Radio frequency fingerprint identification method based on deentanglement

Abstract

The invention discloses a radio frequency fingerprint identification method based on deentanglement, and the method specifically comprises the steps: inputting original IQ data subjected to different channel conditions into a feature extraction module for preliminary feature extraction, and then feeding the extracted signal features to a hardware feature extraction module and a channel feature extraction module for feature decoupling; in a deentanglement module of the network, designing difference loss and similar loss to perform feature decoupling to obtain hardware features and channel features of radio frequency signals; after feature decoupling, an auxiliary module based on cross reconstruction and center loss is designed to optimize the model, finally hardware features are input into a classification module to obtain a classification probability, and training data are used batch by batch to train network parameters. According to different causes and characteristics of a part related to a channel and a part related to hardware in radio frequency fingerprint characteristics, the part related to the channel and the part related to the hardware are separated through a designed de-entanglement loss function, hardware characteristics are reserved, and channel characteristics are eliminated.

Description

Technical Field

[0001] The present invention belongs to the technical field of radio frequency fingerprint recognition methods, and in particular relates to a radio frequency fingerprint recognition method based on deentanglement. Background Art

[0002] Due to the open nature of wireless communications, wireless information transmission faces numerous security threats. To ensure wireless communication security, the primary security barrier in current wireless networks is access authentication. Traditional device access authentication typically relies on key encryption or software-based authentication based on addresses such as MAC / IP addresses. However, these methods are susceptible to fraud, eavesdropping, and impersonation. Therefore, there is an urgent need to explore more secure and reliable device authentication methods. Radio frequency fingerprint recognition can effectively identify different wireless devices and is a reliable and lightweight security authentication method.

[0003] In the field of RF fingerprinting, deep neural networks' powerful nonlinear representation and learning capabilities enable them to learn high-level features directly from complex raw signals, resulting in improved recognition performance. However, most deep neural networks rely on the assumption that the training and test sets come from the same distribution, which is inconsistent with the conditions in many real-world scenarios. In reality, the received signal is affected by both the hardware of the transmitting device and the channel environment. When the channel conditions experienced by the RF signal differ between the training and testing phases, the data distribution will drift, significantly reducing the accuracy and stability of the RF fingerprinting system.

[0004] Existing research to address this issue focuses on the following three areas. First, data augmentation methods. Due to the non-stationary, dynamic, and unpredictable nature of wireless channels, it is difficult to collect training data under all channel conditions. Therefore, data augmentation methods can be used to simulate transmission scenarios across different channels to make the training data as complete as possible. However, such methods not only require prior knowledge of the channel model of the transmission environment; the increased sample size also complicates model training. Second, channel removal methods. Some studies utilize the characteristics of RF signals to design algorithms that minimize the impact of the channel on the signal. However, such methods require prior information about the known signal and may not completely remove the channel influence or obscure the device's original hardware fingerprint in the process. Third, domain adaptation methods treat the inconsistency between the training and test set distributions as a feature distribution shift between domains. Domain adaptation techniques are introduced to improve the cross-channel generalization capabilities of RF fingerprint recognition models. Summary of the Invention

[0005] The purpose of the present invention is to provide a radio frequency fingerprint recognition method based on disentanglement, which solves the problem in the prior art that the data distribution of the training set and the test set drifts, resulting in a significant decrease in the accuracy and stability of the radio frequency fingerprint recognition system.

[0006] The technical solution adopted by the present invention is a radio frequency fingerprint recognition method based on disentanglement, specifically: the original IQ data that has undergone different channel conditions is input into the feature extraction module for preliminary feature extraction, and then the extracted signal mixing features are fed into the hardware feature extraction module and the channel feature extraction module respectively for feature decoupling; in the network's disentanglement module, difference loss and similarity loss are designed for feature decoupling to obtain the hardware features and channel features of the radio frequency signal; after feature decoupling, an auxiliary module based on cross reconstruction and center loss is designed to optimize the model, and finally the hardware features are input into the classification module to obtain the classification probability, and the training data is used batch by batch to train the network parameters.

