Radio frequency fingerprinting method and apparatus based on convolutional neural network

Abstract

The application discloses a radio frequency fingerprint identification method and device based on a convolutional neural network, and relates to the technical field of wireless communication.The method comprises the following steps: a source domain loss function, an adversarial domain loss function and a target domain loss function are respectively designed, and a target function is constructed according to the three loss functions; a cross-domain radio frequency fingerprint identification model is constructed based on the target function and the convolutional neural network, and a controlled adversarial update mechanism is introduced to determine target update parameters; the cross-domain radio frequency fingerprint identification model is trained based on obtained source domain radio frequency sample data, device label data, target domain radio frequency sample data and the target update parameters, and a pre-trained cross-domain radio frequency fingerprint identification model is obtained; the obtained to-be-identified radio frequency signal is sequentially input into a pre-trained feature extractor and a pre-trained classifier for fingerprint identification, target radio frequency fingerprint is obtained, and target transmitting device data is determined, so that the technical problem of low radio frequency fingerprint identification accuracy is solved.

Description

Technical Field

[0001] This application relates to the field of wireless communication technology, and in particular to a radio frequency fingerprinting method and apparatus based on a convolutional neural network. Background Technology

[0002] Radio frequency fingerprinting (RF fingerprinting) technology identifies individual wireless transmitting devices by exploiting inherent differences caused by manufacturing defects, suboptimal RF front-end conditions, and other factors. This technology enables device authentication at the physical layer without relying on high-level communication protocols or key mechanisms, and has been applied in scenarios such as physical layer security, wireless network access control, and IoT terminal authentication.

[0003] Existing technologies typically rely on the same distribution assumption. When the model faces distribution shifts caused by differences in the wireless environment during actual deployment, the recognition performance will significantly decrease. Although adversarial adaptive methods have been introduced to achieve cross-domain feature alignment, in the process of multi-objective joint optimization, the domain alignment objective often conflicts with the category discrimination objective. This can easily lead to the cumulative degradation of the discrimination ability that the model has formed in the source domain, thereby limiting the stability and reliability of cross-domain radio frequency fingerprint recognition and resulting in the technical problem of low recognition accuracy of radio frequency fingerprints. Summary of the Invention

[0004] In view of this, embodiments of this application provide a radio frequency fingerprint recognition method and apparatus based on convolutional neural networks, so as to at least solve the technical problem of low recognition accuracy of radio frequency fingerprints.

[0005] According to one aspect of this application, a radio frequency fingerprinting method based on a convolutional neural network is provided. The method includes: acquiring source domain radio frequency sample data from multiple wireless transmitting devices, device tag data corresponding to the source domain radio frequency sample data, and target domain radio frequency sample data; designing a source domain loss function, an adversarial domain loss function, and a target domain loss function, respectively, and constructing an objective function based on the source domain loss function, the adversarial domain loss function, and the target domain loss function; constructing a cross-domain radio frequency fingerprinting model based on the objective function and the convolutional neural network, and introducing a controlled adversarial update mechanism to determine the objective of the cross-domain radio frequency fingerprinting model. Update parameters; based on source domain RF sample data, device tag data, target domain RF sample data, and target update parameters, train the cross-domain RF fingerprint recognition model to obtain a pre-trained cross-domain RF fingerprint recognition model. The pre-trained cross-domain RF fingerprint recognition model includes at least: a pre-trained feature extractor and a pre-trained classifier; acquire the RF signal to be identified and input the RF signal to be identified into the pre-trained feature extractor to obtain the target feature representation; input the target feature representation into the pre-trained classifier for fingerprint recognition to obtain the target RF fingerprint, and determine the target transmitting device data based on the target RF fingerprint.

[0006] Optionally, the source domain loss function can be designed using the following formula, including:

[0007] in, s Used to represent the source domain; Used to represent source domain data in the target batch within the source domain; Used to represent the cross-entropy loss function; Used to represent classifier parameters; Used to represent a classifier; Used to represent feature extractor parameters; Used to represent a feature extractor; Used to indicate in the source domain s The first in i One sample; Used to represent samples The corresponding device label; Used to represent the source domain loss function.

[0008] Optionally, the pre-trained cross-domain RFID fingerprinting model may further include at least: a pre-trained domain discriminator, and an adversarial domain loss function designed using the following formula, including:

[0009] in, Used to represent the target domain; Used to represent target domain data in the target batch within the target domain; Used to represent the parameters of the domain discriminator; Samples used to represent the target domain; Used to represent samples The corresponding field tag; Used to represent the adversarial domain loss function.

[0010] Optionally, the target domain loss function is designed, including: enhancing the target sample in the target domain radio frequency sample data using a first enhancement operator to obtain a first enhanced sample, and enhancing the target sample using a second enhancement operator to obtain a second enhanced sample, wherein the parameters in the first enhancement operator are different from the parameters in the second enhancement operator; inputting the first enhanced sample and the second enhanced sample into a feature extractor to obtain a first feature representation and a second feature representation, and inputting the first feature representation and the second feature representation into a classifier to obtain a first prediction probability and a second prediction probability; constructing a consistency constraint loss function based on the first prediction probability and the second prediction probability; determining a class equilibrium loss function based on the first prediction probability; and determining the target domain loss function based on the consistency constraint loss function and the class equilibrium loss function.

[0011] Optionally, a consistency constraint loss function is constructed based on the first and second prediction probabilities using the following formula, including:

[0012] in, Used to indicate in batch The first in j The first predicted probability of each sample; Used to indicate in batch The first in j The second predicted probability for each sample; Used to represent the second norm; Used to represent the consistency constraint loss function.

[0013] Optionally, determining the class equilibrium loss function based on the first predicted probability includes: determining the average distribution probability of the batch in which the first augmented sample belongs based on the first predicted probability; and constructing the class equilibrium loss function using the divergence function based on the average distribution probability.

[0014] Optionally, a controlled adversarial update mechanism is introduced to determine the target update parameters of the cross-domain radio frequency fingerprinting model, including: obtaining the initial update parameters of the cross-domain radio frequency fingerprinting model in the first target iteration, and determining the gradient update parameters of the cross-domain radio frequency fingerprinting model based on the initial update parameters; determining the candidate update parameters of the cross-domain radio frequency fingerprinting model based on the initial update parameters and the gradient update parameters; using a classifier, determining the initial class loss change based on the initial update parameters and the candidate update parameters; in response to the initial class loss change being greater than zero, smoothing the initial class loss change to obtain the target class loss change; and determining the target update parameters of the cross-domain radio frequency fingerprinting model in the second target iteration based on the target class loss change, wherein the second target iteration is the next iteration after the first target iteration.

[0015] Optionally, based on the change in target category loss, the target update parameters of the cross-domain radio frequency fingerprinting model in the second target iteration are determined, including: determining the moving mean of the cross-domain radio frequency fingerprinting model based on the source domain loss function, and determining the moving variance of the cross-domain radio frequency fingerprinting model based on the source domain loss function and the moving mean; obtaining the threshold parameter of the cross-domain radio frequency fingerprinting model, and determining the target threshold of the cross-domain radio frequency fingerprinting model based on the threshold parameter and the moving variance; determining the update receiving coefficient of the cross-domain radio frequency fingerprinting model based on the change in target category loss and the target threshold; and determining the target update parameters based on the update receiving coefficient, the initial update parameters, and the candidate update parameters.

[0016] According to another aspect of this application, a radio frequency fingerprinting device based on a convolutional neural network is provided. The device includes: a first acquisition unit, configured to acquire source domain radio frequency sample data from multiple wireless transmitting devices, device tag data corresponding to the source domain radio frequency sample data, and target domain radio frequency sample data; a construction unit, configured to design a source domain loss function, an adversarial domain loss function, and a target domain loss function, and construct an objective function based on the source domain loss function, the adversarial domain loss function, and the target domain loss function; and a first determination unit, configured to construct a cross-domain radio frequency fingerprinting model based on the objective function and the convolutional neural network, and introduce a controlled adversarial update mechanism to determine the target update of the cross-domain radio frequency fingerprinting model. The new parameters; the second acquisition unit is used to train the cross-domain radio frequency fingerprint recognition model based on source domain radio frequency sample data, device tag data, target domain radio frequency sample data and target update parameters to obtain a pre-trained cross-domain radio frequency fingerprint recognition model, wherein the pre-trained cross-domain radio frequency fingerprint recognition model includes at least: a pre-trained feature extractor and a pre-trained classifier; the third acquisition unit is used to acquire the radio frequency signal to be identified and input the radio frequency signal to be identified into the pre-trained feature extractor to obtain the target feature representation; the second determination unit is used to input the target feature representation into the pre-trained classifier for fingerprint recognition to obtain the target radio frequency fingerprint, and determine the target transmitting device data based on the target radio frequency fingerprint.

