A domain-adaptive Wi-Fi radio frequency fingerprint recognition method and device

Abstract

A domain-adaptive Wi-Fi radio frequency fingerprinting method and device comprises: 1) data acquisition and preprocessing; 2) data enhancement: based on the signal, a physical damage-based random data enhancement method is proposed to randomly enhance the frequency and phase of the training data; 3) feature extraction: based on the enhanced data, a method combining steady-state and transient signal log spectra is designed to extract radio frequency fingerprint features as model input; 4) domain adversarial transfer learning network model is constructed; 5) model training and evaluation of detection results; 6) model application: the trained model is applied for recognition in actual electromagnetic scenarios. The proposed method successfully improves the recognition accuracy of existing models for unknown target domain data by introducing an adversarial transfer learning network mechanism into the radio frequency fingerprinting field, while designing a physical damage-based random data enhancement method and a channel-robust time-frequency log spectrum feature. Experiments show that the proposed method improves the accuracy of existing models by approximately 10%-25% in single-domain generalization scenarios compared to existing models, solving the data distribution drift problem of existing models across different temperatures and channel states.

Description

Technical Field

[0001] This invention relates to the field of signal security technology, and in particular to a domain-adaptive Wi-Fi radio frequency fingerprinting method and apparatus. Background Technology

[0002] In recent years, with the rapid development of the Industrial Internet of Things (IIoT), wireless networks and radio frequency signals have been widely used in industrial manufacturing, smart cities, commerce, and military fields. However, most existing wireless network authentication methods are based on link-layer protocols, which are vulnerable to forgery and tampering attacks. For example, attackers can modify device identifiers such as network IP or MAC addresses to impersonate normal users and access the network to steal network data, leading to catastrophic consequences. Therefore, a more stable and effective network security authentication method is needed.

[0003] While physical layer radio frequency (RF) fingerprinting can effectively counter this type of network attack, the robustness of RF fingerprint features is poor due to the complex electromagnetic environment and the influence of temperature and channel conditions in real-world scenarios, making it difficult to cope with different cross-domain scenarios. To address this issue, scholars both domestically and internationally have conducted various studies, which can be mainly divided into: 1) Fingerprint extraction methods based on traditional transient and steady-state signals: These mainly include transient signal features (amplitude envelope, phase, etc.) and steady-state features (carrier frequency offset, steady-state spectrum, etc.), and extract transient fingerprints for device identification through a series of feature transformations; 2) Fingerprint extraction methods based on deep learning: These mainly extract implicit signal features through neural networks, and then achieve RF fingerprint identification and classification through classification networks, improving robustness to different cross-domain scenarios through network parameter optimization. However, these radio frequency fingerprinting methods all have certain drawbacks: for example, radio frequency fingerprints extracted based on traditional transient and steady-state signals rely heavily on expert knowledge and are prone to ignoring potential implicit features in the signal, resulting in low recognition accuracy; while existing fingerprint extraction methods based on deep learning can uncover implicit features in the signal, the distribution of the original data will shift in complex real electromagnetic environments, resulting in poor model robustness and difficulty in dealing with complex cross-domain scenarios.

[0004] Therefore, existing technologies have not achieved significant generalization effects in radio frequency fingerprint recognition. To address these issues, a robust radio frequency fingerprint recognition method is urgently needed in today's complex real-world electromagnetic environments. Summary of the Invention

[0005] To overcome the shortcomings of existing identification methods, this invention proposes a domain-adaptive Wi-Fi radio frequency fingerprint identification method and device to solve the technical problem that the recognition accuracy of deep learning models in existing identification methods drops significantly with changes in ambient temperature or channel.

[0006] The method proposed in this invention introduces an adversarial transfer learning network mechanism in the field of radio frequency fingerprinting, and designs a random data augmentation method based on physical damage and a channel-robust time-frequency logarithmic spectrum feature. This successfully improves the recognition accuracy of existing models for data in unknown target domains. Experiments show that the proposed method improves the accuracy by about 10%-25% compared with existing models in single-domain generalization scenarios, and solves the problem of data distribution drift of existing models across different temperatures and channel conditions.

