Radio frequency fingerprint recognition method based on dual-stream feature embedding and consistency regularization

Abstract

The present invention provides a radio frequency fingerprint recognition method based on dual-stream feature embedding and consistency regularization, which includes the following steps: acquiring and constructing a labeled wireless signal dataset and separating content information from hardware features; importing the wireless signal dataset into a dual-stream feature embedding module to extract hardware features of the original and shielded signals; calculating the cross-entropy classification loss and regularization loss to ensure feature extraction and distribution consistency; optimizing the parameters of the dual-stream feature embedding module based on a comprehensive loss function; dynamically expanding the dual-stream feature embedding module to add new category classifiers to accommodate new devices; deploying the model for online recognition and updating the model weights in real time. Through a consistency-guided robust learning framework, the present invention effectively enhances the recognition performance and robustness of RFF in content tampering scenarios.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication security, and in particular to a radio frequency fingerprint recognition method based on dual-stream feature embedding and consistency regularization, belonging to the technical field of radio frequency fingerprint recognition (RFF). Background Art

[0002] Radio Frequency Fingerprinting (RFF) technology utilizes the physical layer characteristics of wireless transmitting devices (such as minor manufacturing differences in hardware) for identity authentication, offering strong security and unforgeability. However, with the development of deep learning technology, existing RFF methods primarily rely on deep neural networks to automatically extract and classify signal features. While these methods excel in improving recognition accuracy, they also raise new issues, particularly the problem of "content dependency." When wireless signals contain specific content information (such as MAC addresses or other identifiable content), deep learning models tend to rely on this information for classification, thereby ignoring the device's hardware characteristics. This content dependency reduces the model's robustness and security against content tampering or forgery attacks. Summary of the Invention

[0003] The purpose of this invention is to address the content-dependency issue in existing RF fingerprinting methods and provide a content-independent RF fingerprinting method. By introducing a consistency-guided robust learning framework, this invention not only improves recognition accuracy but also enhances the model's robustness against content tampering, enabling RFF technology to function more effectively in more complex application scenarios.

[0004] The specific plan is as follows:

[0005] A radio frequency fingerprint recognition method based on dual-stream feature embedding and consistency regularization, comprising:

[0006] Step 1: Obtain and construct a labeled wireless signal dataset to separate content information and hardware features;

[0007] Step 2: Import the wireless signal dataset into the dual-stream feature embedding module to extract the hardware features of the original signal and the shielded signal;

[0008] Step 3: Calculate the cross entropy classification loss and regularization loss to ensure feature extraction and distribution consistency;

[0009] Step 4: Optimize the parameters of the two-stream feature embedding module based on the comprehensive loss function;

[0010] Step 5: Dynamically expand the two-stream feature embedding module and add new category classifiers to adapt to new devices;

[0011] Step 6: Deploy the model for online recognition and update the model weights in real time.

[0012] Furthermore, in step 1, obtaining and constructing a labeled wireless signal dataset, separating content information and hardware features includes:

[0013] Step 1.1: Collect the original wireless signal from the wireless transmitting device, including the real and imaginary parts of the signal;

[0014] Step 1.2: Parse the signal content (such as MAC address, device ID, etc.) based on the signal transmission protocol and structure to extract the specific content;

[0015] Step 1.3: Preprocess the signal data to remove or mask specific content information so that only features related to the transmitting device hardware are retained in the signal;

[0016] Step 1.4: Based on the device label and signal category, classify the signal samples into positive samples (without specific content) and negative samples (with specific content), and construct a labeled wireless signal dataset, where the positive samples are the original signal streams and the negative samples are the shielded signal streams.

[0017] The core of this step is to ensure that the specific content in the signal data is separated from the hardware fingerprint features and to construct a standardized labeled dataset to support subsequent feature extraction and classification tasks.

[0018] Furthermore, in step 2, the wireless signal dataset is imported into the dual-stream feature embedding module to extract the hardware features of the original signal and the shielded signal, including:

[0019] Step 2.1: Import the wireless signal dataset constructed in step 1 (including the original signal stream and the shielded signal stream) into the dual-stream feature embedding module for processing;

[0020] Step 2.2: The dual-stream feature embedding module processes the original signal stream and the shielded signal stream independently to extract their hardware features. For the shielded signal stream, a content masking matrix is ​​applied to remove specific content before feature extraction.