[0017] According to another aspect of this application, a storage medium is provided that stores a computer program thereon, which, when executed by a processor, implements the above-described radio frequency fingerprint recognition method based on a convolutional neural network.

[0018] According to another aspect of this application, a computer device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the program to implement the above-described radio frequency fingerprint recognition method based on a convolutional neural network.

[0019] Using the above technical solution, this application provides a radio frequency fingerprinting method and apparatus based on a convolutional neural network. The method includes: first, acquiring source domain radio frequency sample data, device tag data corresponding to the source domain radio frequency sample data, and target domain radio frequency sample data; then designing a source domain loss function, an adversarial domain loss function, and a target domain loss function, and constructing a target function using the above three loss functions; then, constructing a cross-domain radio frequency fingerprinting model based on the target function and the convolutional neural network, and introducing a controlled adversarial update mechanism to determine the target update parameters; using the source domain radio frequency sample data, device tag data, target domain radio frequency sample data, and target update parameters obtained in the above steps to train the cross-domain radio frequency fingerprinting model to obtain a pre-trained cross-domain radio frequency fingerprinting model; finally, sequentially inputting the radio frequency signal to be identified into a pre-trained feature extractor and classifier to extract features and complete fingerprint recognition to determine the identity of the target transmitting device. This application introduces adversarial domain loss functions and target domain loss functions to force the cross-domain RF fingerprint recognition model to maintain invariant features of the learning domain, thereby reducing distribution differences and improving cross-domain generalization ability. Considering that blind domain alignment will destroy the class discrimination structure and lead to performance degradation, a controlled adversarial update mechanism is introduced to dynamically evaluate and suppress parameter updates that impair discrimination ability. This ensures that the cross-domain RF fingerprint recognition model maintains class discrimination while aligning domain differences, achieving a balance between domain adaptation and discrimination stability. This solves the technical problem of low RF fingerprint recognition accuracy and improves the technical effect of RF fingerprint recognition accuracy.

[0020] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart illustrating a radio frequency fingerprint recognition method based on a convolutional neural network provided in an embodiment of this application is shown. Figure 2 This illustration shows a structural schematic diagram of a cross-domain radio frequency fingerprint recognition model provided in an embodiment of this application; Figure 3 This illustration shows a schematic diagram of the structure of a feature extractor, classifier, and domain discriminator provided in an embodiment of this application; Figure 4This illustration shows a schematic diagram of the structure of a radio frequency fingerprint recognition device based on a convolutional neural network according to an embodiment of this application; Figure 5 A schematic diagram of the device structure of a computer device provided in an embodiment of this application is shown. Detailed Implementation

[0022] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0023] In this embodiment, Figure 1 The illustration shows a flowchart of a radio frequency fingerprint recognition method based on a convolutional neural network according to an embodiment of this application. Figure 1 As shown, the method includes: Step S101: Obtain source domain radio frequency sample data, device tag data corresponding to the source domain radio frequency sample data, and target domain radio frequency sample data from multiple wireless transmitting devices.

[0024] In the technical solution provided in step S101 of this application, source domain radio frequency sample data of multiple wireless transmitting devices, device tag data corresponding to the source domain radio frequency sample data, and target domain radio frequency sample data can be obtained.

[0025] Optionally, source domain RF sample data can be referred to as source domain RF signal data or source domain RF signal samples. Device tag data can be referred to as transmitting device identification tags. Source domain RF signal data and transmitting device identification tags can be referred to as source domain data. Target domain RF sample data can be referred to as target domain RF signal data, target domain RF signal samples, or target domain data. Source domain RF sample data, device tag data, and target domain RF sample data are collectively referred to as RF signal data.

[0026] Optionally, the radio frequency (RF) signal data can be derived from publicly available RF fingerprint datasets or actual wireless communication signal data. For example, publicly available Wireless Fidelity (WiFi) RF fingerprint datasets can be used for experimental verification. This dataset contains a large number of RF signal samples acquired by multiple wireless transmitting devices under different receiver conditions and acquisition environments, effectively reflecting the impact of wireless channel variations and receiver hardware differences on the statistical characteristics of RF signals in real-world wireless environments.

[0027] Optionally, the source domain RF signal data and the target domain RF signal data correspond to the same transmitting device category space, but their RF signal feature distributions differ. A unified data preprocessing operation is performed on the source domain RF signal data and the target domain RF signal data to convert the RF signals into a digital baseband representation suitable for model input, thereby constructing the input dataset for the cross-domain RF fingerprint recognition model.

[0028] Furthermore, source domain data can be represented by the following formula. , in, Used to indicate the first i One source domain radio frequency signal sample, Used to indicate the first i The source domain RF signal sample corresponds to the transmitting device identification tag; the target domain data can be represented as... , in, Used to indicate the first j The target domain RF signal samples do not contain transmitter identification tags. Data preprocessing is performed on the acquired source and target domain data, and a fixed-length complex baseband in-phase / quadrature (I / Q) sequence is selected as the model input. The data preprocessing operations include at least energy detection, channel equalization, and normalization. In this embodiment, a single RF signal sample can be represented as... , in, Used to indicate the first The in-phase components at each sampling point Used to indicate the first Orthogonal components of each sampling point This is used to represent the number of sampling points. After the above processing, the radio frequency signal samples are set into a fixed-dimensional two-dimensional data format to meet the input requirements of the deep learning model.

[0029] Step S102: Design the source domain loss function, the adversarial domain loss function, and the target domain loss function respectively, and construct the target function based on the source domain loss function, the adversarial domain loss function, and the target domain loss function.

[0030] In the technical solution provided in step S102 of this application, three loss functions need to be designed: a source domain loss function, an adversarial domain loss function, and a target domain loss function. Then, based on the above three loss functions, the objective function is constructed.

[0031] Optionally, the source domain loss function can be called the source domain class discrimination loss, and through... This is represented. The adversarial domain loss function can be called the adversarial domain alignment loss, and it is expressed through... This is represented as the target domain loss function, which can be called the target domain structural constraint loss, and is expressed through... To express.

[0032] Optionally, a target function is constructed based on the source domain loss function, the adversarial domain loss function, and the target domain loss function, including: performing a weighted summation of the source domain loss function, the adversarial domain loss function, and the target domain loss function to obtain the target function.

[0033] Optionally, the source domain loss function, adversarial domain loss function, and target domain loss function are weighted and summed using the following formula to obtain the objective function, including: , in, Used to represent the objective function. Used to represent the source domain category discrimination loss. Used to represent adversarial domain alignment loss. This is used to represent the target domain structure constraint loss, which is composed of the enhanced consistency loss and the class equilibrium loss. This parameter represents the weight of the adversarial domain alignment loss in the overall optimization process. These are the weight parameters used to represent the target domain structural constraint loss in the overall optimization process. These weight parameters can be set according to specific application scenarios and training requirements.

[0034] Furthermore, during the backpropagation phase of the cross-domain RF fingerprinting model, a gradient consistency adjustment mechanism is introduced to coordinate the gradient information generated by the category discrimination target and the domain alignment target. This alleviates potential gradient conflicts during multi-objective joint optimization, thereby improving the stability of the model training process. The model's training data can be divided into a training set and a validation set according to a predetermined ratio. During training, early stopping control is implemented based on the performance of the validation set to prevent overfitting. After training is complete, inputting the target domain RF signal sample into the trained model outputs the corresponding transmitter identification result, thus achieving cross-domain RF fingerprinting.

[0035] Step S103: Based on the objective function and convolutional neural network, a cross-domain radio frequency fingerprinting model is constructed, and a controlled adversarial update mechanism is introduced to determine the target update parameters of the cross-domain radio frequency fingerprinting model.

[0036] In the technical solution provided in step S103 of this application, a cross-domain radio frequency fingerprint recognition model can be constructed based on the objective function and the convolutional neural network (CNN). At the same time, a controlled adversarial update mechanism is introduced to determine the target update parameters of the cross-domain radio frequency fingerprint recognition model.