[0007] The present invention is also characterized in that: Please follow the steps below to implement it: Step 1: Acquire signals transmitted by wireless communication devices at different times, obtain samples containing different channel information, and construct a data set; Step 2: Input the training set in step 1 into the feature extraction module integrating the dual attention mechanism to preliminarily extract the mixed features of the device signal; Step 3: Build a de-entanglement module and input the mixed signal features extracted in step 2 into the de-entanglement module to decouple the signal features and decompose the signal features into hardware features and channel features. Step 4: Use a cross-reconstruction strategy to avoid information loss during feature decoupling. Optimize the hardware features obtained in step 3 using the center loss to reduce intra-class distances and improve the model's separability for easily confused devices. Step 5: Input the hardware features optimized in step 4 into the classification network to obtain the classification probability; Step 6: Repeat steps 2 to 5 to train the network based on all current training samples, and use the test set samples to test the network effect to obtain the recognition accuracy.

[0008] The signal of the wireless communication device is obtained in step 1, which is expressed as: (1) in, It is the digital signal obtained after the received signal is converted from analog to digital; Signals transmitted by wireless communication devices; is additive Gaussian white noise; is the time-varying impulse response of the wireless channel experienced by the transmitted signal; Indicates the impact of transmitter hardware on the signal; Represents the convolution operation; is the sequence number of the received signal; divided into the training set and the test set that the signal has experienced different.

[0009] Step 2 is as follows: For the data set divided in step 1, it is input into the feature extraction module integrating the dual attention mechanism to initially extract features. The dual attention mechanism is expressed as: (2) in, Represented as a channel attention weight map, Represented as a spatial attention weight map, , Represented as the number of channels of the feature map, represents the characteristic length, Represented as replicated feature vectors by rows or columns.

[0010] Step 3 is as follows: The shape of the feature vector after the feature extraction module is , since the inputs are respectively Data under various channel conditions and experiences For data under various channel conditions, a disentanglement loss function is designed for signal features to separate features. The disentanglement loss function consists of two parts: difference loss and similarity loss. The difference loss is expressed as: (3) in, The number of samples input for a batch, and Respectively indicate the experience The hardware characteristics and channel characteristics of the data, Represents a measure of the correlation between the two parties. Since the data is input in the form of sample pairs during training, the two inputs have the same number of samples. , .

[0011] The similarity loss is expressed as: (4) in, , , are two vectors in the input space, represents the Euclidean distance between two vectors, is the width parameter of the Gaussian kernel, which controls the "smoothness" of the kernel function.

[0012] Step 4 is as follows: Designing a signal decoder , the features of the input decoder are the hardware features extracted by the hardware feature extraction module and the channel features extracted by the channel feature extraction module The task of the decoder is to concatenate or combine these two parts of features, and after nonlinear transformation, output the reconstructed signal features. , and then use the mean squared error (MSE) as the reconstruction loss function to measure and The similarity between them is expressed as: (5) in, , The signal mixing features extracted by the feature extraction module.

[0013] In step 4, the hardware features obtained after disentanglement are optimized using the center loss to make similar features close to each other. The center loss is expressed as: (6) When using the center loss to optimize the model, the category center The update of needs to be synchronized with the sample characteristics, and its specific update formula is: (7) in, Expressed as The hardware feature vector of samples, It is The category labels of samples, Represented as a label Category Center, represents the Euclidean norm, is the update rate, used to control the step size of the center update, Expressed as an indicator function, representing the sample Belongs to category .

[0014] The beneficial effects of the present invention are: 1) This paper proposes a feature extraction module that integrates a dual-attention mechanism. This module uses the synergistic effect of the spatial attention mechanism and the channel attention mechanism to enable the model to more accurately focus on features in the signal that are more conducive to classification and recognition, thereby improving the model's classification performance. 2) This paper proposes a deentanglement module, which uses a twin network as its basic framework. By designing similarity loss and difference loss, it effectively decouples signal characteristics, removes the influence of channel characteristics, and only uses hardware features for identification, thereby achieving channel-robust RF fingerprint recognition.

