A radio frequency fingerprinting method based on data augmentation and prototype learning

Abstract

The application is a radio frequency fingerprint identification method based on data enhancement and prototype learning. The method extracts radio frequency fingerprint features by using a convolutional prototype network, simultaneously learns the prototype of each known class and calculates the center point for subsequent unknown class generation. On the other hand, the method uses GAN to generate feature expansion sample space of unknown classes based on the center point, makes the learned known class features far away from the center area through the center distance-based loss function, and reduces the feature overlap between known classes and unknown classes. Finally, only the feature extractor and the known class prototype are reserved for prototype matching, and open set radio frequency fingerprint identification is realized in a high openness case. The method reduces the feature overlap between known classes and unknown classes, and can realize open set radio frequency fingerprint identification in a high openness case.

Description

Technical Field

[0001] This invention belongs to the field of signal security technology in wireless communication, specifically relating to a radio frequency fingerprinting method based on data augmentation and prototype learning. Background Technology

[0002] Since the beginning of the 21st century, the Internet of Things (IoT) technology has flourished. IoT enables end-to-end device connectivity and is currently involved in various fields such as smart homes, healthcare monitoring, and logistics management. However, with the increasing number of nodes connected to wireless networks, network security issues are becoming increasingly prominent. To address network security threats, physical layer security, based on techniques such as maximizing channel differences, physical layer signal processing, and information theory, has achieved lightweight and robust secure communication, and has been applied to complex communication scenarios such as satellites and the IoT.

[0003] Radio frequency fingerprinting (RFF) is an important authentication method in physical layer security. An RRF fingerprint serves as the physical layer identifier for a wireless device, originating from hardware characteristic differences caused by variations in the manufacturing process of electronic components. This hardware characteristic is uncopyable and tamper-resistant, providing a novel solution for wireless device authentication. RRF fingerprinting technology leverages the uniqueness of RRF fingerprints to prevent unauthorized device access, thereby enhancing the security of communication systems.

[0004] Existing radio frequency fingerprinting (RFF) methods are mainly divided into traditional machine learning (TML)-based methods and deep learning (DL)-based methods. TML-based RFF identification utilizes manually extracted statistical information or physical parameters as device features to identify wireless access devices. This method relies on signal processing knowledge and experience, struggles to capture complex defect features, and lacks adaptability to dynamic environments. In recent years, DL-based methods have been applied to the RFF identification field. Compared to traditional feature engineering methods, DL significantly reduces the complexity of manual feature design. Deep learning replaces subjective feature selection and signal compensation processing, and effectively preserves the features of radio frequency signals. Advanced architectures such as convolutional neural networks (CNN), recurrent neural networks (RNN), and Transformers, as well as improved structures of residual networks (ResNet) and long short-term memory (LSTM), have all demonstrated superior recognition accuracy compared to TML in closed-set RFF identification tasks.

[0005] However, the above methods do not consider the identification of unknown classes in open-set scenarios. When applying closed-set classification models to open-set RFID fingerprinting, the presence of unknown classes often leads to classification results biased towards known classes, which may degrade the performance of RFID fingerprinting. Commonly used classifiers with unknown class detection capabilities, such as OpenMax, struggle to quantify the boundary between unknown and known classes, resulting in a high false detection rate. Existing unknown data generators, including Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs), tend to ignore the spatial distribution characteristics of unknown classes, leading to feature overlap between known and unknown classes and hindering accurate detection of unknown samples. Summary of the Invention

[0006] Technical Problem: To address the aforementioned problems, this invention proposes a radio frequency fingerprinting method based on data augmentation and prototype learning. This method utilizes a Generative Adversarial Network (GAN) to generate unknown class features based on center points, expanding the sample space of the unknown class. Furthermore, by employing a loss function based on center distance, it ensures that the known class features learned by the convolutional prototype network are moved away from the central region, thereby reducing feature overlap between known and unknown classes. This method enables open-set radio frequency fingerprinting under highly open conditions.