[0007] The technical solution adopted by the present invention to solve its technical problem is:

[0008] A domain-adaptive Wi-Fi radio frequency fingerprint recognition method, the method comprising the following steps:

[0009] Step S1: Data Acquisition and Preprocessing: Construct a Wi-Fi signal detector to collect radio signals from the air. Detect Wi-Fi signals from these signals through time synchronization and correlation calculations to obtain the original IQ data packets. Slice the data packets to obtain the preamble portion of each data packet as the initial data for extracting RF fingerprint features. Normalize this data to prevent excessive differences in signal amplitude that could lead to overfitting during subsequent model training.

[0010] Furthermore, the data acquisition and preprocessing in step S1 includes the following procedures:

[0011] Step S11: Collect Wi-Fi signals from the air using the USRP-B210 device, and divide the collected long sequence IQ data into windows. Since the short training period of the Wi-Fi signal at a sampling rate of 20MHz has 10 repeated short-lead symbols (STF) and the period is 16 sampling points, shift the signal of each window backward by 16 sampling points, and perform correlation calculation with the original signal without shifting to obtain the correlation coefficient matrix for coarse time synchronization. The calculation method is shown in formula (1):

[0012] r short (n)=y(n+L short )y * (n) (1)

[0013] Among them, L short The period of the short-conducting signal is 16, y * (n) represents the conjugate value of the received signal at the nth sampling point.

[0014] The correlation coefficient matrix is ​​normalized with a threshold of 0.5, and signal start points are determined by finding positions exceeding the threshold. However, since the short lead symbol includes multiple periodic sequences, multiple start points with high correlation will appear, thus requiring further fine-tuning of time synchronization. Fine-tuning time synchronization uses the long lead symbol (LTF) of the Wi-Fi signal preamble, with a period of 64 sampling points at a 20MHz sampling rate. Therefore, the signal in each window is shifted backward by 64 sampling points, and correlation calculation is performed with the original signal without shifting to obtain the long-period correlation coefficient matrix.

[0015] r long (n)=y(n+L long )y * (n) (2)

[0016] Among them, L long The period of the long-conducting signal is 64, y * (n) represents the conjugate value of the received signal at the nth sampling point.

[0017] Select r short (n) and r long (n) Signal points that simultaneously exceed the threshold in the matrix are taken as the starting point of the Wi-Fi data packet. Since the preamble field of the Wi-Fi signal includes 320 sampling points at a sampling rate of 20MHz, and there are certain transient characteristics in the process of the signal rising from the initial noise stage to the signal amplitude, a total of 400 signal sampling points, including the 80 bits before the starting point and the 320 bits after the starting point, are selected for slicing as subsequent model training and testing data.

[0018] Step S12: The data normalization process mentioned in step S1 mainly involves normalizing the data after the final synchronization and slicing in step S11. The specific calculation method is shown in formula (3):

[0019]

[0020] Where M represents the length of the signal, which is set to 400.

[0021] Step S2: Based on the normalized preamble signal data described in Step S1, a data random augmentation method based on physical damage is proposed to randomly augment the frequency and phase of the training set data for subsequent model training. The augmentation method is as follows:

[0022]

[0023] Formula (4) is used for frequency enhancement of the data, and formula (5) is used for phase enhancement of the data. αK and β represent the frequency and phase magnitude of the random enhancement, respectively. k (x b(t),τ) represent the original data expression, Δw k Indicates the original frequency. Let h(t) represent the original phase, and h(t) represent the modeling expression for the wireless channel.

[0024] Step S3: Based on the enhanced dataset from Step S2, design a method that combines the logarithmic spectra of steady-state and transient signals, such as... Figure 1 The diagram shows the structure of the Wi-Fi signal preamble field. The STF signal of the 2nd to 9th period and the LTF signal of the 1st to 2nd period of the signal preamble are selected as steady-state features, as shown in formula (6). At the same time, the STF signal of the 1st to 4th period and the STF and GI signals of the 9th to 10th period of the signal are selected as transient features, as shown in formula (7).

