Cross-receiver radio frequency fingerprint identification method based on feature separation under small sample condition

Abstract

The invention discloses a cross-receiver radio frequency fingerprint identification method based on feature separation under a small sample condition, relates to the technical field of communication radiation source individual identification, and aims to construct a separation model to perform feature extraction and feature separation on IQ data, separate transmitter features, and identify transmitter categories by using the transmitter features. The separation model comprises a radio frequency fingerprint feature extractor, a transmitter feature extractor, a receiver feature extractor, a transmitter classifier and a receiver classifier. IQ data of different transmitters are respectively collected through different receivers, data enhancement is performed on the collected IQ data, the IQ data after data enhancement is used as a sample, and a separation model is trained by using the sample. According to the invention, the feature separation network is used to decouple transmitter features and receiver features, and radio frequency fingerprint identification across receivers can be realized; and meanwhile, through a data enhancement mode, individual identification under a small sample condition is realized, and the robustness of the identification model under different channel conditions is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of individual identification of communication radiation sources, and in particular to a cross-receiver radio frequency fingerprint identification method based on feature separation under small sample conditions. Background Art

[0002] With the rapid development of the Internet of Things, the number of communication terminals in many emerging fields such as driverless cars and smart homes has increased exponentially. While putting tremendous pressure on limited spectrum resources, equipment security and privacy issues have also attracted people's attention. Using RF fingerprint recognition technology to identify individual transmitters by extracting RF fingerprints from radio signals emitted by transmitters is the key to ensuring the security of spectrum resources and device authentication. In practical applications, it is often necessary to identify on multiple platforms for different task requirements, and the data to be identified also comes from different receivers in different places. Therefore, it is of great significance to design a cross-receiver RF fingerprint recognition system to effectively identify RF fingerprints from different receivers.

[0003] Traditional RF fingerprint recognition methods mainly extract features from radiation source signals through artificial design. Common signal features include parameter domain features such as IQ imbalance and carrier frequency deviation that reflect the impact of various device irrational characteristics on the modulated signal, as well as various transform domain features such as short-time Fourier transform, continuous wavelet transform, and bispectral transform. However, traditional RF fingerprint recognition methods are extremely dependent on expert experience and require complex preprocessing of the signal, so they have limitations in practical applications.

[0004] Compared with traditional methods, methods based on deep learning can solve the RF fingerprint recognition problem from higher-dimensional signal features, reduce reliance on expert experience, and improve the robustness of traditional methods. However, most methods only consider the impact of the transmitter and wireless channel on the RF fingerprint, and ignore the impact of the receiver on the RF fingerprint, resulting in a greatly reduced recognition rate when identifying on a new receiver. In related technologies, such as Chinese patent CN117113061A, a cross-receiver radiation source fingerprint recognition method and device is provided. This method uses data from different receivers to train the neural network multiple times, and realizes accurate and efficient recognition of radiation source signals in scenarios across time and even across receivers. However, this solution does not consider solving the cross-receiver RF fingerprint recognition problem from the perspective of the impact of the receiver on the RF fingerprint.

[0005] From the above, it can be seen that the relevant technology does not provide any technical inspiration on how to improve the cross-receiver RF fingerprint recognition problem. Summary of the invention

[0006] In order to overcome the defects in the above-mentioned prior art, the present invention provides a cross-receiver RF fingerprint recognition method based on feature separation under small sample conditions, which identifies the RF fingerprint of the radiation source according to IQ data, and effectively solves the RF fingerprint recognition problem in the cross-receiver scenario by separating the transmitter characteristics and the receiver characteristics.