[0007] Technical Solution: This invention proposes a radio frequency fingerprinting method based on data augmentation and prototype learning. On one hand, the method uses a convolutional prototype network to extract radio frequency fingerprint features, while simultaneously learning the prototype of each known class and calculating the center point for subsequent unknown class generation. On the other hand, a GAN is used to generate features based on the center point to expand the sample space of the unknown class. A loss function based on center distance is used to make the learned features of the known class move away from the central region, reducing feature overlap between the known and unknown classes. Finally, only the feature extractor and the known class prototype are retained for prototype matching, achieving open-set radio frequency fingerprinting under high openness conditions.

[0008] The method includes the following steps:

[0009] The first step is data preprocessing; converting all the data into time-frequency graphs.

[0010] The second step is to obtain a training set containing known classes and a test dataset that mixes known and unknown classes;

[0011] The third step is to use GANs to generate adversarial examples;

[0012] The fourth step is to use a convolutional prototype network to enhance the discriminative power of known classes and clarify the boundaries of unknown classes through prototype learning and feature space optimization.

[0013] The fifth step is to use a dataset loader to extract training data to train the GAN and the convolutional prototype network;

[0014] The sixth step is to retain the residual network obtained from training and the known class prototypes for open set RF fingerprint recognition.

[0015] in,

[0016] The first step, data preprocessing, includes the following steps:

[0017] Using the acquired co-directional / orthogonal signals as the raw dataset, all data is converted into time-frequency graphs to obtain...

[0018]

[0019] Where x(n) is the acquired in-phase / orthogonal signal, n is the time index, k is the frequency index, m is the frame index, N is the number of Fourier transform points, which is also the length of the window function w(n), and L is the window's step size. The window function is...

[0020]

[0021] The second step involves dividing the known class samples into a training set, a validation set, and a closed test set in a 5:2:2 ratio, while retaining the independent unknown class samples as the open test set, thus obtaining a training set that includes the known classes. and test datasets that mix known and unknown classes Where N x and N x′ These represent the number of data sets in the training and testing sets, respectively.

[0022] The third step involves using GANs to generate adversarial examples to address the shortage of unknown class samples in open-set scenarios. The GAN generator generates unknown class features z based on the input random noise z0. uuc The discriminator input of a GAN consists of a set of prototype center points. and the unknown class sample z generated by the generator uuc Output the true or false result; where o h,l It is the center point of the coordinates of the h nearest prototype points to the l-th prototype point, N. p The total number of prototype points, and the set of center points O, provide the basic features of the unknown class for RFID fingerprint recognition. GAN expands the unknown class samples z based on the center points. uuc This improves the ability of the radio frequency fingerprint recognition model to distinguish unknown classes.

[0023] The fourth step involves using a convolutional prototype network to enhance the discriminative power of known classes and clarify the boundaries of unknown classes through prototype learning and feature space optimization; the input to the convolutional prototype network is X. train The data outputs the extracted features f. θ (x) and the learned prototype M, where x is X train One piece of data, feature f θ (x) is used for subsequent prototype matching with the prototype M to determine the category label of x.

[0024] The fifth step involves using a dataset loader from X. train The process of extracting training data to train GANs and convolutional prototype networks includes the following steps:

[0025] Step 5.1: Initialize the residual network parameters θ in the convolutional prototype network and the generator parameters θ in the GAN using the Kaiming uniform distribution initialization method. G and discriminator parameters θ D Random values ​​are selected, and the prototype point M is initialized to zero.

[0026] Step 5.2: Calculate the coordinates of the center point

[0027]

[0028] Where h = 2, 3, ..., N p l = 1, 2, 3, L, N p N p This is the total number of prototype points; the formula is the sum and average of the coordinate values ​​of h prototype points, p i,lIt is the i-th prototype point closest to the l-th prototype point, and the formula is derived from prototype point p. i,l The o in the feature region of the unknown class is calculated. h,l So that o can be used h,l Generate unknown class samples to occupy open space, providing spatial distribution characteristics of unknown classes for radio frequency fingerprint recognition;

[0029] Step 5.3: To enable GANs to generate diverse data to address the problem of insufficient unknown class samples, N is generated. z A set of random noises that follow a standard normal distribution N z The batch size is determined by the batch size set during GAN training, and the batch size is determined experimentally.