[0025]

[0026]

[0027] Among them, F log (·) indicates the operation of taking the logarithm of the signal power spectrum. This represents the first 16 sampled points of the first LTF cycle, and concat(·) indicates the concatenation operation.

[0028] Finally, the steady-state signal characteristic RFF is analyzed. Stable and transient signal characteristics RFF Transient The features are stitched together to form a robust radio frequency fingerprint for subsequent model training and testing.

[0029] Step S4: Construct a domain adversarial transfer learning network model. Since existing RFID fingerprinting models have poor robustness in complex multi-domain environments, a domain adaptation mechanism is introduced to construct the network model. For example... Figure 2 As shown, it includes a feature extractor module, a label classifier module, and a domain classifier module.

[0030] Furthermore, the feature extractor module, label classifier module, and domain classifier module in step S4 respectively include:

[0031] Step S41: Select the ResNet18 network as the feature extractor module of the model. The specific network structure includes:

[0032] (1) Initialize the convolutional layer

[0033] (2) Residual blocks, which consist of 4 stages, each stage consisting of 2 residual blocks. Each residual block consists of 2 1*3 convolutional layers and has a skip connection.

[0034] (3) Global average pooling layer: After the last residual block, a global average pooling layer is used to aggregate features.

[0035] Step S42: The label classifier module includes two fully connected layers, a batch normalization layer, a ReLU activation function, and the last fully connected layer outputs the number of label categories.

[0036] Step S43: The domain classifier module includes a gradient inversion layer, two fully connected layers, a batch normalization layer, a ReLU activation function, and the last fully connected output channel is the number of domains, which is set to 2 here, representing two domains: the source domain and the target domain. The gradient inversion layer added between the feature extractor and the domain classifier enables the domain classifier to distinguish between the data from the source domain and the target domain, while the feature extractor module cannot determine whether the data comes from the source domain or the target domain, thus extracting the common features that are invariant to the domain.

[0037] Step S5: Model Training and Result Evaluation: Using accuracy as the evaluation metric, train the model and evaluate its signal recognition performance in cross-domain RF fingerprint scenarios. This includes the following steps:

[0038] (1) First, collect Wi-Fi signal data for N days with different channel states using the USRP-B210 device.

[0039] (2) The collected data is processed into preprocessed signal data through step S1.

[0040] (3) Randomly select one day's data as the training set data, i.e. the source domain, and randomly select another day as the test data, i.e. the target domain.

[0041] (4) Perform data augmentation on the training set data using the method in step S2 to obtain the augmented source domain data.

[0042] (5) The source domain data and target domain data are respectively processed by the model in step S4 to obtain the label classifier output matrix and the domain classifier output matrix.

[0043] (6) The loss function is the cross-entropy loss, which is used to calculate the classification output loss L of the source domain data. y Domain classifier loss L of source domain data d1 Domain classifier loss L for target domain data d2 .

[0044] (7) L y +L d1 +L d2 The gradient is used as the total loss of the model for gradient updates.

[0045] Step S6: Model Application: Process the source domain data collected in the actual electromagnetic scene and a small portion of the target domain data in the new scene, and obtain the trained model based on steps S1-S5, then perform the following operations:

[0046] Step S61: Obtain Wi-Fi signal data to be identified in the actual new scenario: Collect the Wi-Fi radio frequency signal to be identified through the radio frequency receiving device, extract the domain-invariant radio frequency fingerprint features, and convert them into the model input format.

[0047] Step S62: Use the trained domain adaptive model to classify and determine the device identity category corresponding to the signal to be identified, thereby realizing Wi-Fi radio frequency fingerprint recognition.

[0048] A second aspect of the present invention relates to a domain-adaptive Wi-Fi radio frequency fingerprint recognition device, comprising a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement a domain-adaptive Wi-Fi radio frequency fingerprint recognition method of the present invention.

[0049] A third aspect of the invention relates to a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements a domain-adaptive Wi-Fi radio frequency fingerprinting method of the present invention.