[0037] Optionally, the cross-domain RFID fingerprinting model can be implemented based on a deep learning framework and can be deployed on a general-purpose graphics processing unit (GPU) or other computing platforms. The cross-domain RFID fingerprinting model includes at least: a feature extractor, a classifier, and a domain discriminator. The feature extractor is used to extract feature representations from the input RFID signal; the classifier is used to identify the device identity of the source domain RFID signal data; and the domain discriminator is used to determine the domain category to which the feature representations obtained from the feature extractor belong.

[0038] For example, according to the objective function A cross-domain RFID fingerprinting model, comprising at least a feature extractor, a classifier, and a domain discriminator, is constructed using convolutional neural networks. A controlled adversarial update mechanism is then introduced. When the algorithm calculates a set of candidate update parameters, the controlled adversarial update mechanism first judges: if the candidate update parameter aligns with domain features but impairs the device's classification ability, it is weakened or discarded; if the candidate update parameter aligns and does not compromise classification ability, it is retained. This dynamic filtering process ultimately determines the target update parameter, ensuring that the cross-domain RFID fingerprinting model can balance domain alignment and class discrimination.

[0039] Step S104: Based on source domain RF sample data, device tag data, target domain RF sample data, and target update parameters, train the cross-domain RF fingerprint recognition model to obtain a pre-trained cross-domain RF fingerprint recognition model.

[0040] In the technical solution provided in step S104 of this application, the cross-domain radio frequency fingerprint recognition model is trained based on the source domain radio frequency sample data, device tag data, target domain radio frequency sample data, and target update parameters obtained in the above steps, so as to obtain a pre-trained cross-domain radio frequency fingerprint recognition model. The pre-trained cross-domain radio frequency fingerprint recognition model includes at least a pre-trained feature extractor and a pre-trained classifier. It should be noted that the source domain radio frequency sample data, device tag data, and target domain radio frequency sample data here are all pre-processed data.

[0041] For example, during the training of a cross-domain RFID fingerprint recognition model, the model simultaneously reads source domain RFID sample data with device tag data and target domain RFID sample data without device tag data. The classification loss is calculated using the source domain data and its tags to ensure the model's discriminative power for device categories. Simultaneously, the domain adaptation loss is calculated using the target domain data to narrow the cross-domain feature distribution. Each parameter update is strictly controlled by the target update parameters; if adversarial updates impair classification power, they are suppressed. After multiple iterations, the model can complete pre-training, obtaining a feature extractor and classifier capable of accurately identifying the signal to be identified.

[0042] For another example, a cross-domain radio frequency fingerprint recognition model is jointly trained based on source domain radio frequency signal data and target domain radio frequency signal data. The feature extractor is guided by labeled source domain data to learn radio frequency fingerprint features with device discrimination capabilities. An adversarial learning mechanism is introduced to constrain the consistency of the distribution of source domain feature representation and target domain feature representation in the feature space, thereby obtaining radio frequency fingerprint feature representation with both class discrimination and domain invariance, so as to facilitate the application of the cross-domain radio frequency fingerprint recognition model in subsequent steps.

[0043] In this embodiment, a stable cross-domain RFID fingerprinting model is constructed by jointly training the feature extractor, classifier, and domain discriminator through the synergistic effect of the source domain category discrimination target, the cross-domain adversarial learning process, and the target domain structural constraints. During model training, a controlled adversarial update mechanism based on discriminative performance feedback is introduced to dynamically constrain the parameter update process, including the domain alignment target. This ensures continuous progress in cross-domain alignment while effectively suppressing the adverse effects of adversarial training on the model's discriminative performance, preventing cumulative degradation of discriminative ability during training. Simultaneously, under unlabeled target domain conditions, target domain structural constraints based on enhanced consistency and class balance are introduced to explicitly constrain the stability and class distribution of the target domain prediction results. This ensures consistent discriminative behavior under different enhancement perturbations and acquisition conditions, and prevents the target domain prediction distribution from collapsing towards a minority of classes, thereby stabilizing the target domain's category discrimination structure. Through the aforementioned joint training and structural constraint methods, the radio frequency fingerprint recognition model can balance cross-domain adaptability and discrimination stability under different wireless channel conditions and receiver configurations, achieving stable and reliable cross-domain radio frequency fingerprint recognition even in the case of a labelless target domain.

[0044] Step S105: Acquire the radio frequency signal to be identified and input the radio frequency signal to be identified into a pre-trained feature extractor to obtain the target feature representation.

[0045] In the technical solution provided in step S105 of this application, after obtaining the pre-trained feature extractor, the radio frequency signal to be identified can be acquired, and the radio frequency signal to be identified can be input into the pre-trained feature extractor to obtain the target feature representation of the radio frequency signal to be identified. The radio frequency signal to be identified can be referred to as the target domain radio frequency signal data to be identified.

[0046] For example, after the cross-domain radio frequency fingerprint recognition model is trained, it can enter the inference stage of the model. In the inference stage, the radio frequency signal data of the target domain to be identified is input into the trained feature extractor to obtain the corresponding target feature representation.

[0047] Step S106: Input the target feature representation into a pre-trained classifier for fingerprint recognition to obtain the target radio frequency fingerprint, and determine the target transmitting device data based on the target radio frequency fingerprint.

[0048] In the technical solution provided by step S106 of this application, the obtained target feature representation is input into a pre-trained classifier for fingerprint recognition to obtain the target radio frequency fingerprint, and then the target transmitting device data is determined based on the target radio frequency fingerprint. The target transmitting device data can be referred to as the transmitting device identification tag.

[0049] For example, after obtaining the target feature representation corresponding to the target domain radio frequency signal data to be identified, the target feature representation is input into a pre-trained classifier. Based on the maximum probability output of the classification, the target radio frequency fingerprint is determined, thereby realizing the identification label of the corresponding transmitting device.

[0050] Furthermore, since the domain discriminator and the target domain structure constraint module are only used to guide the feature learning process during the model training phase, there is no need to introduce the domain discriminator and structure constraint calculation during the inference phase. The cross-domain radio frequency fingerprint recognition task can be completed using only the trained feature extractor and classifier.

[0051] In the technical solution provided by steps S101 to S106 of this application, firstly, source domain radio frequency sample data, device tag data corresponding to the source domain radio frequency sample data, and target domain radio frequency sample data are obtained. Then, source domain loss function, adversarial domain loss function, and target domain loss function are designed, and a target function is constructed using the above three loss functions. Then, based on the target function and convolutional neural network, a cross-domain radio frequency fingerprint recognition model is constructed, and a controlled adversarial update mechanism is introduced to determine the target update parameters. Using the source domain radio frequency sample data, device tag data, target domain radio frequency sample data, and target update parameters obtained in the above steps, the cross-domain radio frequency fingerprint recognition model is trained to obtain a pre-trained cross-domain radio frequency fingerprint recognition model. Finally, the radio frequency signal to be identified is sequentially input into the pre-trained feature extractor and classifier to extract features and complete fingerprint recognition to determine the identity of the target transmitting device. This application introduces adversarial domain loss functions and target domain loss functions to force the cross-domain RF fingerprint recognition model to maintain invariant features of the learning domain, thereby reducing distribution differences and improving cross-domain generalization ability. Considering that blind domain alignment will destroy the class discrimination structure and lead to performance degradation, a controlled adversarial update mechanism is introduced to dynamically evaluate and suppress parameter updates that impair discrimination ability. This ensures that the cross-domain RF fingerprint recognition model maintains class discrimination while aligning domain differences, achieving a balance between domain adaptation and discrimination stability. This solves the technical problem of low RF fingerprint recognition accuracy and improves the technical effect of RF fingerprint recognition accuracy.

[0052] The method described in this embodiment will be further described below.

[0053] As an optional implementation method, the source domain loss function is designed using the following formula, including:

[0054] in, s Used to represent the source domain; Used to represent source domain data in the target batch within the source domain; Used to represent the cross-entropy loss function; Used to represent classifier parameters; Used to represent a classifier; Used to represent feature extractor parameters; Used to represent a feature extractor; Used to indicate in the source domain s The first in i One sample; Used to represent samples The corresponding device label; Used to represent the source domain loss function.

[0055] In this embodiment, the source domain sample set is assumed to be... The target domain sample set is The feature extractor parameters are: The classifier parameters are The source domain category discrimination loss can be expressed by the following formula:

[0056] in, Used to represent the cross-entropy loss function.