[0030] Step 5.4: Input the set Z into the generator G to obtain the sample set. Then G z and center point o h,l The input to the GAN discriminator D yields true / false judgments; the noise set Z enables the GAN to generate diverse unknown class features G based on the center points. z Discriminator D judges the G generated by the generator. z Whether the data is real or fake provides feedback to the generator, enabling the generator to adjust its parameters based on the discriminator's feedback and generate data that is closer to the center point, thus expanding the unknown class samples and improving the RF fingerprint recognition model's ability to distinguish unknown classes.

[0031] Step 5.5: Update discriminator parameters

[0032]

[0033] Where L is the number of vectors in the set O of center points, this formula represents the center points O i and the data G(z) generated by the generator i The values ​​are input into the discriminator to obtain the true / false judgment D(O). i ) and D(G(z) i By increasing D(O) i ) is true and D(G(z) is true. i The probability of a result being false is used to adjust the discriminator parameter θ. D This improves the discriminator's ability to distinguish between real and fake data, which in turn feeds back to the generator, prompting it to generate diverse unknown class features and improve the radio frequency fingerprint recognition model's ability to distinguish unknown classes.

[0034] Step 5.6: Update generator parameters

[0035]

[0036] Where Nx It is the number of training samples, d i It is the distance between sample x and the known class, d m It is the distance between sample x and the generated unknown class, and this formula will reduce the noise z. i The input discriminator obtains the true / false judgment D(z). i ), by increasing D(z) i The probability of ) being true enhances the generator's data generation capability, and by reducing d m and increase d i To adjust the generator parameter θ G This ensures that the generated samples are close to the unknown class and far from the known class, occupying the feature space of the unknown class and providing information about the unknown class for the radio frequency fingerprint recognition model.

[0037] Step 5.7: Calculate the total loss

[0038] λ=L P +L center (6)

[0039] Where L p It is the mixture loss of the prototype classifier, expressed as the distance-based cross-entropy loss L. DCE With prototype loss L prototype The sum of; L center It is a centroid-based loss that forces known class features away from the open space, reducing the overlap between known and unknown class features, and thus avoiding the identification of unknown classes as known classes. p The definition of

[0040] L p =L DCE +λL prototype (7)

[0041] Where λ is the weight of the prototype loss, ranging from [0,1], and its specific value is determined experimentally. DCE For distance-based cross-entropy loss, L prototype It is the prototype loss, which is defined as follows:

[0042] L prototype (x,y)=||f θ (x)-m y || 2 (8)

[0043] f θ (x) are the features learned by the deep residual network from sample x, m y It is the prototype of category y, L prototype Prototype loss represents the average radius of the prototype cluster; the smaller the radius, the more compact the class features. Reducing L... prototypeThis helps the model learn the boundaries between known and unknown classes, thereby improving the discrimination ability of the RFID fingerprint recognition model. Distance-based cross-entropy loss L DCE Used to improve the similarity between samples and prototypes, and reduce L DCE This enables the model to learn to better represent the prototypes of known classes, facilitating the updating of centroids and the generation of unknown classes during subsequent iterations.

