of multiple signal features through a multi-channel attention network to further improve the accuracy and robustness of device recognition. Summary of the invention

[0005] In order to solve the above technical problems, the present invention proposes a radio frequency fingerprint recognition method based on a conditional generative adversarial network and a multi-channel attention network. The method first uses a conditional generative adversarial network (CGAN) to generate a simulated received signal preamble sequence, and splices it with the actual received preamble sequence, thereby improving the accuracy of the least squares method for channel estimation and improving the signal quality. In the feature fusion stage, a multi-channel attention network is used to extract the equalized signal and its various transformation features, including short-time Fourier transform, carrier frequency deviation and fast Fourier transform, etc., and weighted fusion is performed to amplify the feature weights that are helpful for identification. In the classification stage, the ResNeXt classification network is used to process the weighted comprehensive features.

[0006] The method specifically comprises the following steps:

[0007] (1) Standardize the received signal;

[0008] (2) Generate a preamble sequence that simulates the receiving end using a conditional generative adversarial network;

[0009] (3) Least squares channel estimation and equalization using the concatenated extended preamble sequence;

[0010] (4) Use a multi-channel attention network to extract multi-channel features from the equalized signal;

[0011] (5) Generate a dynamic modulation factor using the channel response obtained in step (3) to adjust the network feature weight;

[0012] (6) Fusing the weighted features generated by the multi-channel attention network to obtain a comprehensive feature representation;

[0013] (7) Input the comprehensive feature representation into the classification network to obtain the prediction result of device identification;

[0014] (8) Update the network parameters according to the classification results, and repeat steps (1) to (7) until the network converges;

[0015] (9) Output the optimal network parameters for testing.

[0016] Preferably, the specific features of step (1) include:

[0017] (1.1) For the received signal y = [y0, y1, ..., y N ] is standardized, N is the number of signals, and it is achieved by calculating the mean μ and standard deviation σ of the signal:

[0018]

[0019] (1.2) Use the mean and standard deviation to perform standardization and obtain the standardized signal

[0020]

[0021] Preferably, the specific features of step (2) include:

[0022] (2.1) Construct a generative adversarial network, the generator G(z|x p ,l i θ g ) accepts random noise z and known leading sequence x p and device tags i As input, generate a simulated receive preamble sequence θ g are the parameters of the generator network, the discriminator D(y p |x p ,l i θ d ) for the generated sequence and the true sequence y p To distinguish, θ d is the parameter of the discriminator network, and the loss function L of the network D The definition is as follows, where p data (y p ) is y p Obey the data distribution, p(z) is the probability distribution of random noise z

[0023]

[0024] (2.2) The generated simulated receiving preamble sequence The actual received preamble sequence y p Splice to obtain the extended leader sequence

[0025] Preferably, the specific features of step (3) include:

[0026] (3.1) Using the extended leading sequence y ex_p Perform channel estimation, channel response Estimation was performed by the method of least squares;

[0027] (3.2) After the channel estimation is obtained, the received signal y is equalized by the following formula:

[0028]

[0029] Preferably, the specific features of step (4) include:

[0030] (4.1) The equalized signal y eq And three other signal representations: short-time Fourier transform, carrier frequency deviation and fast Fourier transform features are input into the multi-channel attention network for feature extraction; the attention network weights the features of different signal channels.

[0031] Preferably, the specific features of step (5) include:

[0032] (5.1) Based on the channel response obtained in step (3) Generate modulation factors through multi-layer perceptron And compare it with the feature weight s of the attention module i Modulated Attention is obtained by element-by-element multiplication. i :

[0033]

[0034] (5.2) Modulated Attention is weighted by the following formula i Normalize it to get the normalized attention weight α of the i-th feature i , exp is the exponential function:

[0035]

[0036] Preferably, the specific features of step (6) include:

[0037] (6.1) The four different features processed by the attention network are represented as F i Perform weighted fusion to obtain the comprehensive representation F′, α i is the normalized attention weight of the ith feature:

[0038]

[0039] Preferably, the specific features of step (7) include:

[0040] (7.1) The comprehensive feature representation is input into the classification network ResNeXt for device identification and classification.

[0041] Preferably, the specific features of step (8) include:

[0042] (8.1) According to the error between the classification result and the true label, the network parameters are updated through the back propagation algorithm, and the cross entropy is selected as the loss function. The calculation formula is as follows:

[0043]

[0044] (8.2) Repeat steps (1) to (7) until the training process converges and the optimal network parameters are obtained.