[0050] The technical concept of this invention is to propose a data random augmentation method based on physical damage, and to design a radio frequency fingerprint feature that combines the logarithmic spectrum of steady-state and transient signals. Furthermore, a domain adaptive mechanism is introduced into the field of radio frequency fingerprint recognition, and a domain adversarial transfer learning network training model is constructed to enable robust recognition of Wi-Fi radio frequency fingerprints under conditions of changing external environment and channel state.

[0051] The beneficial effects of this invention are mainly reflected in the following aspects: by data augmentation and the introduction of a domain adaptation mechanism, radio frequency fingerprint recognition was successfully performed, and compared with traditional machine learning methods, better detection results were achieved. Furthermore, experiments show that the method improves the accuracy by about 10%-25% in single-domain generalization scenarios compared with existing models, and solves the problem of data distribution drift of existing models across different temperatures and channel states. Attached Figure Description

[0052] Figure 1 This is a flowchart of the method of the present invention.

[0053] Figure 2 This is a general framework diagram of the method of the present invention.

[0054] Figure 3 This is a schematic diagram of the structure of the Wi-Fi signal preamble field.

[0055] Figure 4 It is a domain adversarial transfer learning network model structure. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them.

[0057] Example 1

[0058] First, combine Figure 1 This description illustrates a Wi-Fi radio frequency fingerprint recognition method according to one embodiment of the present invention. The method includes the following steps:

[0059] Step S1: Data Acquisition and Preprocessing: Construct a Wi-Fi signal detector to collect radio signals from the air. Detect Wi-Fi signals from these signals through time synchronization and correlation calculations to obtain the original IQ data packets. Slice the data packets to obtain the preamble portion of each data packet as the initial data for extracting RF fingerprint features. Normalize this data to prevent excessive differences in signal amplitude that could lead to overfitting during subsequent model training.

[0060] Furthermore, the data acquisition and preprocessing in step S1 includes the following procedures:

[0061] Step S11: Collect Wi-Fi signals from the air using the USRP-B210 device, and divide the collected long sequence IQ data into windows. Since the short training period of the Wi-Fi signal at a sampling rate of 20MHz has 10 repeated short-lead symbols (STF) and the period is 16 sampling points, shift the signal of each window backward by 16 sampling points, and perform correlation calculation with the original signal without shifting to obtain the correlation coefficient matrix for coarse time synchronization. The calculation method is shown in formula (1):

[0062] r short (n)=y(n+L short )y * (n) (1)

[0063] Among them, L short The period of the short-conducting signal is 16, y * (n) represents the conjugate value of the received signal at the nth sampling point.

[0064] The correlation coefficient matrix is ​​normalized with a threshold of 0.5, and signal start points are determined by finding positions exceeding the threshold. However, since the short lead symbol includes multiple periodic sequences, multiple start points with high correlation will appear, thus requiring further fine-tuning of time synchronization. Fine-tuning time synchronization uses the long lead symbol (LTF) of the Wi-Fi signal preamble, with a period of 64 sampling points at a 20MHz sampling rate. Therefore, the signal in each window is shifted backward by 64 sampling points, and correlation calculation is performed with the original signal without shifting to obtain the long-period correlation coefficient matrix.

[0065] r long (n)=y(n+L long )y * (n) (2)

[0066] Among them, L long The period of the long-conducting signal is 64, y * (n) represents the conjugate value of the received signal at the nth sampling point.

[0067] Select r short (n) and r long (n) Signal points that simultaneously exceed the threshold in the matrix are taken as the starting point of the Wi-Fi data packet. Since the preamble field of the Wi-Fi signal includes 320 sampling points at a sampling rate of 20MHz, and there are certain transient characteristics in the process of the signal rising from the initial noise stage to the signal amplitude, a total of 400 signal sampling points, including the 80 bits before the starting point and the 320 bits after the starting point, are selected for slicing as subsequent model training and testing data.

[0068] Step S12: The data normalization process mentioned in step S1 mainly involves normalizing the data after the final synchronization and slicing in step S11. The specific calculation method is shown in formula (3):

[0069]

[0070] Where M represents the length of the signal, which is set to 400.