[0044] Distance-based cross-entropy loss L DCE Defined as

[0045] L DCE (x,y)=-logp(y|x)=-logp(f θ (x)∈m y |x) (9)p(f θ (x)∈m y The term |x) is abbreviated as p. y p(x) represents the probability that sample x belongs to class y. y (x) is defined as

[0046]

[0047] C is the number of elements in the known set of category labels, p y (x) is a distance-based loss function that transforms the distance metric space into the negative log-likelihood of the probability space. The distance metric is the square root of the difference between the coordinates in the multidimensional space. The Euclidean distance formula is as follows:

[0048]

[0049] a j With b j They are n-dimensional vectors f θ (x) and m i The value of L in the j-th dimension center The expression for the loss function is:

[0050]

[0051] Where, d y It is the distance d between sample x and the prototype of this class. j d is the distance between sample x and other known class prototypes. m ω is the distance between sample x and the generated unknown class. j and η m

[0052]

[0053] coefficient ω j and η mActing on d j and d m The system assigns a larger repulsive force to samples that are close to each other and a smaller repulsive force to samples that are far apart, causing known samples to tend to be distributed in the edge region, reducing the overlap of features between known and unknown classes, and thus avoiding the identification of unknown classes as known classes.

[0054] Step 5.8: Optimize the parameters θ and prototype point M of the residual network using backpropagation with loss function l.

[0055] The L prototype Prototype loss represents the average radius of the prototype cluster; the smaller the radius, the more compact the class features. Reducing L... prototype This is beneficial for the model to learn the boundaries between known and unknown classes, thereby improving the discrimination ability of the RFID fingerprint recognition model. The distance-based cross-entropy loss L... DCE Used to improve the similarity between samples and prototypes, and reduce L DCE This enables the model to learn to better represent the prototypes of known classes, facilitating the updating of centroids and the generation of unknown classes during subsequent iterations.

[0056] In the sixth step, the residual network and known class prototypes obtained from training are retained for open set RFID fingerprint recognition. Let τ represent the threshold for detecting unknown classes, where 0 < τ < 1. The specific value is determined experimentally. Then X... test The category label r obtained by radio frequency fingerprinting of sample x′ is represented as follows:

[0057] This formula indicates that if the probability value of the best-matching prototype is maxp y If (x′) is less than the threshold τ, it is identified as an unknown class; otherwise, it is identified as the class corresponding to the best matching prototype. C is the number of elements in the known category label set.

[0058] Beneficial Effects: The RF fingerprinting method based on data augmentation and prototype learning proposed in this invention has the following advantages: Compared with GANs and VAEs, this invention generates unknown class features based on center points, and forces the learned known class features to move away from the center region through a loss function based on center distance, thereby reducing the feature overlap between known and unknown classes, and achieving better recognition performance in open set scenarios. Furthermore, the sample features learned in this invention are not limited to nonlinear features, but also include more complex data features learned from the prototype center points. Compared with traditional algorithms, the features are richer, and RF fingerprinting performance can be maintained even when the number of known devices is insufficient. Attached Figure Description

[0059] Figure 1 This is a radio frequency fingerprint recognition system driven by a hybrid of knowledge and data.

[0060] Figure 2 Here is a diagram of the GAN network structure;

[0061] Figure 3 This is a prototype convolutional learner based on ResNet-18. Detailed Implementation

[0062] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0063] The radio frequency fingerprint recognition system based on data augmentation and prototype learning of the present invention is as follows: Figure 1 As shown. The overall process of radio frequency fingerprint recognition is as follows:

[0064] Step 1: Data preprocessing.

[0065] Using LoRa preamble signals collected in an indoor environment as the raw dataset, all data was converted into a time-frequency graph to obtain...

[0066]

[0067] Where x(n) is the LoRa signal in the same direction / orthogonal direction, n is the time index, and k is the frequency index. N is the number of Fourier transform points, which is also the length of the window function w(n), and L is the window's step size. In this embodiment, N and L are both 1024. The window function is...

[0068]

[0069] Step 2: Divide the known class samples into a training set, a validation set, and a closed test set in a 5:2:2 ratio, while retaining the independent unknown class samples as the open test set, thus obtaining a training set containing the known classes. and test datasets that mix known and unknown classes Where N x and N x′ These represent the number of data sets in the training and testing sets, respectively.