[0007] To achieve the above object, the present invention adopts the following technical solutions, including:

[0008] A cross-receiver RF fingerprint recognition method based on feature separation under small sample conditions, which constructs a separation model to extract and separate features from IQ data, separates transmitter features, and uses transmitter features for identification to identify transmitter categories;

[0009] The separation model includes: a radio frequency fingerprint feature extractor, a transmitter feature extractor and a transmitter classifier; the radio frequency fingerprint feature extractor is used to extract features from IQ data to obtain a common radio frequency fingerprint feature including transmitter features and receiver features; the transmitter feature extractor is used to perform feature separation on the common radio frequency fingerprint feature to separate the transmitter features; the transmitter classifier is used to identify according to the transmitter features to identify the transmitter category.

[0010] Preferably, the separation model further includes a receiver feature extractor and a receiver classifier during the training process, and the training process of the separation model is as follows:

[0011] S1, collecting IQ data of different transmitters through different receivers, and performing data enhancement on the collected IQ data, taking the data-enhanced IQ data as samples, and taking the labels of the receiver and transmitter used for collecting and transmitting the IQ data as labels of the samples, that is, the labels of the samples include the receiver label and the transmitter label;

[0012] S2, the separation model includes a radio frequency fingerprint feature extractor, a transmitter feature extractor and a receiver feature extractor, as well as a transmitter classifier and a receiver classifier; the separation model is trained using training samples, as shown below:

[0013] S21, training of RF fingerprint feature extractor:

[0014] The data-enhanced IQ data is input into a radio frequency fingerprint feature extractor to obtain a common radio frequency fingerprint feature including a transmitter feature and a receiver feature;

[0015] S22, training of transmitter feature extractor and receiver feature extractor:

[0016] The common RF fingerprint features obtained by the RF fingerprint feature extractor are respectively input into the transmitter feature extractor and the receiver feature extractor to extract the transmitter features and the receiver features respectively, and the similarity loss is used to reduce the correlation between the transmitter features and the receiver features to achieve feature separation, thereby obtaining the transmitter features and the receiver features respectively;

[0017] S23, training of transmitter classifier and receiver classifier:

[0018] The extracted transmitter features and receiver features are input into the transmitter classifier and receiver classifier respectively, and the cross entropy loss and information entropy loss are used to control the validity of the extracted features to achieve transmitter classification.

[0019] Preferably, in step S1, data enhancement includes noise addition and simulated channel processing; the noise addition is achieved by adding Gaussian noise with a specified signal-to-noise ratio; and the model channel adopts a multipath channel and a Doppler frequency shift channel.

[0020] Preferably, the signal after noise addition and multipath channel enhancement is expressed as:

[0021]

[0022] Among them, num_paths is the number of paths, n(t) is additive Gaussian white noise, and y i,j (t) is the time delay τ i , amplitude attenuation a i and phase shift φ i The output signal component after i,j (t) is expressed as:

[0023] y i,j (t) = a i cos(φ i )x i,j (t-τ i )

[0024] Among them, x i,j (t-τ i ) is the input signal x i,j (t) After a time delay τ i The signal obtained.

[0025] Preferably, the signal after noise addition and Doppler frequency shift channel enhancement is expressed as:

[0026]

[0027] Where Δf i is the Doppler frequency shift, h(t) is the channel impulse response, j is the imaginary unit, x i(t) is the input signal.

[0028] Preferably, in step S21, the radio frequency fingerprint feature extractor is composed of a convolutional layer and five residual blocks.

[0029] Preferably, in step S22, the similarity loss The calculation method is as follows:

[0030]

[0031] Where C represents the covariance matrix of the transmitter features and the receiver features, ||C|| F represents the Frobenius norm of the covariance matrix, c ab represents the element in the ath row and bth column of the covariance matrix C, and A and B represent the number of rows and columns of the covariance matrix C.