[0057] As an optional implementation, the pre-trained cross-domain RFID fingerprinting model may at least include: a pre-trained domain discriminator, and an adversarial domain loss function designed using the following formula, including:

[0058] in, Used to represent the target domain; Used to represent target domain data in the target batch within the target domain; Used to represent the parameters of the domain discriminator; Samples used to represent the target domain; Used to represent samples The corresponding field tag; Used to represent the adversarial domain loss function.

[0059] In this embodiment, the adversarial domain loss function can be simply referred to as the domain alignment loss, which can be defined as:

[0060] in, Used to represent field labels.

[0061] As an optional embodiment, the target domain loss function is designed as follows: A first enhanced sample is obtained by enhancing a target sample in the target domain radio frequency sample data using a first enhancement operator; a second enhanced sample is obtained by enhancing the target sample using a second enhancement operator, wherein the parameters in the first enhancement operator are different from the parameters in the second enhancement operator; the first and second enhanced samples are respectively input into a feature extractor to obtain a first feature representation and a second feature representation; the first and second feature representations are respectively input into a classifier to obtain a first prediction probability and a second prediction probability; a consistency constraint loss function is constructed based on the first and second prediction probabilities; a class equilibrium loss function is determined based on the first prediction probability; and a target domain loss function is determined based on the consistency constraint loss function and the class equilibrium loss function.

[0062] In this embodiment, the target sample in the target domain radio frequency sample data is first enhanced by a first enhancement operator to obtain a first enhanced sample. Then, the target sample is enhanced by a second enhancement operator to obtain a second enhanced sample. The first and second enhanced samples are then input into a feature extractor to obtain a first feature representation and a second feature representation. The first and second feature representations are then input into a classifier to obtain a first prediction probability and a second prediction probability. Based on the first and second prediction probabilities, a consistency constraint loss function is constructed. Simultaneously, based on the first prediction probability, a class equilibrium loss function can be determined. Finally, based on the consistency constraint loss function and the class equilibrium loss function, the target domain loss function is determined.

[0063] Optionally, the first enhancement operator can be determined by a first enhancement function, which can be called the first identity-preserving random perturbation operator, and is determined by... This is represented. The target sample can be any target domain sample in the target domain RF sample data, and is represented by... The second enhancement operator is determined by the second enhancement function and can be called the second identity-preserving random perturbation operator. To express.

[0064] Optionally, and These are used to represent different random enhancement transformation operators, used to simulate channel variation information and receiver difference information that may occur in cross-domain environments. The first enhancement sample can be called the first enhancement view, and is obtained through... This is represented. The second augmented sample can be called the second augmented view, and is expressed through... To express.

[0065] Optionally, the first enhancement operator and the second enhancement operator can be obtained by selecting one of the following: random amplitude scaling operator, random phase rotation operator, carrier frequency offset perturbation operator, additive complex Gaussian noise operator, and lightweight multipath convolution operator, but the operators selected by the first enhancement operator and the second enhancement operator are different.

[0066] Optionally, the first predicted probability is obtained by... The second predicted probability is represented. This is represented. The consistency constraint loss function is simply called the consistency constraint, and it is expressed through... This is represented. The class-equalization loss function is simply called class-equalization loss, and it is expressed through... To express.

[0067] In this embodiment of the application, the target domain sample set is assumed to be For any target domain sample Two enhanced views, namely the first enhanced view and the second enhanced view, are generated using two different identity-preserving random perturbation operators:

[0068] The enhancement operators are used to simulate channel variations and receiver differences that may occur in cross-domain environments. In this embodiment, the enhancement operators include the following: a random amplitude scaling operator, which is expressed by the following formula: ,

[0069] in, A time index used to represent discrete sampling. Used to represent the original radio frequency signal in the 1st Complex baseband signals at each sampling point; Used to represent signals that have undergone random amplitude scaling; Used to represent the random amplitude scaling factor; Used to represent uniform distribution; and These represent the minimum and maximum values ​​of the amplitude scaling factor, respectively. The random phase rotation operator is expressed by the following formula: ,

[0070] in, Used to represent signals after random phase rotation; Used to represent random phase offset; Used to represent the imaginary unit. The carrier frequency offset perturbation operator is expressed by the following formula:

[0071] in, Used to represent the signal after carrier frequency offset disturbance; Used to represent the normalized carrier frequency offset parameter. The additive complex Gaussian noise operator is expressed by the following formula: ,

[0072] in, Used to represent the signal after adding complex Gaussian noise; Used to represent complex Gaussian random noise; Used to represent a mean of 0 and a variance of 0. The complex Gaussian distribution; Used to represent noise power. The lightweight multipath convolution operator is expressed by the following formula:

[0073] in, Used to represent the signal after lightweight multipath convolution; K is used to represent the number of taps in the multipath channel; Used to indicate the first k The complex fading coefficients corresponding to each path. It should be noted that the above perturbation parameters are randomly sampled within a preset physical reasonable range to ensure that the enhancement operation simulates changes in the real wireless environment while avoiding alteration of the device's inherent radio frequency characteristics. This enhancement method allows the model to be exposed to cross-domain perturbation modes in advance during the training phase, thereby reducing discrimination instability caused by channel changes during actual deployment.

[0074] Furthermore, the enhanced first and second samples are sequentially input into the feature extractor and classifier, respectively, to obtain the corresponding predicted probabilities:

[0075] in, Used to represent the normalized exponential function.

[0076] In this embodiment, the target domain loss function is obtained by weighted summation of the consistency constraint loss function and the class equilibrium loss function, and is expressed by the following formula:

[0077] in, The balance coefficient is used to adjust the weight between sample-level stability and batch-level structure unfolding. Furthermore, by jointly optimizing the consistency constraint loss function and the class equilibrium loss function, without relying on hard or pseudo-label supervision in the target domain, the target domain samples maintain predictive stability under enhanced perturbations and preserve class structure unfolding at the overall distribution level. This complements the aforementioned controlled adversarial update mechanism; the former ensures that the discriminative ability does not degrade during domain alignment, while the latter explicitly shapes a stable class structure within the target domain. Together, these mechanisms enhance the stability and robustness of the cross-domain RF fingerprinting model in complex wireless environments.

[0078] In this embodiment, after completing adversarial cross-domain feature learning and controlling the adversarial update process of the model, to address the issues of unstable prediction results and aliasing or collapse of the class structure of target domain radio frequency signal samples under unlabeled conditions, a target domain structure constraint mechanism based on enhanced consistency and class equalization is introduced to explicitly stabilize the class discrimination structure of the target domain. Specifically, for each target domain radio frequency signal sample, at least two enhanced views are generated that maintain the device identity information. These enhanced views are obtained by applying simulated channel and receiver perturbations to the original radio frequency signal. Transformations include, but are not limited to, random amplitude scaling, random phase rotation, carrier frequency offset perturbation, time alignment offset, additive noise perturbation, and lightweight multipath filtering, thereby obtaining a first enhanced sample and a second enhanced sample for the same target domain sample. The first and second enhanced samples are then input into the feature extractor and classifier, respectively, to obtain the corresponding class prediction distribution.

[0079] Furthermore, at the target domain sample level, by imposing consistency constraints on the class prediction distributions of the first and second enhanced samples, the prediction results of the same target domain samples under different enhancement perturbations are kept consistent, thereby improving the model's robustness to channel noise and receiver-side differential perturbations and enhancing the stability of the target domain representation. At the target domain batch level, by statistically analyzing the class prediction distributions of target domain samples, a class distribution balance constraint is constructed to suppress the distribution bias and collapse caused by the concentration of target domain prediction results in a few classes, thus avoiding trivial solutions under consistency constraints and maintaining the expansiveness of the overall class structure of the target domain. By jointly optimizing the consistency constraint and class balance constraint, without relying on the target domain device identity label, the intra-class compactness and inter-class separability of the target domain RF fingerprint features are explicitly enhanced, improving the discriminative performance, training stability, and result reliability of the cross-domain RF fingerprint recognition model.

[0080] As an optional implementation method, a consistency constraint loss function is constructed based on a first prediction probability and a second prediction probability using the following formula, including:

[0081] in, Used to indicate in batch The first in j The first predicted probability of each sample; Used to indicate in batch The first in j The second predicted probability for each sample; Used to represent the second norm; Used to represent the consistency constraint loss function.