[0015] 3) This paper proposes an auxiliary module based on cross-domain reconstruction and center loss. This module prevents information loss by constraining cross-domain feature reconstruction, and at the same time introduces center loss to optimize hardware features, so that the features of similar devices are more clustered, effectively improving the model's ability to distinguish easily confused devices, and enhancing the recognition accuracy and robustness of the radio frequency fingerprint recognition system. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is the overall network architecture diagram of the radio frequency fingerprint recognition method based on disentanglement of the present invention; Figure 2 This is a comparison chart of the recognition accuracy of the present invention with the traditional method, the traditional method + disentanglement method, and the traditional method + disentanglement + attention under different signal-to-noise ratios. DETAILED DESCRIPTION

[0017] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] The present invention is based on the radio frequency fingerprint recognition method of de-entanglement, such as Figure 1 As shown in the figure, the original IQ data that has undergone different channel conditions is input into the feature extraction module for preliminary feature extraction, and then the extracted signal mixing features are fed into the hardware feature extraction module and the channel feature extraction module respectively for feature decoupling; in the network's disentanglement module, difference loss and similarity loss are designed for feature decoupling to obtain the hardware features and channel features of the RF signal; after feature decoupling, an auxiliary module based on cross reconstruction and center loss is designed to optimize the model, and finally the hardware features are input into the classification module to obtain the classification probability, and the training data is used batch by batch to train the network parameters.

[0019] Example 1 The present invention is based on the radio frequency fingerprint recognition method of de-entanglement, which is specifically implemented according to the following steps: Step 1: Acquire signals transmitted by wireless communication devices at different times, obtain samples containing different channel information, and construct a data set; Step 2: Input the training set in step 1 into the feature extraction module integrating the dual attention mechanism to preliminarily extract the mixed features of the device signal; Step 3: Build a de-entanglement module and input the mixed signal features extracted in step 2 into the de-entanglement module to decouple the signal features and decompose the signal features into hardware features and channel features. Step 4: Use a cross-reconstruction strategy to avoid information loss during feature decoupling. Optimize the hardware features obtained in step 3 using the center loss to reduce intra-class distances and improve the model's separability for easily confused devices. Step 5: Input the hardware features optimized in step 4 into the classification network to obtain the classification probability; Step 6: Repeat steps 2 to 5 to train the network based on all current training samples, and use the test set samples to test the network effect to obtain the recognition accuracy.

[0020] Example 2 The present invention is based on the radio frequency fingerprint recognition method of disentanglement, wherein the signal of the wireless communication device is obtained in step 1, which is expressed as: (1) in, It is the digital signal obtained after the received signal is converted from analog to digital; Signals transmitted by wireless communication devices; is additive Gaussian white noise; is the time-varying impulse response of the wireless channel experienced by the transmitted signal; Indicates the impact of transmitter hardware on the signal; Represents the convolution operation; is the sequence number of the received signal; divided into the training set and the test set that the signal has experienced different.

[0021] Example 3 The present invention is based on the radio frequency fingerprint recognition method of disentanglement, wherein step 2 is specifically as follows: For the data set divided in step 1, it is input into the feature extraction module integrating the dual attention mechanism to initially extract features. The dual attention mechanism is expressed as: (2) in, Represented as a channel attention weight map, Represented as a spatial attention weight map, , Represented as the number of channels of the feature map, represents the characteristic length, Represented as replicated feature vectors by rows or columns.

[0022] Example 4 The present invention is based on the radio frequency fingerprint recognition method of disentanglement, wherein step 3 is specifically as follows: The shape of the feature vector after the feature extraction module is , since the inputs are respectively Data under various channel conditions and experiences For data under various channel conditions, a disentanglement loss function is designed for signal features to separate features. The disentanglement loss function consists of two parts: difference loss and similarity loss. The difference loss is expressed as: (3) in, The number of samples input for a batch, and Respectively indicate the experience The hardware characteristics and channel characteristics of the data, Represents a measure of the correlation between the two parties. Since the data is input in the form of sample pairs during training, the two inputs have the same number of samples. , .