[0021] Step 2.3: Output two feature vectors, representing the characteristics of the original signal stream And the feature z′ of the shielded signal flow provides data for subsequent regularization;

[0022] The dual-stream feature embedding module ensures that the model can simultaneously process original signals and content-shielded signals, extract feature information related to device hardware, and lay the foundation for reducing dependence on specific content.

[0023] Furthermore, in step 3, the cross entropy classification loss and regularization loss are calculated to ensure feature extraction and distribution consistency, including:

[0024] Step 3.1: Substitute the feature vector z of the original signal stream and the corresponding label value into the cross entropy classification loss function, calculate the classification accuracy, and obtain the classification loss value L CE ;

[0025] Cross-Entropy Loss: Cross-entropy loss is used to measure the difference between the predicted distribution and the true distribution. The formula is as follows:

[0026]

[0027] Where N is the number of samples, C is the number of categories, and y i，c is the true label (0 or 1) of sample i p i，c is the predicted probability that sample i corresponds to category c.

[0028] Step 3.2: Substitute the feature vectors z and z′ of the original signal stream and the masked signal stream into the semantic consistency regularization (SCR) module, use the maximum mean difference (MMD) technique to compare the feature distributions of the two, calculate the distribution difference, and obtain the semantic consistency regularization loss value L SCR ;

[0029] Maximum Mean Discrepancy (MMD): MMD is used to measure the difference between two distributions, especially for unsupervised learning. Its calculation formula is:

[0030]

[0031] Wherein, k(x, y) is a kernel function, such as a Gaussian kernel.

[0032] Step 3.3: Generate the attention matrices A and A′ of the original signal stream and the masked signal stream through the Class Activation Mapping (CAM) technique, align their distributions through the MMD technique, and calculate the attention consistency regularization loss L ACR ;

[0033] Class Activation Mapping (CAM): CAM is used to visualize the area of ​​interest of the CNN model for the input image. The calculation formula is as follows:

[0034]

[0035] Among them, w k is the weight of the last fully connected layer, A k (x) is the feature map of the last convolutional layer.

[0036] In this step, by calculating the cross entropy loss and regularization loss, we ensure that the model can reduce its dependence on specific content through consistency regularization while maintaining classification accuracy, thereby increasing the model's attention to hardware features.

[0037] Furthermore, in step 4, the parameters of the dual-stream feature embedding module are optimized based on the comprehensive loss function, including:

[0038] Step 4.1: Construct the overall loss function, including the classification loss L CE , semantic consistency regularization loss L SCR and attention consistency regularization loss L ACR , the formula is:

[0039] L overall =L CE +λ1L SCR +λ2L ACR

[0040] Among them, λ1 and λ2 are the weights for adjusting each loss term;

[0041] Step 4.2: Optimize the model parameters using the gradient descent algorithm. Update the weights of the two-stream feature embedding module based on minimizing the overall loss function until the loss value converges and the optimal model is obtained.

[0042] This step optimizes the model parameters by minimizing the comprehensive loss function, so that the model can effectively reduce content dependence and improve its robust recognition ability of hardware features while maintaining classification accuracy.

[0043] Furthermore, in step 5, the dual-stream feature embedding module is dynamically expanded and new category classifiers are added to adapt to new devices, including:

[0044] Step 5.1: When a new wireless transmitting device or signal category is detected, dynamically expand the model, add new category labels, and generate a device fingerprint classifier adapted to the new device;

[0045] Step 5.2: In the two-stream feature embedding module, keep the stability module unchanged and introduce the classification ability of new categories through the plasticity module to enhance the recognition of new devices;

[0046] Step 5.3: The dynamically expanded plasticity module continuously improves its adaptability to new devices and new signal types through real-time learning of new signal samples;

[0047] By dynamically expanding new category classifiers, the model has the ability to learn and adapt to new device fingerprints in real time, maintaining efficient recognition of new devices and enhancing the scalability and versatility of the model.

[0048] Furthermore, in step 6, the model is deployed for online recognition and the model weights are updated in real time, including:

[0049] Step 6.1: Deploy the optimal two-stream feature embedding model into the online wireless identification system to process the received wireless signal data in real time and authenticate the transmitter.

[0050] Step 6.2: Based on the prediction results output by the model, determine whether the input signal comes from a known device. If a new device signal is detected, adaptively learn and classify the new signal;

[0051] Step 6.3: Dynamically update model weights through online training to ensure that the system can promptly adapt to new devices and changes in signal content, thereby improving recognition robustness and system security.