[0082] In this embodiment, to enhance the invariance of the cross-domain radio frequency fingerprinting model to environmental disturbances, a consistency constraint is imposed on the prediction results of the two views of the same sample, which is expressed by the following formula:

[0083] in, This is used to represent the consistency constraint loss function, simply called the consistency constraint. The purpose of this consistency loss is to suppress the model's oversensitivity to input perturbations, making the feature representation more focused on inherent device differences rather than channel factors. As the model trains, this constraint encourages the target domain samples to form a more stable discrimination boundary in the latent space.

[0084] As an optional implementation method, determining the class equilibrium loss function based on the first predicted probability includes: determining the average distribution probability of the batch in which the first augmented sample belongs based on the first predicted probability; and constructing the class equilibrium loss function using a divergence function based on the average distribution probability.

[0085] In this embodiment, after obtaining the first predicted probability, the average distribution probability of the batch to which the first augmented sample belongs can be determined based on the first predicted probability. Then, based on the average distribution probability, a class equilibrium loss function is constructed using the divergence function. The average distribution probability is determined by... The divergence function can be a preset divergence (KL divergence), and can be expressed through... This is represented. The class equilibrium loss function can be called the class equilibrium constraint.

[0086] In this embodiment, relying solely on the consistency constraints designed using the above steps may result in a "prediction collapse" phenomenon, where the model predicts most target domain samples as belonging to the same category to meet the consistency requirements. To avoid this trivial solution, a class equilibrium constraint is introduced at the batch level. In this case, the average distribution of mini-batch predictions for the target domain can be set to determine the class equilibrium loss function.

[0087] Optionally, the average distribution probability of the batch to which the first enhanced sample belongs is determined based on the first predicted probability using the following formula, including:

[0088] in, This represents the average probability distribution of the batch to which the first augmented sample belongs; the average probability distribution can be simply referred to as the average distribution. Further, based on the average probability distribution, a class-equilibrium loss function is constructed using the divergence function, as shown in the following formula:

[0089] in, This is used to represent a uniform distribution vector. This constraint encourages the predicted distribution of the target domain to remain spread across the categories, thereby maintaining the diversity of the category structure at the overall level.

[0090] In this embodiment, after completing the controlled adversarial update process, to further stabilize the target domain category structure under the condition of no target domain label, an enhanced consistency constraint and a class equalization constraint are introduced. These constraints are designed based on the physical feature formation mechanism of RF fingerprints. Since the RF fingerprint features of transmitting devices mainly originate from the non-ideal nature of the RF front-end hardware and differences in manufacturing processes, and factors such as channel fading, receiver gain changes, and carrier frequency offset primarily affect the external observation form of the signal, enhancing the signal within a reasonable perturbation range should not change its device identity attributes. Based on this principle, by constructing enhanced samples with multiple views for the target domain samples and applying consistency constraints to the prediction results, the robustness of the model to environmental perturbations can be effectively improved, and the category discrimination structure of the target domain can be explicitly stabilized.

[0091] As an optional implementation, a controlled adversarial update mechanism is introduced to determine the target update parameters of the cross-domain RF fingerprinting model. This includes: obtaining the initial update parameters of the cross-domain RF fingerprinting model in the first target iteration, and determining the gradient update parameters of the cross-domain RF fingerprinting model based on the initial update parameters; determining candidate update parameters of the cross-domain RF fingerprinting model based on the initial update parameters and the gradient update parameters; using a classifier, determining the initial class loss change based on the initial update parameters and the candidate update parameters; smoothing the initial class loss change in response to the initial class loss change being greater than zero to obtain the target class loss change; and determining the target update parameters of the cross-domain RF fingerprinting model in the second target iteration based on the target class loss change, wherein the second target iteration is the next iteration after the first target iteration.

[0092] In this embodiment, the initial update parameters of the cross-domain RF fingerprint recognition model in the first target iteration are first obtained. Based on the initial update parameters, the gradient update parameters of the cross-domain RF fingerprint recognition model are determined. Then, based on the initial update parameters and the gradient update parameters, the candidate update parameters of the cross-domain RF fingerprint recognition model are determined. Next, using a classifier, the initial class loss change is determined based on the initial update parameters and the candidate update parameters obtained above. At this time, the initial class loss change is compared with zero. If the initial class loss change is greater than zero, the initial class loss change is smoothed to obtain the target class loss change. Finally, based on the target class loss change, the target update parameters of the cross-domain RF fingerprint recognition model in the second target iteration can be determined.

[0093] Optionally, the initial update parameters can be called the current parameters. The first target iteration can be achieved through...t If we represent it, then the initial update parameters at this point can be expressed as: The smoothing process can be called moving average smoothing.

[0094] Optionally, the gradient update parameters of the cross-domain radio frequency fingerprinting model are determined based on the initial update parameters using the following formula, including:

[0095] in, The gradient update parameters used to represent the cross-domain radio frequency fingerprinting model at the t-th iteration are the joint update direction of the model. Used to indicate a pair Perform the differentiation operation. The domain alignment weight coefficient is used to control the strength of the influence of the domain alignment target on parameter updates. Using the above formula, the cross-domain RFID fingerprint recognition model can be realized in the [missing information - likely a specific region or area]. t In the next iteration, based on the current parameters Calculate the joint update direction.

[0096] Optionally, candidate update parameters for the cross-domain RFID fingerprinting model are determined based on the initial update parameters and gradient update parameters using the following formula, including:

[0097] in, Used to represent candidate update parameters Used to represent the learning rate.

[0098] Furthermore, to determine whether the candidate update parameters disrupt the discriminative structure, the change in class discriminative loss before and after the update is calculated on a mini-batch of samples in the source domain. Specifically, the change in initial class loss is determined using the classifier based on the initial update parameters and candidate update parameters, according to the following formula:

[0099] in, Used to represent the change in initial category loss. Used to indicate when the parameter is Feature extractor for time. Used to indicate when the parameter is The feature extractor at that time. At this point, it checks whether the change in the initial class loss is greater than zero. This indicates that the candidate updated parameters cause a decrease in the model's discriminative performance in the current batch. Since single-batch fluctuations may be affected by random noise, a moving average smoothing process is needed to smooth the initial class loss variation in order to improve the stability of the judgment.

[0100] Optionally, in response to a positive change in the initial class loss, the change in the initial class loss is smoothed using the following formula to obtain the change in the target class loss, including:

[0101] in, This is used to represent the change in target category loss and can be called the change in discrimination performance. Used to represent the smoothing coefficient.

[0102] In this embodiment, during the adversarial cross-domain RF fingerprint feature learning process, the parameter update process of the feature extraction model containing the domain alignment target is controlled and adjusted to suppress the adverse effects of adversarial training on the model's discrimination performance. Specifically, in each parameter update process containing the domain alignment target, candidate update parameters are first generated based on the current model parameters, and the change in discrimination performance before and after the candidate update is evaluated on the source domain RF signal data to obtain the corresponding change in discrimination performance. Based on historical discrimination performance changes, a dynamic threshold for determining the acceptability of candidate updates is adaptively determined.

[0103] Furthermore, when the change in discrimination performance is within the dynamic threshold range, candidate update parameters are accepted; when the change in discrimination performance exceeds the dynamic threshold, a soft backoff adjustment mechanism is triggered. The candidate update parameters are weighted and fused with the parameters before the update to obtain the final parameters used to update the feature extractor, i.e., the target update parameters. The fusion weights are adaptively determined based on the relative relationship between the change in discrimination performance and the dynamic threshold. Through this method, while ensuring the continuous advancement of the domain alignment process, the adversarial update process of the model is effectively constrained, preventing cumulative degradation of discrimination performance during training. This improves the stability and recognition reliability of the cross-domain RF fingerprint recognition model during adversarial adaptive training.

[0104] As an optional embodiment, the target update parameters of the cross-domain radio frequency fingerprinting model in the second target iteration are determined based on the change in target category loss, including: determining the moving mean of the cross-domain radio frequency fingerprinting model based on the source domain loss function, and determining the moving variance of the cross-domain radio frequency fingerprinting model based on the source domain loss function and the moving mean; obtaining the threshold parameter of the cross-domain radio frequency fingerprinting model, and determining the target threshold of the cross-domain radio frequency fingerprinting model based on the threshold parameter and the moving variance; determining the update receiving coefficient of the cross-domain radio frequency fingerprinting model based on the change in target category loss and the target threshold; and determining the target update parameters based on the update receiving coefficient, the initial update parameters, and the candidate update parameters.