[0032] Preferably, in step S23,

[0033] Transmitter Classifier Cross Entropy Loss The calculation method is as follows:

[0034]

[0035] Among them, z tx is the transmitter characteristic; y tx is the transmitter label; M is the number of samples; H(·) is the cross entropy loss function; Xm is the mth sample; x represents the IQ data of the sample; y represents the label of the sample;

[0036] Receiver Classifier Cross Entropy The calculation method is as follows:

[0037]

[0038] Among them, z rx is the receiver characteristic, y rx is the receiver tag;

[0039] Transmitter classifier information entropy loss The calculation method is as follows:

[0040]

[0041] Among them, y mk represents the probability that the mth sample is classified as the kth transmitter; N is the number of transmitters; P(y mk ) represents probability distribution;

[0042] Receiver classifier information entropy loss The calculation method is as follows:

[0043]

[0044] Among them, y mw represents the probability that the mth sample is classified as the wth receiver; J is the number of receivers; P(y mw ) represents the probability distribution.

[0045] The present invention also provides a computer program product, which includes a computer program / instruction. When the computer program / instruction is executed by a processor, the cross-receiver radio frequency fingerprint recognition method based on feature separation under small sample conditions is implemented.

[0046] The present invention also provides an electronic device, comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the cross-receiver radio frequency fingerprint recognition method based on feature separation under small sample conditions is implemented.

[0047] The advantages of the present invention are:

[0048] (1) The present invention realizes the identification of the RF fingerprint of the radiation source based on IQ data under small sample conditions, and effectively solves the RF fingerprint identification problem in cross-receiver scenarios by separating the transmitter characteristics and the receiver characteristics.

[0049] (2) The present invention uses noise addition and simulated channels for data enhancement, thereby solving the problem of sample scarcity during training. Simultaneously, the simulated channels can effectively enhance the robustness of the model system under different channel environments, thereby realizing individual recognition under small sample conditions and enhancing the robustness of the recognition model under different channel conditions.

[0050] (3) The present invention uses feature separation to train a separation model through IQ data, and separates transmitter features from receiver features, thereby effectively reducing the impact of the receiver on the RF fingerprint and improving the recognition performance of the model in cross-receiver scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 The figure is a flow chart of a cross-receiver radio frequency fingerprint recognition method under a small sample condition in one embodiment of the present invention.

[0052] Figure 2 Schematic diagram of a feature separation process in one embodiment of the present invention.

[0053] Figure 3 The figure is a schematic diagram comparing the performance of radio frequency fingerprint recognition using analog channel enhancement and non-enhancement in one embodiment of the present invention.

[0054] Figure 4The figure is a schematic diagram comparing the recognition performance of the separation model and the adversarial model algorithms in one embodiment of the present invention. DETAILED DESCRIPTION

[0055] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] Depend on Figure 1 As shown, the present invention provides a cross-receiver radio frequency fingerprint recognition method based on feature separation under small sample conditions, comprising the following steps:

[0057] S1, collect IQ data of different transmitters through different receivers, and perform data enhancement on the collected IQ data, use the data-enhanced IQ data as samples, and use the labels of the receiver and transmitter used to collect and transmit the IQ data as labels of the samples, that is, the labels of the samples include the receiver label and the transmitter label. The samples include training samples and test samples.

[0058] Among them, the collected IQ data is first processed to a size that meets the input neural network model, and then the collected IQ data is enhanced. The data enhancement includes noise addition and analog channel processing; noise addition is achieved by adding Gaussian noise with a specified signal-to-noise ratio, and the model channel uses a multipath channel and a Doppler frequency shift channel.

[0059] The signal after noise addition and multipath channel enhancement is expressed as:

[0060]

[0061] Among them, num_paths is the number of paths, n(t) is additive Gaussian white noise, and y i,j (t) is the time delay τ i , amplitude attenuation a i and phase shift φ i The output signal component after i,j (t) is expressed as:

[0062] y i,j (t) = a i cos(φ i )x i,j (t-τ i )

[0063] Among them, x i,j (t-τ i ) is the input signal x i,j(t) After a time delay τ i The signal obtained.

[0064] The signal after noise addition and Doppler frequency shift channel enhancement is expressed as:

[0065]

[0066] Where Δf i is the Doppler frequency shift, h(t) is the channel impulse response, j is the imaginary unit, x i (t) is the input signal.