[0071] Specifically, the embodiment based on step S1 can be briefly described as follows:

[0072] Using a USRP-B210 device, the acquisition center frequency was set to 2457MHz, the bandwidth to 20MHz, and the sampling rate to 20MHz. Wi-Fi signals from 10 mobile devices with different channel conditions were collected over 9 days as the dataset. Each device collected 1000 training data points and 200 test data points per day, for a total of 108,000 Wi-Fi signal data points. Each data point was normalized according to the method in step S12 to facilitate subsequent feature extraction.

[0073] Step S2: Based on the normalized preamble signal data described in Step S1, a data random augmentation method based on physical damage is proposed to randomly augment the frequency and phase of the training set data for subsequent model training. The augmentation method is as follows:

[0074]

[0075] Formula (4) is used for frequency enhancement of the data, and formula (5) is used for phase enhancement of the data. αK and β represent the frequency and phase magnitude of the random enhancement, respectively. k (x b (t),τ) represent the original data expression, Δw k Indicates the original frequency. Let h(t) represent the original phase, and h(t) represent the modeling expression for the wireless channel.

[0076] Specifically, the embodiment based on step S2 can be briefly described as follows:

[0077] Based on the preprocessed data obtained in step S1, the training data is randomly enhanced by frequency offset and phase offset according to the method in step S2. The frequency offset enhancement range can be set to [0, 2000] Hz, and the phase offset enhancement range can be set to [0, 2] rad, to obtain the enhanced training data.

[0078] Step S3: Based on the enhanced dataset from Step S2, design a method that combines the logarithmic spectra of steady-state and transient signals, such as... Figure 1 The diagram shows the structure of the Wi-Fi signal preamble field. The STF signal of the 2nd to 9th period and the LTF signal of the 1st to 2nd period of the signal preamble are selected as steady-state features, as shown in formula (6). At the same time, the STF signal of the 1st to 4th period and the STF and GI signals of the 9th to 10th period of the signal are selected as transient features, as shown in formula (7).

[0079]

[0080] Among them, F log (·) indicates the operation of taking the logarithm of the signal power spectrum. This represents the first 16 sampled points of the first LTF cycle, and concat(·) indicates the concatenation operation.

[0081] Finally, the steady-state signal characteristic RFF is analyzed. Stable and transient signal characteristics RFF Transient The features are stitched together to form a robust radio frequency fingerprint for subsequent model training and testing.

[0082] Specifically, the embodiment based on step S3 can be briefly described as follows:

[0083] Based on the enhanced data obtained in step S2, the transient and steady-state features of the RF fingerprint that are invariant to the neighborhood are extracted according to the method in step S3. The steady-state and transient features are then combined and stitched together as the input for the subsequent model.

[0084] Step S4: Construct a domain adversarial transfer learning network model. Since existing RFID fingerprinting models have poor robustness in complex multi-domain environments, a domain adaptation mechanism is introduced to construct the network model. For example... Figure 2 As shown, it includes a feature extractor module, a label classifier module, and a domain classifier module.

[0085] Furthermore, the feature extractor module, label classifier module, and domain classifier module in step S4 respectively include:

[0086] Step S41: Select the ResNet18 network as the feature extractor module of the model. The specific network structure includes:

[0087] (1) Initialize the convolutional layer

[0088] (2) Residual blocks, which consist of 4 stages, each stage consisting of 2 residual blocks. Each residual block consists of 2 1*3 convolutional layers and has a skip connection.

[0089] (3) Global average pooling layer: After the last residual block, a global average pooling layer is used to aggregate features.

[0090] Step S42: The label classifier module includes two fully connected layers, a batch normalization layer, a ReLU activation function, and the last fully connected layer outputs the number of label categories.

[0091] Step S43: The domain classifier module includes a gradient inversion layer, two fully connected layers, a batch normalization layer, a ReLU activation function, and the last fully connected output channel is the number of domains, which is set to 2 here, representing two domains: the source domain and the target domain. The gradient inversion layer added between the feature extractor and the domain classifier enables the domain classifier to distinguish between the data from the source domain and the target domain, while the feature extractor module cannot determine whether the data comes from the source domain or the target domain, thus extracting the common features that are invariant to the domain.