[0070] Step 3: Use GAN to generate adversarial examples to address the lack of unknown class samples in open-set scenarios. The GAN network structure is shown in Figure 2. The discriminator consists of an input layer, two Dense layers, and a Softmax layer to output a [0,1] or [1,0] vector representing the discriminator's judgment as false or true. The generator contains two Dense layers in the reverse order of the discriminator's layers, facilitating the derivation of false data from the discriminator's judgment. The GAN generator generates unknown class features z based on the input random noise z0. uuc The discriminator input of a GAN consists of a set of prototype center points. and the unknown class sample z generated by the generator uucOutput the true or false result. Where o h,l It is the center point of the coordinates of the h nearest prototype points to the l-th prototype point, N. p This represents the total number of prototype points. The set of center points O provides the basic features of the unknown class for RFID fingerprint recognition. GAN expands the unknown class samples z based on the center points. uuc This is beneficial for improving the ability of radio frequency fingerprint recognition models to distinguish unknown classes.

[0071] Step 4: Use a convolutional prototype network to enhance the discriminative power of known classes and clarify the boundaries of unknown classes through prototype learning and feature space optimization. The structure of the convolutional prototype network is as follows: Figure 3 As shown, the feature extractor F is the classic ResNet-18 model. The input is X. train The data outputs the extracted features f. θ (x) and the learned prototype M, where x is X train One data point from the dataset. Feature f θ (x) is used for subsequent prototype matching with the prototype M to determine the category label of x.

[0072] Step 5: Use the dataset loader to retrieve data from X train The training data is extracted to train the GAN and the convolutional prototype network. This step includes the following process.

[0073] (1) The Kaiming uniform distribution initialization method is used to initialize the residual network parameters θ in the convolutional prototype network and the generator parameters θ in the GAN. G and discriminator parameters θ D Random values ​​are selected, and the prototype point M is initialized to zero.

[0074] (2) Calculate the coordinates of the center point

[0075]

[0076] Where h = 2, 3, ..., N p l = 1, 2, 3, L, N p N p This is the total number of prototype points. The formula is the sum and average of the coordinate values ​​of h prototype points, p i,l It is the i-th prototype point that is closest to the l-th prototype point. This formula is derived from prototype point p. i,l The o in the feature region of the unknown class is calculated. h,l So that o can be used h,l The generation of unknown class samples occupies open space, providing spatial distribution characteristics of unknown classes for radio frequency fingerprint recognition.

[0077] (3) In order for GAN to generate diverse data to solve the problem of insufficient unknown class samples, N zA set of random noises that follow a standard normal distribution N z The batch size is determined by the batch size set during GAN training; in this embodiment, the batch size is 64.

[0078] (4) Input the set Z into the generator G to obtain the sample set. Then G z and center point o h,l The input to the GAN discriminator D yields true / false judgments. The noise set Z enables the GAN to generate diverse unknown class features G based on the center points. z Discriminator D judges the G generated by the generator. z Whether the data is real or fake provides feedback to the generator, allowing it to adjust its parameters based on the discriminator's feedback and generate data that is closer to the center point, thus expanding the unknown class samples and improving the RF fingerprint recognition model's ability to distinguish unknown classes.

[0079] (5) Update the discriminator parameters

[0080]

[0081] Where L is the number of vectors in the set O of center points. This formula represents the center points O. i and the data G(z) generated by the generator i The values ​​are input into the discriminator to obtain the true / false judgment D(O). i ) and D(G(z) i By increasing D(O) i ) is true and D(G(z) is true. i The probability of a result being false is used to adjust the discriminator parameter θ. D This improves the discriminator's ability to distinguish between real and fake data, which in turn feeds back to the generator, prompting it to generate diverse unknown class features, thereby enhancing the RFID fingerprint recognition model's ability to distinguish unknown classes.