[0067] Specifically, the IQ data is processed into 2×256 I-channel and Q-channel sequences, where 2 is the number of channels and 256 is the sampling point data.

[0068] Specifically, noise is added by adding Gaussian noise with a specified signal-to-noise ratio, and the signal-to-noise ratio is uniformly distributed in [15dB, 30dB]. The analog channel includes a multipath channel and a Doppler frequency shift channel. The number of paths and the maximum delay in the multipath channel are uniformly distributed in [1, 5] and [5ns, 10ns], respectively. The Doppler frequency shift in the Doppler frequency shift channel is uniformly distributed in [-15Hz, 15Hz].

[0069] S2, constructing a separation model, including a radio frequency fingerprint feature extractor, a transmitter feature extractor and a receiver feature extractor, as well as a transmitter classifier and a receiver classifier; using training samples to train the separation model, as shown below:

[0070] S21, training of RF fingerprint feature extractor:

[0071] The data-enhanced IQ data is input into a radio frequency fingerprint feature extractor to obtain a common radio frequency fingerprint feature including transmitter features and receiver features.

[0072] Specifically, the sample is input into a feature extractor consisting of a convolutional layer and five residual blocks to obtain a public RF fingerprint feature with a size of 1×512.

[0073] S22, training of transmitter feature extractor and receiver feature extractor:

[0074] The common RF fingerprint features obtained by the RF fingerprint feature extractor are respectively input into the transmitter feature extractor and the receiver feature extractor to extract the transmitter features and the receiver features respectively, and the similarity loss is used to reduce the correlation between the transmitter features and the receiver features to achieve feature separation, thereby obtaining the transmitter features and the receiver features respectively.

[0075] Specifically, the extracted common RF fingerprint features are passed through a fully connected layer to obtain transmitter features and receiver features of size 256;

[0076] Using similarity loss The similarity between the transmitter and receiver features is reduced as much as possible to achieve feature separation. The calculation method is as follows:

[0077]

[0078] Where C represents the covariance matrix of the transmitter features and the receiver features, ||C|| F represents the Frobenius norm of the covariance matrix, c ab represents the element in the ath row and bth column of the covariance matrix C, and A and B represent the number of rows and columns of the covariance matrix C.

[0079] S23, training of transmitter classifier and receiver classifier:

[0080] The transmitter features extracted by the transmitter feature extractor and the receiver features extracted by the receiver feature extractor are input into the transmitter classifier and the receiver classifier respectively, and the cross entropy loss and the information entropy loss are used to control the validity of the extracted features, improve the accuracy of individual classification, and realize individual classification.

[0081] Specifically, the transmitter features and receiver features are respectively passed through the fully connected layer to obtain the transmitter features of size 6 and the receiver features of size 2; using the cross entropy loss and information entropy loss Control the effectiveness of extracting fingerprint features, the goal is to make the cross entropy loss and information entropy loss The sum is minimized.

[0082] Transmitter Classifier Cross Entropy Loss The calculation method is as follows:

[0083]

[0084] Among them, z tx is the transmitter characteristic; y tx is the transmitter label; M is the number of samples; H(·) is the cross entropy loss function; Xm is the mth sample; x represents the IQ data of the sample; y represents the label of the sample.

[0085] Receiver Classifier Cross Entropy The calculation method is as follows:

[0086]

[0087] Among them, z rx is the receiver characteristic, y rx Receiver tag.

[0088] Transmitter classifier information entropy loss The calculation method is as follows:

[0089]

[0090] Where M is the number of samples, y mk represents the probability that the mth sample is classified as the kth transmitter; N is the number of transmitters; P(y mk ) represents the probability distribution.

[0091] Receiver classifier information entropy loss The calculation method is as follows:

[0092]

[0093] Among them, y mw represents the probability that the mth sample is classified as the wth receiver; J is the number of receivers; P(y mw ) represents the probability distribution.

[0094] In step S3, the test sample is input into the separation model, and the individual identification of the RF fingerprint (transmitter) across receivers is realized through the separated transmitter features.