[0092] Step S5: Model training and result evaluation. This includes the following steps:

[0093] (1) First, collect Wi-Fi signal data for N days with different channel states using the USRP-B210 device.

[0094] (2) The collected data is processed into preprocessed signal data through step S1.

[0095] (3) Randomly select one day's data as the training set data, i.e. the source domain, and randomly select another day as the test data, i.e. the target domain.

[0096] (4) Perform data augmentation on the training set data using the method in step S2 to obtain the augmented source domain data.

[0097] (5) The source domain data and target domain data are respectively processed by the model in step S4 to obtain the label classifier output matrix and the domain classifier output matrix.

[0098] (6) The loss function is the cross-entropy loss, which is used to calculate the classification output loss L of the source domain data. y Domain classifier loss L of source domain data d1 Domain classifier loss L for target domain data d2 .

[0099] (7) L y +L d1 +L d2 The gradient is used as the total loss of the model for gradient updates.

[0100] Specifically, the embodiment based on step S5 can be briefly described as follows:

[0101] Based on the domain-invariant RF fingerprint features obtained in step S3, the features are input into the domain adversarial transfer learning network model built in step S4 according to the training and testing process in step S5. The learning rate is set to 0.001, the training rounds are set to 200 rounds, the Adam optimizer is selected as the optimizer, and the cross-entropy loss is selected as the loss function for model training.

[0102] Step S6: Model Application: Process the source domain data collected in the actual electromagnetic scene and a small portion of the target domain data in the new scene, and obtain the trained model based on steps S1-S5, then perform the following operations:

[0103] Step S61: Obtain Wi-Fi signal data to be identified in the actual new scenario: Collect the Wi-Fi radio frequency signal to be identified through the radio frequency receiving device, extract the domain-invariant radio frequency fingerprint features, and convert them into the model input format.

[0104] Step S62: Use the trained domain adaptive model to classify and determine the device identity category corresponding to the signal to be identified, thereby realizing Wi-Fi radio frequency fingerprint recognition.

[0105] Specifically, based on the embodiment of step S6, an experiment is set up, using the data from the first day as the source domain data for training, and then using a small amount of data from the remaining 8 days as the target domain for testing to achieve single-domain generalization. New data from the remaining 8 days is then collected for model recognition. The test recognition results for the target domain on the remaining 8 days are 94.10%, 93.70%, 94.00%, 89.90%, 96.55%, 95.75%, 86.20%, and 93.70%, respectively. Simultaneously, experiments are conducted in comparison with methods in existing literature. The results show that the method of this invention improves the accuracy by an average of approximately 10%-25% compared to existing methods.

[0106] This invention proposes a data randomization method based on physical damage, and designs an RF fingerprint feature combining the logarithmic spectrum of steady-state and transient signals, while introducing a domain adaptation mechanism into the field of RF fingerprint recognition. This method aims to improve the cross-domain recognition accuracy of RF fingerprints. Belonging to the field of signal security technology, this method has successfully achieved robust cross-domain recognition of Wi-Fi RF fingerprints.

[0107] Example 2

[0108] This embodiment relates to a domain-adaptive Wi-Fi radio frequency fingerprint recognition device, including a memory and one or more processors. The memory stores executable code, and when the one or more processors execute the executable code, they are used to implement a domain-adaptive Wi-Fi radio frequency fingerprint recognition method according to Embodiment 1.

[0109] Example 3

[0110] This embodiment relates to a computer-readable storage medium storing a program that, when executed by a processor, implements a domain-adaptive Wi-Fi radio frequency fingerprint recognition method according to the present invention.