[0105] In this embodiment, the moving average of the cross-domain RF fingerprinting model is first determined based on the source domain loss function. Then, the moving variance of the cross-domain RF fingerprinting model is determined based on the source domain loss function and the moving average. Next, the threshold parameters of the cross-domain RF fingerprinting model are obtained. Based on the threshold parameters and the moving variance, the target threshold of the cross-domain RF fingerprinting model is determined. Then, based on the target category loss change and the target threshold obtained above, the updated receiving coefficients of the cross-domain RF fingerprinting model are determined. Finally, based on the updated receiving coefficients, the initial update parameters, and the candidate update parameters, the target update parameters can be determined. The moving average can be obtained through… This can be represented. The sliding variance can be expressed as follows: To express.

[0106] Optionally, to avoid overly restricting the exploration space in the early stages of model training, an adaptive dynamic threshold is constructed for the source domain discrimination loss. First, the mean and fluctuation amplitude of the loss are calculated using a sliding statistics method. Specifically, the sliding mean of the cross-domain RF fingerprint recognition model is determined based on the source domain loss function using the following formula:

[0107] in, Used to indicate the current training iteration t When the source domain determines the moving average of the loss, the central tendency of the loss can be estimated. Used to represent the sliding statistical smoothing coefficient, with a value range of [value missing]. This is used to control the degree to which the current iteration loss affects the mean and variance. Used to indicate that in the training iteration t When -1, the moving mean of the source domain discrimination loss. The moving variance of the cross-domain RF fingerprint recognition model is determined using the following formula, based on the source domain loss function and the moving mean, including:

[0108] in, Used to indicate the current training iteration t At that time, the sliding variance (or fluctuation range) of the source domain is used to determine the loss, which is used to quantify the magnitude of the loss change. Used to indicate the current training iteration t When -1, the sliding variance of the source domain discrimination loss. The discriminative loss used to represent the source domain category in the iteration t The value at that time.

[0109] Optionally, the threshold parameter can be obtained through This is represented. The target threshold can be called the adaptive dynamic threshold, or simply the dynamic threshold, and is expressed through... To express.

[0110] Optionally, the target threshold for the cross-domain radio frequency fingerprinting model is determined based on the threshold parameter and the sliding variance using the following formula, including:

[0111] in, Used to indicate the current training iteration t An adaptive dynamic threshold is used to determine whether updating candidate parameters will disrupt the discriminant structure. This is used to represent the sensitivity coefficient, which controls the weight of loss fluctuation amplitude in the dynamic threshold calculation. This represents the minimum threshold constant, ensuring that the dynamic threshold is not too small in the early stages of training, allowing the model to explore the cross-domain alignment space. It should be noted that this threshold remains loose in the early stages of training due to its large fluctuations, allowing the model to fully explore the cross-domain alignment space; as training stabilizes and loss fluctuations decrease, the threshold automatically tightens, thereby strengthening the protection of the discriminative structure.

[0112] Optionally, based on the change in discrimination performance after smoothing With dynamic threshold The relationship is defined to establish a continuous update acceptance coefficient, which is determined by the following formula based on the change in target category loss and the target threshold, to establish the update acceptance coefficient of the cross-domain radio frequency fingerprinting model, including:

[0113] in, Used to represent sensitivity parameters. Used to represent the stability constant. Used to indicate the updated receive coefficient.

[0114] Optionally, the target update parameter is determined based on the updated reception coefficient, the initial update parameter, and the candidate update parameter using the following formula, including:

[0115] in, This is used to represent the target update parameters, which are used to represent the model's performance in... t +1 training iterations of parameters. The soft backoff mechanism described above offers better continuity and stability compared to full backoff or hard threshold pruning. When candidate parameter updates do not trigger discriminative degradation, When the value approaches 1, the model performs a full adversarial update; as the degradation increases, It automatically reduces the adversarial update magnitude, thus smoothly decreasing it rather than completely rejecting updates, thereby avoiding training stagnation.

[0116] Furthermore, the domain alignment weights can be adjusted synchronously, for example:

[0117] When the discrimination performance degrades significantly, the domain alignment strength is automatically reduced, allowing the model to strengthen the category discrimination structure before gradually restoring the alignment process.

[0118] In this embodiment, the controlled adversarial update mechanism based on discriminative performance feedback transforms the potentially negative transfer of the multi-objective joint optimization process into a dynamically controlled parameter evolution process, achieving a balance between domain-invariant learning and class discrimination capability. This mechanism works in synergy with the subsequent target domain structure constraint module; the former protects the discriminative structure during the cross-domain alignment phase, while the latter shapes a stable class structure within the target domain. Together, they improve the training stability and recognition reliability of the cross-domain RF fingerprint recognition model in complex wireless environments, thus addressing the technical problem of low RF fingerprint recognition accuracy.

[0119] Furthermore, during the training of the adversarial cross-domain RF fingerprinting model, the feature extractor parameters are jointly constrained by both the source domain class discrimination objective and the domain alignment objective. Since the domain alignment objective aims to suppress feature components related to channel and receiver differences, while the class discrimination objective emphasizes preserving the inherent RF fingerprint differences of the device, the two may have different emphases on feature representation during parameter updates. When the domain alignment strength is too high, it may excessively weaken certain feature dimensions related to class distinction, leading to a degradation in the model's discriminative ability in the source domain. If this degradation accumulates during training, it will affect the final recognition performance in the target domain. Therefore, this application implements adversarial training by controlling the update result layer, introducing a controlled adversarial update mechanism based on discriminative performance feedback, which dynamically protects the class discrimination structure while ensuring continuous progress in domain alignment.

[0120] By applying the technical solution of this embodiment, firstly, source domain radio frequency sample data, device tag data corresponding to the source domain radio frequency sample data, and target domain radio frequency sample data are obtained. Then, source domain loss function, adversarial domain loss function, and target domain loss function are designed, and a target function is constructed using the above three loss functions. Then, based on the target function and convolutional neural network, a cross-domain radio frequency fingerprint recognition model is constructed, and a controlled adversarial update mechanism is introduced to determine the target update parameters. Using the source domain radio frequency sample data, device tag data, target domain radio frequency sample data, and target update parameters obtained in the above steps, the cross-domain radio frequency fingerprint recognition model is trained to obtain a pre-trained cross-domain radio frequency fingerprint recognition model. Finally, the radio frequency signal to be identified is sequentially input into the pre-trained feature extractor and classifier to extract features and complete fingerprint recognition, thereby determining the identity of the target transmitting device. This application introduces adversarial domain loss functions and target domain loss functions to force the cross-domain RF fingerprint recognition model to maintain invariant features of the learning domain, thereby reducing distribution differences and improving cross-domain generalization ability. Considering that blind domain alignment will destroy the class discrimination structure and lead to performance degradation, a controlled adversarial update mechanism is introduced to dynamically evaluate and suppress parameter updates that impair discrimination ability. This ensures that the cross-domain RF fingerprint recognition model maintains class discrimination while aligning domain differences, achieving a balance between domain adaptation and discrimination stability. This solves the technical problem of low RF fingerprint recognition accuracy and improves the technical effect of RF fingerprint recognition accuracy.

[0121] Furthermore, as Figure 1 In the specific implementation of the method, in the embodiments of this application, Figure 2 The diagram shows a structural schematic of a cross-domain radio frequency fingerprint recognition model provided in an embodiment of this application, as follows: Figure 2 As shown, the cross-domain RFID fingerprinting model includes at least: a feature extractor, a classifier, a controlled adversarial update module, and a domain discriminator. These network modules work collaboratively to support supervised learning in the source domain and unsupervised adaptive training in the target domain. Figure 3 This illustration shows a schematic diagram of the structure of a feature extractor, classifier, and domain discriminator provided in an embodiment of this application, as shown below. Figure 3 As shown, the feature extractor includes: identity convolution (1D convolution), activation function (ReLU function), batch normalization (BN), max pooling, block, average pooling, and flattening operation. The block further includes: 1D convolution, ReLU function, and batch normalization. The domain discriminator includes: gradient inversion operation, linear layer, and activation function. The classifier includes: linear layer and activation function. The ReLU function can be simply referred to as ReLU. The flattening operation can be simply referred to as flattening.