[0095] Specifically, the test sample is input into the RF fingerprint feature extractor to obtain a public RF fingerprint feature including transmitter features and receiver features; the public RF fingerprint feature is input into the transmitter feature extractor to extract the transmitter feature; the transmitter feature is input into the transmitter classifier to finally obtain the individual identification result of the RF fingerprint (transmitter).

[0096] After training and testing, the separation model is used to extract and separate features from the IQ data, separate the transmitter features, and use the transmitter features for identification to identify the transmitter category.

[0097] Table 1 is a performance comparison between the present invention and the traditional method without considering the receiver effect on the Wisig data set. The data in the table are the correct identification results of the radiation source.

[0098] Table 1

[0099]

[0100]

[0101] As can be seen from Table 1, the accuracy performance of the present invention on the real data set is better than that of the traditional method that does not consider the influence of the receiver, and the average accuracy is improved by about 9.51%.

[0102] Table 2 shows the performance of the present invention under different test dates on the Wisig dataset, and the data in the table are the correct identification results of the radiation source.

[0103] Table 2

[0104]

[0105] As can be seen from Table 2, in the real data set, the present invention can still maintain a certain model robustness under the conditions of testing on different dates, and the average recognition rate remains at 77.7%.

[0106] like Figure 2 As shown in the figure, Class0-Class5 represent a total of 6 different types of transmitter individuals. The left figure shows that before feature separation, the features of each type of transmitter individual are mixed together and cannot be effectively classified and identified; the right figure shows that after feature separation, the features of each type of transmitter individual are separated to a certain extent, which is conducive to classification and identification, proving the effectiveness of feature separation of the model.

[0107] like Figure 3 As shown in FIG. 1 , the channel enhancement adopted by the model of the present invention has a significant improvement in the cross-receiver recognition effect, which is about 21.9% higher than that without enhancement. Figure 4 As shown, the feature separation method of the present invention improves the performance by about 4.2% compared with the existing adversarial method.

[0108] In summary, the cross-receiver RF fingerprint recognition method under small sample conditions of the present invention effectively decouples the transmitter characteristics and the receiver characteristics through feature separation, realizes RF fingerprint recognition in cross-receiver scenarios, and enhances the robustness of the model in different channel environments by using simulated channels.

[0109] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A cross-receiver radio frequency fingerprint recognition method based on feature separation under small sample conditions, characterized in that: Construct a separation model to extract and separate features from IQ data, separate transmitter features, and use transmitter features to identify transmitter categories; The separation model includes: a radio frequency fingerprint feature extractor, a transmitter feature extractor and a transmitter classifier; the radio frequency fingerprint feature extractor is used to extract features from IQ data to obtain a common radio frequency fingerprint feature including transmitter features and receiver features; the transmitter feature extractor is used to perform feature separation on the common radio frequency fingerprint feature to separate the transmitter features; the transmitter classifier is used to identify according to the transmitter features to identify the transmitter category.

2. According to the method of claim 1, the cross-receiver radio frequency fingerprint recognition method based on feature separation under small sample conditions is characterized in that: The separation model also includes a receiver feature extractor and a receiver classifier during the training process. The training process of the separation model is as follows: S1, collecting IQ data of different transmitters through different receivers, and performing data enhancement on the collected IQ data, taking the data-enhanced IQ data as samples, and taking the labels of the receiver and transmitter used for collecting and transmitting the IQ data as labels of the samples, that is, the labels of the samples include the receiver label and the transmitter label; S2, the separation model includes a radio frequency fingerprint feature extractor, a transmitter feature extractor and a receiver feature extractor, as well as a transmitter classifier and a receiver classifier; the separation model is trained using training samples, as shown below: S21, training of RF fingerprint feature extractor: The data-enhanced IQ data is input into a radio frequency fingerprint feature extractor to obtain a common radio frequency fingerprint feature including a transmitter feature and a receiver feature; S22, training of transmitter feature extractor and receiver feature extractor: The common RF fingerprint features obtained by the RF fingerprint feature extractor are respectively input into the transmitter feature extractor and the receiver feature extractor to extract the transmitter features and the receiver features respectively, and the similarity loss is used to reduce the correlation between the transmitter features and the receiver features to achieve feature separation, thereby obtaining the transmitter features and the receiver features respectively; S23, training of transmitter classifier and receiver classifier: The extracted transmitter features and receiver features are input into the transmitter classifier and receiver classifier respectively, and the cross entropy loss and information entropy loss are used to control the validity of the extracted features to achieve transmitter classification.