[0111] The embodiments described in this specification are merely examples of implementations of the inventive concept. The scope of protection of this invention should not be considered as limited to the specific forms stated in the embodiments. The scope of protection of this invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A domain-adaptive Wi-Fi radio frequency fingerprint recognition method, characterized in that, Includes the following steps: Step S1: Data Acquisition and Preprocessing: Construct a Wi-Fi signal detector to collect radio signals from the air. Detect Wi-Fi signals from the signals through time synchronization and correlation calculations to obtain the original IQ data packets. Slice the data packets to obtain the preamble portion of each data packet as the initial data for extracting RF fingerprint features. Normalize the data to prevent excessive differences in signal amplitude from causing overfitting in subsequent model training. Step S2: Based on the normalized preamble signal data described in Step S1, a data random augmentation method based on physical damage is proposed to randomly augment the training set data in terms of frequency and phase for subsequent model training. Step S3: Based on the enhanced dataset from Step S2, design a method that combines the logarithmic spectra of steady-state and transient signals. Select the STF signal of the 2nd to 9th cycles and the LTF signal of the 1st to 2nd cycles of the signal preamble as steady-state features, and simultaneously select the STF signal of the 1st to 4th cycles and the STF and GI signals of the 9th to 10th cycles as transient features. Then the steady-state signal characteristic RFF Stable and transient signal characteristics RFF Transient The features are stitched together to form a robust radio frequency fingerprint for subsequent model training and testing; Step S4: Construct a domain adversarial transfer learning network model. Since the existing RFID fingerprint recognition model has poor robustness in complex multi-domain environments, a domain adaptation mechanism is introduced to construct a network model. The network model includes a feature extractor module, a label classifier module, and a domain classifier module. Step S5: Model training and result evaluation process: Use accuracy as the evaluation metric to train the model and evaluate its signal recognition performance in cross-domain RF fingerprint scenarios; Step S6: Model Application: Process the source domain data collected in the actual electromagnetic scene and a small portion of the target domain data in the new scene, and obtain the trained model based on steps S1-S5, then perform the following operations: Step S61: Obtain Wi-Fi signal data to be identified in the actual new scenario: Collect the Wi-Fi radio frequency signal to be identified through the radio frequency receiving device, extract the domain-invariant radio frequency fingerprint features, and convert them into the model input format; Step S62: Use the trained domain adaptive model to classify and determine the device identity category corresponding to the signal to be identified, thereby realizing Wi-Fi radio frequency fingerprint recognition.

2. The domain-adaptive Wi-Fi radio frequency fingerprint recognition method as described in claim 1, characterized in that, The data acquisition and preprocessing in step S1 includes the following processes: Step S11: Collect Wi-Fi signals from the air using the USRP-B210 device, and divide the collected long sequence IQ data into windows. Since the short training period of the Wi-Fi signal at a sampling rate of 20MHz has 10 repeated short-lead symbols (STF) and the period is 16 sampling points, shift the signal of each window backward by 16 sampling points, and perform correlation calculation with the original signal without shifting to obtain the correlation coefficient matrix for coarse time synchronization. The calculation method is shown in formula (1): r short (n)=y(n+L short )y * (n) (1) Among them, L short The period of the short-conducting signal is 16, y * (n) represents the conjugate value of the received signal at the nth sampling point; The correlation coefficient matrix is ​​normalized with a threshold of 0.5, and signal start points are determined by finding positions exceeding the threshold. However, since the short lead symbol includes multiple periodic sequences, multiple start points with high correlation will appear, thus requiring further fine-tuning of time synchronization. Fine-tuning time synchronization uses the long lead symbol (LTF) of the Wi-Fi signal preamble, with a period of 64 sampling points at a 20MHz sampling rate. Therefore, the signal in each window is shifted backward by 64 sampling points, and correlation calculation is performed with the original signal without shifting to obtain the long-period correlation coefficient matrix. r long (n)=y(n+L long )y * (n) (2) Among them, L long The period of the long-conducting signal is 64, y * (n) represents the conjugate value of the received signal at the nth sampling point; Select r short (n) and r long (n) Signal points that simultaneously exceed the threshold in the matrix are taken as the starting point of the Wi-Fi data packet. Since the preamble field of the Wi-Fi signal includes 320 sampling points at a sampling rate of 20MHz, and there are certain transient characteristics in the process of the signal rising from the initial noise stage to the signal amplitude, a total of 400 signal sampling points, including the 80 bits before the signal starting point and the 320 bits after the starting point, are selected for slicing as subsequent model training and testing data. Step S12: The data normalization process described in step S1 is to normalize the data after the final synchronization and slicing in step S11. The specific calculation method is as shown in formula (3): Where M represents the length of the signal, which is set to 400.