[0082] (6) Update generator parameters

[0083]

[0084] Where N x It is the number of training samples, d i It is the distance between sample x and the known class, d m It is the distance between sample x and the generated unknown class. This formula will reduce the noise z. i The input discriminator obtains the true / false judgment D(z). i ), by increasing D(z) i The probability of ) being true enhances the generator's data generation capability, and by reducing d m and increase d iTo adjust the generator parameter θ G This ensures that the generated samples are close to the unknown class and far from the known class, occupying the feature space of the unknown class and providing information about the unknown class for the radio frequency fingerprint recognition model.

[0085] (7) Calculate the total loss

[0086] λ=L P +L center (6)

[0087] Where L p It is the mixture loss of the prototype classifier, expressed as the distance-based cross-entropy loss L. DCE With prototype loss L prototype The sum of L. center It is a centroid-based loss function that forces known class features away from the open space, reducing the overlap between known and unknown class features, and thus avoiding the identification of unknown classes as known classes. p The definition of

[0088] L p =L DCE +λL prototype (7)

[0089] Where λ is the weight of the prototype loss, ranging from [0,1], and its specific value is determined experimentally. L DCE This is a distance-based cross-entropy loss. L prototype It is the prototype loss, which is defined as follows:

[0090] L prototype (x,y)=||f θ (x)-m y || 2 (8)

[0091] f θ (x) are the features learned by the deep residual network from sample x, m y It is the prototype of category y. L prototype Prototype loss represents the average radius of the prototype cluster; the smaller the radius, the more compact the class features. Reducing L... prototype This helps the model learn the boundaries between known and unknown classes, thereby improving the discrimination ability of the RFID fingerprint recognition model. Distance-based cross-entropy loss L DCE Used to improve the similarity between samples and prototypes, and reduce L DCE This enables the model to learn prototypes that better represent known classes, facilitating the updating of centroids and the generation of unknown classes during subsequent iterations. DCE Defined as

[0092] L DCE (x,y)=-logp(y|x)=-logp(fθ (x)∈m y |x) (9)p(f θ (x)∈m y The term |x) can be simply written as p y p(x) represents the probability that sample x belongs to class y. y (x) is defined as

[0093]

[0094] C is the number of elements in the known category label set; in this embodiment, the number of known category labels is 10. y (x) is a distance-based loss function that transforms the distance metric space into the negative log-likelihood of the probability space. The distance metric is the square root of the difference between the coordinates in the multidimensional space, and the Euclidean distance formula is as follows:

[0095]

[0096] Among them, a j With b j They are n-dimensional vectors f θ (x) and m i The value in the j-th dimension. L center The expression for the loss function is:

[0097]

[0098] Where, d y It is the distance d between sample x and the prototype of this class. j d is the distance between sample x and other known class prototypes. m It is the distance between sample x and the generated unknown class. Define ω. j and η m

[0099] coefficient ω j and η m Acting on d j and d m The system assigns a larger repulsive force to samples that are close to each other and a smaller repulsive force to samples that are far apart, causing known samples to tend to be distributed in the edge region, reducing the overlap of features between known and unknown classes, and thus avoiding the identification of unknown classes as known classes.

[0100] (8) Optimize the parameters θ and prototype point M of the residual network by backpropagation using the loss function l.

[0101] Step 5: Retain the residual network obtained from training and the known class prototypes for open set RFID fingerprint recognition. Let τ represent the threshold for detecting unknown classes, where 0 < τ < 1, and the specific value is determined experimentally. Then Xtest The category label r obtained by radio frequency fingerprinting of sample x′ can be represented as:

[0102] This formula indicates that if the probability value of the best-matching prototype is maxp y If (x′) is less than the threshold τ, it is identified as an unknown class. Otherwise, it is identified as the class corresponding to the best matching prototype.