[0045] The beneficial effects of the present invention are:

[0046] 1. The conditional generative adversarial network is used to generate a simulated preamble sequence, which is then concatenated with the actual received sequence for equalization, effectively enhancing the robustness of the model under time-varying channel conditions and reducing the impact of channel changes on device recognition accuracy;

[0047] 2. Using a multi-channel attention network to perform weighted extraction of multiple features of the equalized signal can adaptively enhance key features and further improve the accuracy of device recognition;

[0048] 3. In the experiment, the features weighted by the multi-channel attention network significantly improved the final recognition accuracy, which is better than the traditional single signal processing method, proving the effectiveness and application potential of this method. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the examples of the present invention, the following briefly introduces the drawings required for the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative work.

[0050] Attached Figure 1 It is a flow chart of the radio frequency fingerprint recognition method based on conditional generative adversarial network and multi-channel attention network of the present invention;

[0051] Attached Figure 2 A schematic diagram of the structure of the conditional generative adversarial network of the present invention;

[0052] Attached Figure 3 This is a network structure diagram of the multi-channel attention feature extraction network of the present invention. DETAILED DESCRIPTION

[0053] In order to make the purpose and technical solution of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be described more clearly and completely in combination with the drawings of the embodiments in the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present application.

[0054] Since the present invention is based on a deep neural network, the description of the specific implementation method will be described separately from the two aspects of the training process and the testing process.

[0055] like Figure 1 The flowchart of the training process of the radio frequency fingerprint recognition method based on the conditional generative adversarial network and the multi-channel attention network of the present invention is shown, which includes the following steps:

[0056] Step 1: For the received signal y=[y0,y1,…,y N ] for standardization.

[0057] (1.3) The mean μ and standard deviation σ of the signal are calculated according to the following formula:

[0058]

[0059] (1.4) Use the mean and standard deviation to perform standardization and obtain the standardized signal

[0060]

[0061] Step 2: Use the conditional generative adversarial network to generate a leading sequence that simulates the receiving end.

[0062] (2.3) Using Figure 2 The generative adversarial network shown, the generator G(z|x p ,l i θ g ) accepts random noise z and known leading sequence x p and device tags i As input, generate a simulated receive preamble sequence The discriminator D distinguishes between the generated sequence and the real sequence. The loss function of the network is defined as:

[0063]

[0064] (2.4) The generated simulated receiving preamble sequence The actual received preamble sequence y p Splice to obtain the extended leader sequence

[0065] Step 3: Use the concatenated extended preamble sequence to perform least squares channel estimation and equalization.

[0066] (3.3) Using the extended leading sequence y ex_p Perform channel estimation, channel response Use the least squares method to estimate;

[0067] (3.4) After obtaining the channel estimate, the received signal y is equalized using the following formula:

[0068]

[0069] Step 4: Use the multi-channel attention network to extract multi-channel features from the equalized signal.

[0070] (4.2) The equalized signal y eq The other three signal representations (short-time Fourier transform, carrier frequency deviation and fast Fourier transform features) are input into the multi-channel attention network for feature extraction. The attention network weights the features of different signal channels to enhance the influence of important features.

[0071] Step 5: Generate a dynamic modulation factor based on the channel response obtained in step (3) to adjust the network feature weight.

[0072] (5.3) Based on the channel response obtained in step (3) Generate modulation factors through multi-layer perceptron And modulate it with the feature weight of the attention module through element-wise multiplication:

[0073]

[0074] (5.4) The attention weight is normalized by the following formula to obtain the normalized attention weight α i :

[0075]

[0076] Step six, fuse the weighted features generated by the multi-channel attention network to obtain a comprehensive feature representation.

[0077] (6.2) The four different features processed by the attention network are represented as F i Perform weighted fusion to obtain the comprehensive representation F′:

[0078]

[0079] Step 7: Input the comprehensive feature representation into the classification network to obtain the prediction result of device identification.

[0080] (7.2) The comprehensive feature representation is input into the classification network ResNeXt for device identification and classification.

[0081] Step 8: Update the network parameters according to the classification results, repeat steps (1) to (7) until the network converges, and save the optimal network parameters for subsequent testing.

[0082] (8.3) According to the error between the classification result and the true label, the network parameters are updated through the back propagation algorithm, and the cross entropy is selected as the loss function. The calculation formula is as follows:

[0083]

[0084] (8.4) Repeat steps (1) to (7) until the training process converges and the optimal network parameters are obtained.

[0085] The relevant parameters of the network are determined during the training phase, while steps 1 to 6 during the testing process are the same as the training process.