3. The domain-adaptive Wi-Fi radio frequency fingerprint recognition method as described in claim 1, characterized in that, Step S2 specifically includes: Based on the normalized preamble signal data described in step S1, a data augmentation method based on physical damage is proposed to randomly augment the training set data in terms of frequency and phase for subsequent model training. The augmentation method is as follows: Formula (4) is used for frequency enhancement of the data, and formula (5) is used for phase enhancement of the data. αK and β represent the frequency and phase magnitude of the random enhancement, respectively. k (x b (t),τ) represent the original data expression, Δw k Indicates the original frequency. Let h(t) represent the original phase, and h(t) represent the modeling expression for the wireless channel.

4. The domain-adaptive Wi-Fi radio frequency fingerprint recognition method as described in claim 1, characterized in that, In step S3, the STF signal of the 2nd to 9th cycles and the LTF signal of the 1st to 2nd cycles of the preamble portion of the signal are selected as steady-state features using formula (6). The STF signal of the 1st to 4th cycles and the STF and GI signals of the 9th to 10th cycles of the signal are selected as transient features using formula (7). Among them, F log (·) indicates the operation of taking the logarithm of the signal power spectrum. This represents the first 16 sampled points of the first LTF cycle, and concat(·) indicates the concatenation operation.

5. The domain-adaptive Wi-Fi radio frequency fingerprint recognition method as described in claim 1, characterized in that, Step S4 specifically includes: A domain adversarial transfer learning network model is constructed. Due to the poor robustness of existing RFID fingerprint recognition models in complex multi-domain environments, a domain adaptation mechanism is introduced to construct the network model, which includes a feature extractor module, a label classifier module, and a domain classifier module. In step S4, the ResNet18 network is selected as the feature extractor module of the model; the specific network structure includes: (1) Initialize the convolutional layer (2) Residual blocks, which consist of 4 stages, each stage consisting of 2 residual blocks; each residual block consists of 2 1*3 convolutional layers and has a skip connection; (3) Global average pooling layer: After the last residual block, a global average pooling layer is used to aggregate features; The label classifier module described in step S4 includes two fully connected layers, a batch normalization layer, a ReLU activation function, and the output channel of the last fully connected layer is the number of label categories. The domain classifier module described in step S4 includes a gradient inversion layer, two fully connected layers, a batch normalization layer, a ReLU activation function, and the last fully connected output channel is the number of domains, which is set to 2 here, representing two domains: the source domain and the target domain. The gradient inversion layer added between the feature extractor and the domain classifier enables the domain classifier to distinguish between the data from the source domain and the target domain, while the feature extractor module cannot determine whether the data comes from the source domain or the target domain, thus extracting common features that are invariant to the domain.

6. The domain-adaptive Wi-Fi radio frequency fingerprint recognition method as described in claim 1, characterized in that, Step S5 specifically includes: (1) First, collect Wi-Fi signal data for N days with different channel states using the USRP-B210 device; (2) The collected data is processed into preprocessed signal data through step S1; (3) Randomly select one day's data as the training set data, i.e. the source domain, and randomly select another day as the test data, i.e. the target domain; (4) Perform data augmentation on the training set data using the method in step S2 to obtain the augmented source domain data. (5) Obtain the label classifier output matrix and the domain classifier output matrix by passing the source domain data and target domain data through the model in step S4, respectively; (6) The loss function is the cross-entropy loss, which is used to calculate the classification output loss L of the source domain data. y Domain classifier loss L of source domain data d1 Domain classifier loss L for target domain data d2 ; (7) L y +L d1 +L d2 The gradient is used as the total loss of the model for gradient updates.

7. A domain-adaptive Wi-Fi radio frequency fingerprint recognition device, characterized in that, The device includes a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement a domain-adaptive Wi-Fi radio frequency fingerprint recognition method according to any one of claims 1-6.

8. A computer-readable storage medium, characterized in that, It stores a program that, when executed by a processor, implements a domain-adaptive Wi-Fi radio frequency fingerprint recognition method according to any one of claims 1-6.

Images