[0103] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A radio frequency fingerprint recognition method based on data augmentation and prototype learning, characterized in that, This method uses a convolutional prototype network to extract RF fingerprint features, while simultaneously learning the prototype of each known class and calculating the center point for subsequent unknown class generation. On the other hand, it uses a GAN to generate features based on the center point to expand the sample space of the unknown class. By using a loss function based on center distance, the learned features of the known class are made to move away from the central region, reducing the feature overlap between the known and unknown classes. Finally, only the feature extractor and the prototype of the known class are retained for prototype matching, realizing open set RF fingerprint recognition under high openness conditions. The method includes the following steps: The first step is data preprocessing; converting all the data into time-frequency graphs. The second step is to obtain a training set containing known classes and a test dataset that mixes known and unknown classes; The third step is to use GANs to generate adversarial examples; The fourth step is to use a convolutional prototype network to enhance the discriminative power of known classes and clarify the boundaries of unknown classes through prototype learning and feature space optimization. The fifth step is to use a dataset loader to extract training data to train the GAN and the convolutional prototype network; Step 6: Retain the residual network obtained from training and the known class prototypes for open set RFID fingerprint recognition; The first step, data preprocessing, includes the following steps: Using the acquired co-directional / orthogonal signals as the raw dataset, all data is converted into time-frequency graphs to obtain... （1）； in These are the collected signals in the same direction or orthogonal direction. It is a time index. It is a frequency index. It is a frame index. It is the number of points in the Fourier transform, and also the window function. Length, It is the window's step size, and the window function is... （2）； The second step involves dividing the known class samples into a training set, a validation set, and a closed test set in a 5:2:2 ratio, while retaining the independent unknown class samples as the open test set, thus obtaining a training set that includes the known classes. and test datasets that mix known and unknown classes ;in and These are the number of data sets in the training and testing sets, respectively. The third step involves using GANs to generate adversarial examples to address the shortage of unknown class samples in open-set scenarios; the GAN generator generates examples based on the input random noise. Generate unknown class features The discriminator input of a GAN consists of a set of prototype center points. and unknown class samples generated by the generator Output the true or false value; where, It is the distance from the first The nearest prototype point The center point of the coordinates of the prototype points. It represents the total number of prototype points and the set of center points. To provide basic features for unknown classes in radio frequency fingerprint recognition, GANs expand the samples of unknown classes based on center points. This improves the ability of the radio frequency fingerprint recognition model to distinguish unknown classes; The fourth step involves using a convolutional prototype network to enhance the discriminative power of known classes and clarify the boundaries of unknown classes through prototype learning and feature space optimization; the input to the convolutional prototype network is... Data, output extracted features and the prototypes learned ,in yes One piece of data, features For subsequent and prototype use Perform prototype matching to determine Category tags.