[0122] In this embodiment, a feature extractor is used to extract latent feature representations from the input RF signal I / Q baseband sequence. The feature extractor can employ a residual neural network structure (ResNet-18) as the backbone network to enhance the model's feature representation capability during the feature extraction process and improve the stability of model training. It should be noted that the specific number of network layers, structural form, and parameter configuration of the feature extractor are not limited to the above examples and can be replaced or adjusted according to actual application requirements.

[0123] Optionally, the classifier is used to identify the transmitting device based on the feature representations corresponding to the source domain radio frequency signal samples. The classifier can employ a classification network structure consisting of multiple fully connected layers, and calculate the category discrimination loss based on the known device identity labels of the source domain samples, thereby guiding the feature extractor to learn radio frequency fingerprint feature representations with device discrimination capabilities.

[0124] Optionally, the domain discriminator is used to determine the data domain to which the feature representation belongs, to distinguish whether the feature comes from the source domain or the target domain, thereby imposing cross-domain alignment constraints on the feature extractor. In this embodiment, an adversarial propagation mechanism can be introduced between the feature extractor and the domain discriminator network, so that the domain discriminator loss has an adversarial effect on the feature extractor parameters during backpropagation, thereby suppressing feature information related to the data domain and promoting the consistency of feature distribution between the source domain and the target domain.

[0125] Optionally, during adversarial training, the model incorporates a controlled adversarial update module to dynamically adjust the parameter update process that includes the domain alignment target, so as to avoid the adversarial training from adversely affecting the model's discrimination ability. At the same time, it incorporates target domain structure constraints to constrain the stability of the prediction results and the class distribution pattern under the condition that the target domain is unlabeled, thereby enhancing the domain adaptability while maintaining the stability of the class discrimination structure.

[0126] In this embodiment, during the training phase of the cross-domain radio frequency fingerprint recognition model, the model parameters are updated using an iterative optimization method. In each training iteration, radio frequency signal samples with device identification tags are sampled from the source domain dataset, and unlabeled radio frequency signal samples are sampled from the target domain dataset. A feature extractor is used to perform feature mapping on the source domain samples and the target domain samples to obtain the corresponding source domain feature representation (…). ) and target domain feature representation ( The source domain dataset can be simply referred to as the source domain. The target domain dataset can be simply referred to as the target domain.

[0127] Furthermore, based on the source domain feature representation, the source domain category discrimination loss is calculated using a classifier. ), i.e., classification loss; based on the source domain feature representation and the target domain feature representation, the adversarial domain alignment loss is calculated through the domain discriminator ( ), namely, domain discrimination loss; at the same time, an enhanced view is constructed for the target domain samples, based on the aforementioned enhanced consistency constraints ( ) and class equilibrium constraints ( ) Calculate the structural constraint loss of the target domain ( The above loss terms are weighted and combined to form the overall training objective for the current training iteration.

[0128] Furthermore, during the parameter update process of the cross-domain RFID model, the parameters of the classifier and domain discriminator are updated via backpropagation according to the overall training objective. For the parameter update of the feature extractor, after obtaining candidate update parameters through conventional gradient calculation, a controlled adversarial update mechanism is introduced from the controlled adversarial update module. This mechanism dynamically evaluates whether the candidate update parameters lead to a degradation in the source domain discrimination performance and performs controlled updates through a hybrid approach of adaptive thresholds and parameters. This achieves a dynamic balance between domain invariant learning and class discrimination ability during the unified optimization process. Through this training method, the model avoids the cumulative degradation of discrimination ability during cross-domain alignment and forms a stable class structure within the target domain, thereby improving the training stability and generalization performance of the cross-domain RFID model.

[0129] In this embodiment, to further verify the cross-domain adaptability of the radio frequency fingerprinting method based on convolutional neural networks under different wireless conditions, various cross-condition target domain scenarios are constructed, including cross-signal-to-noise ratio scenarios and cross-channel scenarios, to simulate the differences in wireless environment between the model training stage and the actual deployment stage.

[0130] Optionally, cross-signal-noise ratio scenario construction: In the cross-signal-noise ratio scenario, using the source domain radio frequency signal sample as a reference, target domain radio frequency signal samples with different signal-noise ratio levels are constructed by superimposing random noise onto it, wherein the random noise follows a complex Gaussian distribution to simulate additive noise interference in the actual wireless environment.

[0131] Optionally, for a given target signal-to-noise ratio (SNR) level, the average signal power of the source domain RF signal samples is first estimated, and the noise power is adaptively adjusted according to a preset SNR value to generate target domain signal samples that meet the target SNR condition. The noise injection process can be represented as follows:

[0132] in, Used to represent source domain radio frequency signal samples Used to represent the constructed target domain radio frequency signal sample This term is used to represent random noise that follows a complex Gaussian distribution, and its power is adjusted according to the target signal-to-noise ratio level.

[0133] Optionally, cross-channel scenario construction: In a cross-channel scenario, parameterized channel transformations are applied to the source domain RF signal samples to simulate potential channel mismatches between the training and deployment phases. Channel transformations include, but are not limited to, Rayleigh fading, Doppler shift, and combinations thereof.

[0134] Furthermore, the cross-channel transformation process can be represented as:

[0135] in, Used to represent the channel fading coefficient Used to represent the Doppler frequency shift parameter, Used to represent the sampling time index and These represent source domain RF signal samples and target domain RF signal samples, respectively.

[0136] In this embodiment, the target domain data constructed through the above-mentioned cross-signal-to-noise ratio and cross-channel methods are consistent with the source domain data in the transmitting device category space, but there are significant differences in signal-to-noise ratio level or channel conditions, thereby forming multiple target domain scenarios for verifying the performance of cross-domain radio frequency fingerprint recognition.

[0137] Furthermore, as Figure 1 In the specific implementation of the method, in the embodiments of this application, Figure 4 This illustration shows a schematic diagram of a radio frequency fingerprint recognition device based on a convolutional neural network, as provided in an embodiment of this application. Figure 4 As shown, the radio frequency fingerprint recognition device 400 based on convolutional neural network includes: a first acquisition unit 401, a construction unit 402, a first determination unit 403, a second acquisition unit 404, a third acquisition unit 405, and a second determination unit 406.

[0138] The first acquisition unit 401 is used to acquire source domain radio frequency sample data of multiple wireless transmitting devices, device tag data corresponding to the source domain radio frequency sample data, and target domain radio frequency sample data.

[0139] The construction unit 402 is used to design the source domain loss function, the adversarial domain loss function, and the target domain loss function respectively, and to construct the target function based on the source domain loss function, the adversarial domain loss function, and the target domain loss function.

[0140] The first determining unit 403 is used to construct a cross-domain radio frequency fingerprint recognition model based on an objective function and a convolutional neural network, and to introduce a controlled adversarial update mechanism to determine the target update parameters of the cross-domain radio frequency fingerprint recognition model.

[0141] The second acquisition unit 404 is used to train the cross-domain radio frequency fingerprint recognition model based on source domain radio frequency sample data, device tag data, target domain radio frequency sample data and target update parameters to obtain a pre-trained cross-domain radio frequency fingerprint recognition model. The pre-trained cross-domain radio frequency fingerprint recognition model includes at least a pre-trained feature extractor and a pre-trained classifier.

[0142] The third acquisition unit 405 is used to acquire the radio frequency signal to be identified and input the radio frequency signal to be identified into a pre-trained feature extractor to obtain the target feature representation.

[0143] The second determining unit 406 is used to input the target feature representation into a pre-trained classifier for fingerprint recognition, obtain the target radio frequency fingerprint, and determine the target transmitting device data based on the target radio frequency fingerprint.

[0144] In this embodiment, a first acquisition unit acquires source domain radio frequency sample data, device tag data corresponding to the source domain radio frequency sample data, and target domain radio frequency sample data from multiple wireless transmitting devices. A construction unit designs source domain loss functions, adversarial domain loss functions, and target domain loss functions, and constructs a target function based on these functions. A first determination unit constructs a cross-domain radio frequency fingerprint recognition model based on the target function and a convolutional neural network, and introduces a controlled adversarial update mechanism to determine the target update parameters of the cross-domain radio frequency fingerprint recognition model. A second acquisition unit then acquires source domain radio frequency sample data, device tag data, and target domain radio frequency sample data. The data and target update parameters are used to train the cross-domain radio frequency fingerprint recognition model, resulting in a pre-trained cross-domain radio frequency fingerprint recognition model. This pre-trained model includes at least a pre-trained feature extractor and a pre-trained classifier. A third acquisition unit acquires the radio frequency signal to be identified and inputs it into the pre-trained feature extractor to obtain a target feature representation. A second determination unit inputs the target feature representation into the pre-trained classifier for fingerprint recognition, obtaining the target radio frequency fingerprint. Based on this fingerprint, the target transmitting device data is determined. This solves the technical problem of low accuracy in radio frequency fingerprint recognition and improves the accuracy of radio frequency fingerprint recognition.