[0086] (Because the network parameters have been determined during the training process, there is no need to update the parameters), and modify step 7 as follows:

[0087] Step seven: output the predicted category of the signal.

[0088] Take the category with the highest probability output in step 6 as the category predicted by the network

[0089]

Claims

1. A radio frequency fingerprint recognition method based on conditional generative adversarial network and multi-channel attention network, characterized in that: The steps include: (1) Standardize the received signal; (2) Generate a preamble sequence that simulates the receiving end using a conditional generative adversarial network; (3) Least squares channel estimation and equalization using the concatenated extended preamble sequence; (4) Use a multi-channel attention network to extract multi-channel features from the equalized signal; (5) Generate a dynamic modulation factor using the channel response obtained in step (3) to adjust the network feature weight; (6) Fusing the weighted features generated by the multi-channel attention network to obtain a comprehensive feature representation; (7) Input the comprehensive feature representation into the classification network to obtain the prediction result of device identification; (8) Update the network parameters according to the classification results, and repeat steps (1) to (7) until the network converges; (9) Output the optimal network parameters for testing.

2. The radio frequency fingerprint recognition method based on conditional generative adversarial network and multi-channel attention network according to claim 1 is characterized in that: Specific features of step (1) include: (1.1) For the received signal y = [y0, y1, ..., y N ] is standardized, N is the number of signals, and it is achieved by calculating the mean μ and standard deviation σ of the signal: (1.2) Use the mean and standard deviation to perform standardization and obtain the standardized signal 3. The radio frequency fingerprint recognition method based on conditional generative adversarial network and multi-channel attention network according to claim 1 is characterized in that: Specific features of step (2) include: (2.1) Construct a generative adversarial network, the generator G(z|x p ,l i θ g ) accepts random noise z and known leading sequence x p and device tags i As input, generate a simulated receive preamble sequence θ g are the parameters of the generator network, the discriminator D(y p |x p ,l i θ d ) for the generated sequence and the true sequence y p To distinguish, θ d is the parameter of the discriminator network, and the loss function L of the network D The definition is as follows, where p data (y p ) is y p Obey the data distribution, p(z) is the probability distribution of random noise z (2.2) The generated simulated receiving preamble sequence The actual received preamble sequence y p Splice to obtain the extended leader sequence 4. The radio frequency fingerprint recognition method based on conditional generative adversarial network and multi-channel attention network according to claim 1 is characterized in that: Specific features of step (3) include: (3.1) Using the extended leading sequence y ex_p Perform channel estimation, channel response Estimation was performed by the method of least squares; (3.2) After the channel estimation is obtained, the received signal y is equalized by the following formula:

5. The radio frequency fingerprint recognition method based on conditional generative adversarial network and multi-channel attention network according to claim 1 is characterized in that: Specific features of step (4) include: (4.1) The equalized signal y eq And three other signal representations: short-time Fourier transform, carrier frequency deviation and fast Fourier transform features are input into the multi-channel attention network for feature extraction; the attention network weights the features of different signal channels.

6. The radio frequency fingerprint recognition method based on conditional generative adversarial network and multi-channel attention network according to claim 1 is characterized in that: Specific features of step (5) include: (5.1) Based on the channel response obtained in step (3) Generate modulation factors through multi-layer perceptron And compare it with the feature weight s of the attention module i Modulated Attention is obtained by element-by-element multiplication. i : (5.2) Modulated Attention is weighted by the following formula i Normalize it to get the normalized attention weight α of the i-th feature i , exp is the exponential function:

7. The radio frequency fingerprint recognition method based on conditional generative adversarial network and multi-channel attention network according to claim 1 is characterized in that: Specific features of step (6) include: (6.1) The four different features processed by the attention network are represented as F i Perform weighted fusion to obtain the comprehensive representation F′, α i is the normalized attention weight of the ith feature:

8. The radio frequency fingerprint recognition method based on conditional generative adversarial network and multi-channel attention network according to claim 1 is characterized in that: Specific features of step (7) include: (7.1) The comprehensive feature representation is input into the classification network ResNeXt for device identification and classification.

9. The radio frequency fingerprint recognition method based on conditional generative adversarial network and multi-channel attention network according to claim 1 is characterized in that: Specific features of step (8) include: (8.1) According to the error between the classification result and the true label, the network parameters are updated through the back propagation algorithm, and the cross entropy is selected as the loss function. The calculation formula is as follows: (8.2) Repeat steps (1) to (7) until the training process converges and the optimal network parameters are obtained.

Images