2. The radio frequency fingerprint recognition method based on data augmentation and prototype learning according to claim 1, characterized in that, The fifth step involves using a dataset loader from... The process of extracting training data to train GANs and convolutional prototype networks includes the following steps: Step 5.1: Initialize the residual network parameters in the convolutional prototype network using the Kaiming uniform distribution method. GAN generator parameters Discriminator parameters Random value selection, prototype point Initialize to zero; Step 5.2: Calculate the coordinates of the center point （3）； in , , It is the total number of prototype points; this formula applies to Sum and average the coordinates of the prototype points. It is the distance from the first The nearest prototype point The equation is derived from the prototype points. Calculate the feature region of the unknown class So that it can be used Generate unknown class samples to occupy open space, providing spatial distribution characteristics of unknown classes for radio frequency fingerprint recognition; Step 5.3: To enable GANs to generate diverse data to address the problem of insufficient samples of unknown classes, the following steps are taken: A set of random noises that follow a standard normal distribution , The batch size is determined by the batch size set during GAN training, and the batch size is determined experimentally. Step 5.4: Set The sample set is obtained by inputting the data into the generator G. Then and center point Input to the GAN discriminator D to obtain true / false judgments; noise set This enables GANs to generate diverse unknown class features based on center points. Discriminator D judges the generator's output by... Whether the data is real or fake provides feedback to the generator, allowing it to adjust its parameters based on the discriminator's feedback and generate data that is closer to the center point, thus expanding the unknown class samples and improving the RF fingerprint recognition model's ability to distinguish unknown classes. Step 5.5: Update discriminator parameters （4）； in, It is the set of center points The number of vectors in the formula, the center point and the data generated by the generator The inputs are fed into a discriminator to determine whether the result is true or false. and By increasing For truth and Adjust the discriminator parameters to determine the probability of a false result. This improves the discriminator's ability to distinguish between real and fake data, which in turn feeds back to the generator, prompting it to generate diverse unknown class features and improve the radio frequency fingerprint recognition model's ability to distinguish unknown classes. Step 5.6: Update generator parameters （5）； in It is the number of training samples. It is a sample Distance between and known classes It is a sample The distance between the generated unknown class and the noise is expressed by this formula. The input discriminator determines whether the input is true or false. By improving The probability of truth is increased to enhance the generator's data generation capability, and by narrowing down... and increase To adjust generator parameters This ensures that the generated samples are close to the unknown class and far from the known class, occupying the feature space of the unknown class and providing information about the unknown class for the radio frequency fingerprint recognition model. Step 5.7: Calculate the total loss (6) ； in It is the mixture loss of the prototype classifier, expressed as a distance-based cross-entropy loss. With prototype loss sum; It is a centroid-based loss mechanism that forces known class features away from the open space, reducing the overlap between known and unknown class features, and thus avoiding the identification of unknown classes as known classes. The definition of (7) ； in These are the weights of the prototype loss, ranging from [0,1], with specific values ​​determined experimentally. For distance-based cross-entropy loss, It is the prototype loss, which is defined as follows: (8) ； It is the sample learned by the deep residual network. Features It is a category The prototype, The prototype loss represents the average radius of the prototype cluster. Distance-based cross-entropy loss Defined as (9) ； This item is abbreviated as , indicating sample Belongs to class The probability, Defined as (10) ； It is the number of elements in the known category label set. It is a distance-based loss function that transforms the distance metric space into the negative log-likelihood of the probability space. The distance metric is the square root of the difference between the coordinates in the multidimensional space. The Euclidean distance formula is as follows: (11) ； and They are dimensional vector and In the Dimension value, The expression for the loss function is: (12) ； in, It is a sample Distance from the prototype of this class It is a sample Distance between and other known class prototypes It is a sample The distance between the generated unknown class and the unknown class is defined. and (13) ； (14) ； coefficient and Effect on and The system assigns a larger repulsive force to samples that are close to each other and a smaller repulsive force to samples that are far apart, causing known samples to tend to be distributed in the edge region, reducing the overlap of features between known and unknown classes, and thus avoiding the identification of unknown classes as known classes. Step 5.8: Use the loss function Backpropagation optimizes the parameters of the residual network. and prototype point .

3. The radio frequency fingerprint recognition method based on data augmentation and prototype learning according to claim 2, characterized in that, The Prototype loss represents the average radius of the prototype cluster; the smaller the radius, the more compact the class features. This helps the model learn the boundaries between known and unknown classes, thereby improving the discrimination ability of the RFID fingerprint recognition model, based on distance-based cross-entropy loss. Used to improve the similarity between samples and prototypes, and reduce This enables the model to learn to better represent the prototypes of known classes, facilitating the updating of centroids and the generation of unknown classes during subsequent iterations.

4. The radio frequency fingerprint recognition method based on data augmentation and prototype learning according to claim 3, characterized in that, In the sixth step, the residual network obtained from training and the known class prototypes are retained for open-set radio frequency fingerprint recognition. This represents the threshold for detecting unknown classes, and The specific value is determined by experiment. Samples in Category labels obtained through radio frequency fingerprint recognition Represented as (15) ； This formula represents the probability value of the best matching prototype. Less than the threshold If the value is unknown, it is identified as an unknown class; otherwise, it is identified as the class corresponding to the best matching prototype. C is the number of elements in the known category label set.

Images