3. According to the method of cross-receiver radio frequency fingerprint recognition based on feature separation under small sample conditions in claim 2, it is characterized in that: In step S1, data enhancement includes noise addition and simulated channel processing; the noise addition is achieved by adding Gaussian noise with a specified signal-to-noise ratio; and the model channel uses a multipath channel and a Doppler frequency shift channel.

4. According to the method of cross-receiver radio frequency fingerprint recognition based on feature separation under small sample conditions in claim 3, it is characterized in that: The signal after noise addition and multipath channel enhancement is expressed as: Among them, num_paths is the number of paths, n(t) is additive Gaussian white noise, and y i,j (t) is the time delay τ i , amplitude attenuation a i and phase shift φ i The output signal component after i,j (t) is expressed as: y i,j (t)=a i cos(φ i )x i,j (t-t i ) Among them, x i,j (t-τ i ) is the input signal x i,j (t) After a time delay τ i The signal obtained.

5. According to the method of cross-receiver radio frequency fingerprint recognition based on feature separation under small sample conditions in claim 3, it is characterized in that: The signal after noise addition and Doppler frequency shift channel enhancement is expressed as: Where Δf i is the Doppler frequency shift, h(t) is the channel impulse response, j is the imaginary unit, x i (t) is the input signal.

6. According to the method of cross-receiver radio frequency fingerprint recognition based on feature separation under small sample conditions in claim 2, it is characterized in that: In step S21, the radio frequency fingerprint feature extractor is composed of a convolutional layer and five residual blocks.

7. According to the method of cross-receiver radio frequency fingerprint recognition based on feature separation under small sample conditions in claim 2, it is characterized in that: In step S22, the similarity loss The calculation method is as follows: Where C represents the covariance matrix of the transmitter features and the receiver features, ||C|| F represents the Frobenius norm of the covariance matrix, c ab represents the element in the ath row and bth column of the covariance matrix C, and A and B represent the number of rows and columns of the covariance matrix C.

8. According to the method of cross-receiver radio frequency fingerprint recognition based on feature separation under small sample conditions in claim 2, it is characterized in that: In step S23, Transmitter Classifier Cross Entropy Loss The calculation method is as follows: Among them, z tx is the transmitter characteristic; y tx is the transmitter label; M is the number of samples; H(·) is the cross entropy loss function; Xm is the mth sample; x represents the IQ data of the sample; y represents the label of the sample; Receiver Classifier Cross Entropy The calculation method is as follows: Among them, z rx is the receiver characteristic, y rx is the receiver tag; Transmitter classifier information entropy loss The calculation method is as follows: Among them, y mk represents the probability that the mth sample is classified as the kth transmitter; N is the number of transmitters; P(y mk ) represents probability distribution; Receiver classifier information entropy loss The calculation method is as follows: Among them, y mw represents the probability that the mth sample is classified as the wth receiver; J is the number of receivers; P(y mw ) represents the probability distribution.

9. A computer program product, characterized in that It includes a computer program / instruction, which, when executed by a processor, implements a cross-receiver radio frequency fingerprint recognition method based on feature separation under small sample conditions as described in any one of claims 1 to 8.

10. An electronic device, characterized in that: It includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a cross-receiver radio frequency fingerprint recognition method based on feature separation under small sample conditions as described in any one of claims 1 to 8.

Images