[0052] This step ensures that the model can continuously learn and self-optimize in a real-time wireless signal environment, maintains efficient recognition capabilities, ensures that the system can adapt to device changes, and enhances system stability and security.

[0053] The beneficial effects of the present invention are:

[0054] High robustness: By reducing the dependence on specific content, the method of the present invention significantly improves the recognition accuracy and robustness of the model in content tampering scenarios, avoiding the defect of traditional RFF methods that are easily affected by content tampering.

[0055] Improved recognition performance: Combining dual-stream feature embedding and consistency regularization technology, the method of the present invention can better extract hardware fingerprint features in the signal, thereby improving the overall recognition performance of the model.

[0056] Wide applicability: The method of the present invention is applicable to a variety of wireless communication systems, and has important application value in systems involving specific content (such as WiFi, ADS-B, etc.). BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 A schematic diagram of the structure of a robust learning framework based on consistency guidance provided by an embodiment of the present invention.

[0058] Figure 2 Figure 3 is a structural diagram of the dual-stream feature embedding module, showing the feature extraction process of the original signal stream and the content-masked signal stream. DETAILED DESCRIPTION

[0059] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0060] As shown in the figure, the present invention provides a radio frequency fingerprint recognition method based on dual-stream feature embedding and consistency regularization, including:

[0061] Step 1: Obtain and construct a labeled wireless signal dataset to separate content information and hardware features;

[0062] Step 2: Import the wireless signal dataset into the dual-stream feature embedding module to extract the hardware features of the original signal and the shielded signal;

[0063] Step 3: Calculate the cross entropy classification loss and regularization loss to ensure feature extraction and distribution consistency;

[0064] Step 4: Optimize the parameters of the two-stream feature embedding module based on the comprehensive loss function;

[0065] Step 5: Dynamically expand the two-stream feature embedding module and add new category classifiers to adapt to new devices;

[0066] Step 6: Deploy the model for online recognition and update the model weights in real time.

[0067] In this embodiment, obtaining and constructing a labeled wireless signal dataset in step 1, and separating content information from hardware features includes:

[0068] Step 1.1: Collect the original wireless signal from the wireless transmitting device, including the real and imaginary parts of the signal;

[0069] Step 1.2: Parse the signal content (such as MAC address, device ID, etc.) based on the signal transmission protocol and structure to extract the specific content;

[0070] Step 1.3: Preprocess the signal data to remove or mask specific content information so that only features related to the transmitting device hardware are retained in the signal;

[0071] Step 1.4: Based on the device label and signal category, classify the signal samples into positive samples (without specific content) and negative samples (with specific content), and construct a labeled wireless signal dataset, where the positive samples are the original signal streams and the negative samples are the shielded signal streams.

[0072] The core of this step is to ensure that the specific content in the signal data is separated from the hardware fingerprint features and to construct a standardized labeled dataset to support subsequent feature extraction and classification tasks.

[0073] In this embodiment, in step 2, the wireless signal dataset is imported into the dual-stream feature embedding module, and the hardware features of the original signal and the shielded signal are extracted, including:

[0074] Step 2.1: Import the wireless signal dataset constructed in step 1 (including the original signal stream and the shielded signal stream) into the dual-stream feature embedding module for processing;

[0075] Step 2.2: The dual-stream feature embedding module processes the original signal stream and the shielded signal stream independently to extract their hardware features. For the shielded signal stream, a content masking matrix is ​​applied to remove specific content before feature extraction.

[0076] Step 2.3: Output two feature vectors, representing the feature z of the original signal stream and the feature z′ of the shielded signal stream, respectively, to provide data for subsequent regularization;

[0077] The dual-stream feature embedding module ensures that the model can simultaneously process original signals and content-shielded signals, extract feature information related to device hardware, and lay the foundation for reducing dependence on specific content.