[0023] The similarity loss is expressed as: (4) in, , , are two vectors in the input space, represents the Euclidean distance between two vectors, is the width parameter of the Gaussian kernel, which controls the "smoothness" of the kernel function.

[0024] Example 5 The present invention is based on the radio frequency fingerprint recognition method of disentanglement, wherein step 4 is specifically as follows: Designing a signal decoder , the features of the input decoder are the hardware features extracted by the hardware feature extraction module and the channel features extracted by the channel feature extraction module The task of the decoder is to concatenate or combine these two parts of features, and after nonlinear transformation, output the reconstructed signal features. , and then use the mean square error as the reconstruction loss function to measure and The similarity between them is expressed as: (5) in , The signal mixing features extracted by the feature extraction module.

[0025] The hardware features obtained after disentanglement are optimized by the center loss so that similar features are close to each other. The center loss is expressed as: (6) When using the center loss to optimize the model, the category center The update of needs to be synchronized with the sample characteristics, and its specific update formula is: (7) in, Expressed as The hardware feature vector of samples, It is The category labels of samples, Represented as a label Category Center, represents the Euclidean norm, is the update rate, used to control the step size of the center update, Expressed as an indicator function, representing the sample Belongs to category .

[0026] Example 6 In the field of RF fingerprinting, deep neural networks can effectively extract RF fingerprints that characterize the essential characteristics of a device. To achieve better generalization and recognition performance, RF fingerprinting methods based on deep neural networks require higher-quality training data and deeper network models. However, in real-world scenarios, when RF signals during the training and recognition phases experience different channel conditions due to factors such as acquisition time and location, the distribution of the collected data can drift, significantly reducing RF fingerprinting accuracy. Therefore, the robustness of RF fingerprinting methods needs to be improved.

[0027] The method of the present invention conducts ablation experiments on four methods: traditional method, traditional method + disentanglement, traditional method + disentanglement + attention, and the method proposed by the present invention under different signal-to-noise ratios. Figure 2 As shown in the figure, the traditional method uses a twin network as the network model; the traditional method + disentanglement introduces a disentanglement loss function based on the traditional architecture to decouple and separate signal features, and only uses hardware features for recognition in the classification stage; the traditional method + disentanglement + attention further introduces a dual attention mechanism based on the previous method to enhance the feature extraction capability of the model; the proposed method is the complete method proposed in this invention.

[0028] The data used comes from the WiSig public dataset, which contains data subsets D1, D2, D3, and D4 collected by six devices on March 1, 8, 15, and 23, 2022. Each transmitter contains 1,000 preamble data points, totaling 6,000. Because the interval between each data set is at least one week, the channel conditions experienced by the transmitted signals in different data subsets inevitably vary.

[0029] Since the signal-to-noise ratio of the signal in the data set is greater than 10dB, Gaussian white noise is added to the original data. Figure 2The signal-to-noise ratio (SNR) in each case is the ratio of the power of the original received signal to the power of the added artificial noise. The actual SNR of the signal after adding noise is less than this value. Data from D1 and D2 were used for training, and data from D4 was used for testing. The results show that the recognition accuracy of the method using only disentanglement is higher than that of the traditional method, and the proposed method has the highest recognition accuracy. At low SNRs, the proposed method achieves a more significant improvement than the other three methods.

[0030] The present invention is based on a radio frequency fingerprint recognition method based on disentanglement. Based on the different causes and characteristics of the channel-related and hardware-related parts of the radio frequency fingerprint characteristics, the two are separated by a designed disentanglement loss function, retaining the hardware characteristics while eliminating the channel characteristics.

Claims

1. A radio frequency fingerprint recognition method based on disentanglement, characterized in that: Specifically, the original IQ data that has undergone different channel conditions is input into the feature extraction module for preliminary feature extraction, and then the extracted signal mixing features are fed into the hardware feature extraction module and the channel feature extraction module respectively for feature decoupling; in the network's disentanglement module, difference loss and similarity loss are designed for feature decoupling to obtain the hardware features and channel features of the RF signal; after feature decoupling, an auxiliary module based on cross reconstruction and center loss is designed to optimize the model, and finally the hardware features are input into the classification module to obtain the classification probability, and the training data is used batch by batch to train the network parameters.