[0145] It should be noted that other corresponding descriptions of the functional units involved in the radio frequency fingerprint recognition device based on convolutional neural networks provided in this application embodiment can be found in the following references. Figure 1 The corresponding descriptions in [the document] will not be repeated here.

[0146] This application also provides a computer device, which may specifically be a personal computer, a server, a network device, etc. Figure 5This application provides a schematic diagram of the device structure of a computer device according to an embodiment of the present application. Figure 5 As shown, the computer device includes a bus, a processor, memory, and a communication interface, and may also include an input / output interface and a display device. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores location information. The network interface allows communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the various method embodiments.

[0147] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0148] In one embodiment, a computer-readable storage medium is provided, which may be non-volatile or volatile, having stored thereon a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0149] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0150] It should be noted that the user personal information involved in the embodiments of this application is all authorized (with the knowledge and consent) by the relevant parties or fully authorized by all parties, and the executing entity can obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with the relevant laws and regulations of the relevant countries and regions, and do not violate public order and good morals. It should be noted that if any software tools or components other than those of this company appear in the embodiments of this application, they are merely illustrative examples and do not represent actual use.

[0151] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, graphics processors, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0152] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0153] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A radio frequency fingerprint recognition method based on a convolutional neural network, characterized in that, The method includes: Acquire source domain radio frequency sample data from multiple wireless transmitting devices, device tag data corresponding to the source domain radio frequency sample data, and target domain radio frequency sample data; Design source domain loss function, adversarial domain loss function and target domain loss function respectively, and construct target function based on source domain loss function, adversarial domain loss function and target domain loss function; Based on the objective function and convolutional neural network, a cross-domain radio frequency fingerprinting model is constructed, and a controlled adversarial update mechanism is introduced to determine the target update parameters of the cross-domain radio frequency fingerprinting model. Based on the source domain radio frequency sample data, the device tag data, the target domain radio frequency sample data, and the target update parameters, the cross-domain radio frequency fingerprint recognition model is trained to obtain a pre-trained cross-domain radio frequency fingerprint recognition model, wherein the pre-trained cross-domain radio frequency fingerprint recognition model includes at least: a pre-trained feature extractor and a pre-trained classifier. The radio frequency signal to be identified is acquired, and the radio frequency signal to be identified is input into the pre-trained feature extractor to obtain the target feature representation; The target feature representation is input into the pre-trained classifier for fingerprint recognition to obtain the target radio frequency fingerprint, and the target transmitting device data is determined based on the target radio frequency fingerprint.

2. The method according to claim 1, characterized in that, The source domain loss function is designed using the following formula, including: in, s Used to represent the source domain; Source domain data used to represent the target batch in the source domain; Used to represent the cross-entropy loss function; Used to represent classifier parameters; Used to represent the classifier; Used to represent feature extractor parameters; Used to represent the feature extractor; Used to indicate in the source domain s The first in i One sample; Used to represent the sample The corresponding device label; Used to represent the source domain loss function.

3. The method according to claim 1, characterized in that, The pre-trained cross-domain RF fingerprinting model further includes at least: a pre-trained domain discriminator, and an adversarial domain loss function designed using the following formula, including: in, Used to represent the target domain; Target domain data used to represent the target batch in the target domain; Used to represent the parameters of the domain discriminator; Samples used to represent the target domain; Used to represent the sample The corresponding field tag; This is used to represent the adversarial domain loss function.

4. The method according to claim 1, characterized in that, Design the target domain loss function, including: The target sample in the target domain radio frequency sample data is enhanced by a first enhancement operator to obtain a first enhanced sample, and the target sample is enhanced by a second enhancement operator to obtain a second enhanced sample, wherein the parameters in the first enhancement operator are different from the parameters in the second enhancement operator; The first enhanced sample and the second enhanced sample are respectively input into the feature extractor to obtain a first feature representation and a second feature representation, and the first feature representation and the second feature representation are respectively input into the classifier to obtain a first prediction probability and a second prediction probability; Based on the first prediction probability and the second prediction probability, a consistency constraint loss function is constructed; and based on the first prediction probability, a class equilibrium loss function is determined. The target domain loss function is determined based on the consistency constraint loss function and the class equilibrium loss function.

5. The method according to claim 4, characterized in that, Based on the first prediction probability and the second prediction probability, a consistency constraint loss function is constructed using the following formula, including: in, Used to indicate in batch The first in j The first predicted probability of each sample; Used to indicate in batch The first in j The second predicted probability for each sample; Used to represent the second norm; This is used to represent the consistency constraint loss function.

6. The method according to claim 4, characterized in that, Based on the first predicted probability, the class equilibrium loss function is determined, including: Based on the first predicted probability, the average distribution probability of the batch in which the first enhanced sample belongs is determined; Based on the average distribution probability, the class-equilibrium loss function is constructed using the divergence function.

7. The method according to claim 1, characterized in that, A controlled adversarial update mechanism is introduced to determine the target update parameters of the cross-domain radio frequency fingerprinting model, including: Obtain the initial update parameters of the cross-domain radio frequency fingerprint recognition model in the first target iteration, and determine the gradient update parameters of the cross-domain radio frequency fingerprint recognition model based on the initial update parameters; Based on the initial update parameters and the gradient update parameters, candidate update parameters for the cross-domain radio frequency fingerprint recognition model are determined. Using the classifier, the initial class loss change is determined based on the initial update parameters and the candidate update parameters; In response to the initial class loss change being greater than zero, the initial class loss change is smoothed to obtain the target class loss change. Based on the change in the target category loss, the target update parameters of the cross-domain radio frequency fingerprinting model are determined in the second target iteration, wherein the second target iteration is the next iteration after the first target iteration.

8. The method according to claim 7, characterized in that, Based on the change in the target category loss, the target update parameters of the cross-domain radio frequency fingerprinting model in the second target iteration are determined, including: Based on the source domain loss function, the moving mean of the cross-domain radio frequency fingerprint recognition model is determined, and based on the source domain loss function and the moving mean, the moving variance of the cross-domain radio frequency fingerprint recognition model is determined. Obtain the threshold parameters of the cross-domain radio frequency fingerprint recognition model, and determine the target threshold of the cross-domain radio frequency fingerprint recognition model based on the threshold parameters and the sliding variance; Based on the target category loss change and the target threshold, the updated receiving coefficients of the cross-domain radio frequency fingerprinting model are determined; The target update parameter is determined based on the update reception coefficient, the initial update parameter, and the candidate update parameter.

9. A radio frequency fingerprint recognition device based on a convolutional neural network, characterized in that, The device includes: The first acquisition unit is used to acquire source domain radio frequency sample data of multiple wireless transmitting devices, device tag data corresponding to the source domain radio frequency sample data, and target domain radio frequency sample data. The construction unit is used to design the source domain loss function, the adversarial domain loss function, and the target domain loss function respectively, and to construct the target function based on the source domain loss function, the adversarial domain loss function, and the target domain loss function; The first determining unit is used to construct a cross-domain radio frequency fingerprint recognition model based on the objective function and the convolutional neural network, and to introduce a controlled adversarial update mechanism to determine the target update parameters of the cross-domain radio frequency fingerprint recognition model. The second acquisition unit is used to train the cross-domain radio frequency fingerprint recognition model based on the source domain radio frequency sample data, the device tag data, the target domain radio frequency sample data and the target update parameters to obtain a pre-trained cross-domain radio frequency fingerprint recognition model, wherein the pre-trained cross-domain radio frequency fingerprint recognition model includes at least: a pre-trained feature extractor and a pre-trained classifier. The third acquisition unit is used to acquire the radio frequency signal to be identified and input the radio frequency signal to be identified into the pre-trained feature extractor to obtain the target feature representation; The second determining unit is used to input the target feature representation into the pre-trained classifier for fingerprint recognition to obtain the target radio frequency fingerprint, and to determine the target transmitting device data based on the target radio frequency fingerprint.

10. A computer device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 8.

Images