[0078] In this embodiment, step 3 calculates the cross entropy classification loss and regularization loss to ensure feature extraction and distribution consistency, including:

[0079] Step 3.1: Substitute the feature vector z of the original signal stream and the corresponding label value into the cross entropy classification loss function, calculate the classification accuracy, and obtain the classification loss value L CE ;

[0080] Step 3.2: Substitute the feature vectors z and z′ of the original signal stream and the masked signal stream into the semantic consistency regularization (SCR) module, use the maximum mean difference (MMD) technique to compare the feature distributions of the two, calculate the distribution difference, and obtain the semantic consistency regularization loss value L SCR ;

[0081] Step 3.3: Generate the attention matrices A and A′ of the original signal stream and the masked signal stream through the Class Activation Mapping (CAM) technique, align their distributions through the MMD technique, and calculate the attention consistency regularization loss L ACR ;

[0082] In this step, by calculating the cross entropy loss and regularization loss, we ensure that the model can reduce its dependence on specific content through consistency regularization while maintaining classification accuracy, thereby increasing the model's attention to hardware features.

[0083] In this embodiment, the optimization of the parameters of the dual-stream feature embedding module based on the comprehensive loss function in step 4 includes:

[0084] Step 4.1: Construct the overall loss function, including the classification loss L CE , semantic consistency regularization loss L SCR and attention consistency regularization loss L ACR, the formula is:

[0085] L overall =L CE +λ1L SCR +λ2L ACR

[0086] Among them, λ1 and λ2 are the weights for adjusting each loss term;

[0087] Step 4.2: Optimize the model parameters using the gradient descent algorithm. Update the weights of the two-stream feature embedding module based on minimizing the overall loss function until the loss value converges and the optimal model is obtained.

[0088] This step optimizes the model parameters by minimizing the comprehensive loss function, so that the model can effectively reduce content dependence and improve its robust recognition ability of hardware features while maintaining classification accuracy.

[0089] In this embodiment, in step 5, the dual-stream feature embedding module is dynamically expanded and a new category classifier is added to adapt to the new device, including:

[0090] Step 5.1: When a new wireless transmitting device or signal category is detected, dynamically expand the model, add new category labels, and generate a device fingerprint classifier adapted to the new device;

[0091] Step 5.2: In the two-stream feature embedding module, keep the stability module unchanged and introduce the classification ability of new categories through the plasticity module to enhance the recognition of new devices;

[0092] Step 5.3: The dynamically expanded plasticity module continuously improves its adaptability to new devices and new signal types through real-time learning of new signal samples;

[0093] By dynamically expanding new category classifiers, the model has the ability to learn and adapt to new device fingerprints in real time, maintaining efficient recognition of new devices and enhancing the scalability and versatility of the model.

[0094] In this embodiment, deploying the model for online recognition and updating the model weights in real time in step 6 includes:

[0095] Step 6.1: Deploy the optimal two-stream feature embedding model into the online wireless identification system to process the received wireless signal data in real time and authenticate the transmitter.

[0096] Step 6.2: Based on the prediction results output by the model, determine whether the input signal comes from a known device. If a new device signal is detected, adaptively learn and classify the new signal;

[0097] Step 6.3: Dynamically update model weights through online training to ensure that the system can promptly adapt to new devices and changes in signal content, thereby improving recognition robustness and system security.

[0098] This step ensures that the model can continuously learn and self-optimize in a real-time wireless signal environment, maintains efficient recognition capabilities, ensures that the system can adapt to device changes, and enhances system stability and security.

[0099] The technical means disclosed in the solutions of the present invention are not limited to those disclosed in the above-mentioned embodiments, but also include technical solutions composed of any combination of the above-mentioned technical features. It should be noted that those skilled in the art may make various improvements and modifications without departing from the principles of the present invention, and such improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A radio frequency fingerprint recognition method based on dual-stream feature embedding and consistency regularization, characterized in that: include: Step 1: Obtain and construct a labeled wireless signal dataset and separate content information from hardware features; Specifically include: Step 1.1: Collect the original wireless signal from the wireless transmitting device, including the real and imaginary parts of the signal; Step 1.2: Parse the signal content information, including MAC address and device ID, based on the signal transmission protocol and structure, and extract these specific contents; Step 1.3: Preprocess the signal to remove or mask specific content information so that only features related to the transmitting device hardware are retained in the signal; Step 1.4: Based on the device label and signal category, the signals are classified into positive samples and negative samples, and a labeled wireless signal dataset is constructed. The positive samples do not contain specific content and are shielded signal streams, while the negative samples contain specific content and are original signal streams. Step 2: Import the wireless signal dataset into the dual-stream feature embedding module to extract the hardware features of the original signal and the shielded signal; Step 3: Calculate the cross entropy classification loss and regularization loss to ensure feature extraction and distribution consistency; specifically, substitute the feature vectors z and z′ of the original signal stream and the shielded signal stream into the semantic consistency regularization SCR module, use the maximum mean difference (MMD) technique to compare the feature distributions of the two, calculate the distribution difference, and obtain the semantic consistency regularization loss value L SCR Generate the attention matrices A and A′ of the original signal stream and the masked signal stream through the class activation mapping CAM technology, align their distributions through the MMD technology, and calculate the attention consistency regularization loss L ACR ; Step 4: Construct a comprehensive loss function, including classification loss L CE , semantic consistency regularization loss L SCR and attention consistency regularization loss L ACR ,Optimize the parameters of the two-stream feature embedding module based on the comprehensive loss function; Step 5: Dynamically expand the two-stream feature embedding module and add new category classifiers to adapt to new devices; Step 6: Deploy the model for online recognition and update the model weights in real time.