2. The radio frequency fingerprint recognition method based on disentanglement according to claim 1, characterized in that: Please follow the steps below to implement: Step 1: Acquire signals transmitted by wireless communication devices at different times, obtain samples containing different channel information, and construct a data set; Step 2: Input the training set in step 1 into the feature extraction module integrating the dual attention mechanism to preliminarily extract the mixed features of the device signal; Step 3: Build a de-entanglement module and input the mixed signal features extracted in step 2 into the de-entanglement module to decouple the signal features and decompose the signal features into hardware features and channel features. In step 4, a cross-reconstruction strategy is used to avoid information loss during feature decoupling. The hardware features obtained in step 3 are optimized using the center loss to reduce the intra-class distance and improve the separability of the model for easily confused devices. Step 5: Input the hardware features optimized in step 4 into the classification network to obtain the classification probability; Step 6: Repeat steps 2 to 5 to train the network based on all current training samples, and use the test set samples to test the network effect to obtain the recognition accuracy.

3. The radio frequency fingerprint recognition method based on disentanglement according to claim 2, characterized in that: Acquiring the signal of the wireless communication device in step 1 is represented as follows: （1） in, It is the digital signal obtained after the received signal is converted from analog to digital; Signals transmitted by wireless communication devices; is additive Gaussian white noise; is the time-varying impulse response of the wireless channel experienced by the transmitted signal; Indicates the impact of transmitter hardware on the signal; Represents the convolution operation; is the sequence number of the received signal; divided into the training set and the test set that the signal has experienced different.

4. The radio frequency fingerprint recognition method based on disentanglement according to claim 2, characterized in that: The step 2 is specifically as follows: For the data set divided in step 1, it is input into the feature extraction module integrating the dual attention mechanism to initially extract features. The dual attention mechanism is expressed as: （2） in, Represented as a channel attention weight map, Represented as a spatial attention weight map, , Represented as the number of channels of the feature map, represents the characteristic length, Represented as replicated feature vectors by rows or columns.

5. The radio frequency fingerprint recognition method based on disentanglement according to claim 2, characterized in that: The step 3 is specifically as follows: The shape of the feature vector after the feature extraction module is , since the inputs are respectively Data under various channel conditions and experiences For data under various channel conditions, a disentanglement loss function is designed for signal features to separate features. The disentanglement loss function consists of two parts: difference loss and similarity loss. The difference loss is expressed as: （3） in, The number of samples input for a batch, and Respectively represent the channel subdomains Hardware characteristics and channel characteristics, Represents the measure of the correlation between the two sides. Since the data is input in the form of sample pairs during the training process, the two channel subdomain data have the same number of samples. , .

6. The radio frequency fingerprint recognition method based on disentanglement according to claim 5, characterized in that: The similarity loss is expressed as: （4） in, , , are two vectors in the input space, represents the Euclidean distance between two vectors, is the width parameter of the Gaussian kernel.

7. The radio frequency fingerprint recognition method based on disentanglement according to claim 2, characterized in that: The step 4 is specifically as follows: Designing a signal decoder , the features of the input decoder are the hardware features extracted by the hardware feature extraction module and the channel features extracted by the channel feature extraction module The task of the decoder is to concatenate or combine these two parts of features, and after nonlinear transformation, output the reconstructed signal features. , and then use the mean square error as the reconstruction loss function to measure and The similarity between them is expressed as: （5） in, , The signal mixing features extracted by the feature extraction module.

8. The radio frequency fingerprint recognition method based on disentanglement according to claim 7, characterized in that: In step 4, the hardware features obtained after de-entanglement are optimized using the center loss so that similar features are close to each other. The center loss is expressed as: （6） When using the center loss to optimize the model, the category center The update of needs to be synchronized with the sample characteristics, and its specific update formula is: （7） in, Expressed as The hardware feature vector of samples, It is The category labels of samples, Represented as a label Category Center, represents the Euclidean norm, is the update rate, used to control the step size of the center update, Expressed as an indicator function, representing the sample Belongs to category .

Images