2. The radio frequency fingerprint recognition method based on dual-stream feature embedding and consistency regularization according to claim 1 is characterized in that: Step 2 specifically includes: Step 2.1: Import the wireless signal dataset constructed in step 1 into the dual-stream feature embedding module for processing; Step 2.2: The dual-stream feature embedding module processes the original signal stream and the shielded signal stream independently to extract their hardware features. For the shielded signal stream, a content masking matrix is ​​applied to remove specific content before feature extraction. Step 2.3: Output two feature vectors, representing the feature z of the original signal stream and the feature z′ of the shielded signal stream, respectively, to provide data for subsequent regularization.

3. The radio frequency fingerprint recognition method based on dual-stream feature embedding and consistency regularization according to claim 2 is characterized in that: Step 3 specifically includes: Step 3.1: Substitute the feature vector z of the original signal stream and the corresponding label value into the cross entropy classification loss function, calculate the classification accuracy, and obtain the classification loss value L CE ; Cross entropy classification loss function: Cross entropy loss is suitable for measuring the difference between the predicted distribution and the true distribution. The formula is as follows: Where N is the number of samples, C is the number of categories, and y i，c is the true label of sample i, p i，c is the predicted probability that sample i corresponds to category c; Step 3.2: Maximum mean difference: used to measure the difference between two distributions, especially for unsupervised learning, and its calculation formula is: Among them, k(x, y) is the kernel function; Step 3.3: Class Activation Map: This is used to visualize the area of ​​interest of the CNN model for the input image. The calculation formula is as follows: Among them, w k is the weight of the last fully connected layer, A k (x) is the feature map of the last convolutional layer.

4. The radio frequency fingerprint recognition method based on dual-stream feature embedding and consistency regularization according to claim 3 is characterized in that: Step 4 specifically includes: Step 4.1: The comprehensive loss function formula is: L overall =L CE +λ1L SCR +λ2L ACR Among them, λ1 and λ2 are the weights for adjusting each loss term; Step 4.2: Optimize the model parameters using the gradient descent algorithm. Update the weights of the two-stream feature embedding module based on the minimization of the overall loss function until the loss value converges and the optimal model is obtained.

5. The radio frequency fingerprint recognition method based on dual-stream feature embedding and consistency regularization according to claim 4 is characterized in that: Step 5 specifically includes: Step 5.1: When a new wireless transmitting device or signal category is detected, dynamically expand the model, add new category labels, and generate a device fingerprint classifier adapted to the new device; Step 5.2: In the two-stream feature embedding module, keep the stability module unchanged and introduce the classification ability of new categories through the plasticity module to enhance the recognition of new devices; Step 5.3: The dynamically expanded plasticity module continuously improves its adaptability to new devices and new signal types through real-time learning of new signal samples.

6. The radio frequency fingerprint recognition method based on dual-stream feature embedding and consistency regularization according to claim 5 is characterized in that: Step 6 specifically includes: Step 6.1: Deploy the optimal two-stream feature embedding model into the online wireless identification system to process the received wireless signal data in real time and authenticate the transmitter. Step 6.2: Based on the prediction results output by the model, determine whether the input signal comes from a known device. If a new device signal is detected, adaptively learn and classify the new signal; Step 6.3: Dynamically update the model weights through the online training mechanism to ensure that the system can adapt to new devices and signal content changes in a timely manner, thereby improving the robustness of recognition and the security of the system.

Images