A radio frequency fingerprinting method, device, medium and product

Abstract

This application discloses a radio frequency fingerprinting method, device, medium, and product, relating to the fields of wireless communication and physical layer security. The method includes: acquiring a registration dataset; using a pre-trained similarity-guided variational information bottleneck neural network to perform feature mapping on the IQ signal samples of each registered device in the registration dataset, obtaining a feature embedding representation for each registered device; associating and storing the feature embedding representation of each registered device in the registration dataset with the device identity identifier of the corresponding registered device, generating a registration template library; using the similarity-guided variational information bottleneck neural network to perform feature mapping on the IQ signal samples of the device to be identified, obtaining a feature embedding representation to be identified; and determining the device identity identifier of the device to be identified based on the feature embedding representation to be identified and the registration template library. This application improves the accuracy and anti-interference capability of radio frequency fingerprinting while reducing storage and computational overhead.

Description

Technical Field

[0001] This application relates to the field of wireless communication and physical layer security, and in particular to a radio frequency fingerprinting method, device, medium and product. Background Technology

[0002] With the rapid development of the Internet of Things (IoT) and wireless communication technologies, a massive number of wireless devices are continuously connecting to the network, making IoT security a critical issue. Traditional authentication methods based on keys or digital certificates are susceptible to risks such as certificate leakage and high computational overhead. Radio Frequency Fingerprinting (RFFI) technology utilizes the unavoidable non-ideal characteristics introduced during the manufacturing process of wireless devices (such as power amplifier nonlinearity, crystal oscillator frequency offset, I / Q imbalance, etc.) as unique identifiers for devices. It has advantages such as being difficult to forge and requiring no additional computational resources, providing a new solution for device authentication. In practical deployments, RFFI systems face the typical open-set identification problem: the number of device categories that can be acquired during the training phase is limited, while the deployment phase needs to support the dynamic access of new legitimate devices, and also be able to distinguish new devices not appearing in the training set (i.e., unknown devices). This scenario is particularly common in practical applications where IoT devices dynamically access the network and network boundaries continuously expand. Therefore, developing an open-set identification method that supports dynamic device registration and does not require model retraining has significant practical application value.

[0003] Existing research has employed numerous methods to improve the performance of RFFI in open-set scenarios. Typical approaches include metric learning frameworks based on triplet loss and feature representation learning methods based on contrastive loss. These methods construct sample pairs or triples, utilize the class label information of the training set to learn discriminative feature representations, and perform open-set classification using distance metrics (such as K-nearest neighbor classifiers). Furthermore, the Variational Information Bottleneck (VIB) method has also been used for feature compression, removing redundancy by minimizing the mutual information between the input and representation, while maximizing the mutual information between the representation and the label to preserve discriminative information. However, the training process of these methods relies heavily on the class label information of the training set, and in open-set scenarios (where the test set includes unseen devices), the generalization ability of the feature representations still has room for improvement.

[0004] Neural networks have been widely applied in radio frequency fingerprinting (RFFI), such as open-set device authentication. The limited range of training devices coupled with the need to identify unknown devices during testing is a key factor affecting the model's generalization ability. However, existing open-set identification methods (such as those based on Triplet Loss or contrastive learning) ignore the similarity relationships between samples, relying more on training set label information, and still have room for improvement in handling unknown devices. Summary of the Invention

[0005] The purpose of this application is to provide a radio frequency fingerprint recognition method, device, medium, and product that extracts robust feature embedding through a similarity-guided variational information bottleneck neural network, thereby improving the accuracy and anti-interference capability of radio frequency fingerprint recognition while reducing storage and computational overhead.

[0006] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a radio frequency fingerprint recognition method, including: Obtain the registration dataset; the registration dataset includes IQ signal samples and device identification identifiers from multiple registered devices; A similarity-guided variational information bottleneck neural network is used to perform feature mapping on the IQ signal samples of each registered device in the registration dataset to obtain the feature embedding representation of each registered device. The similarity-guided variational information bottleneck neural network is pre-trained using a training dataset, which includes training samples. Each training sample includes an IQ signal sample of a known device, a feature embedding representation, and a similarity pseudo-label. The feature embedding representation of each registered device in the registration dataset is associated with and stored with the device identity identifier of the corresponding registered device to generate a registration template library; The similarity-guided variational information bottleneck neural network is used to perform feature mapping on the IQ signal samples of the device to be identified, and the embedded representation of the feature to be identified is obtained. The device identity identifier of the device to be identified is determined based on the embedded representation of the feature to be identified and the registration template library.

[0007] Secondly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described radio frequency fingerprint recognition method.

[0008] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described radio frequency fingerprint recognition method.

[0009] Fourthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described radio frequency fingerprint recognition method.

[0010] According to the specific embodiments provided in this application, this application achieves the following technical effects: By using a similarity-guided variational information bottleneck neural network to perform feature mapping on the IQ signal, it can effectively extract highly discriminative and robust feature embedding representations from the registered device and the device to be identified, significantly improving the accuracy of RF fingerprint recognition. During the training process, the network introduces similarity pseudo-labels, strengthening the similarity between features of similar devices and the difference between features of dissimilar devices, thereby reducing the impact of channel noise, environmental interference, and other factors on recognition performance. Simultaneously, by associating the feature embeddings of the registered device with the identity identifier to construct a registration template library, the device identity can be determined through simple matching during the identification stage, reducing storage overhead and online computational complexity, and supporting fast and efficient device authentication. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a diagram illustrating the application environment of a radio frequency fingerprint recognition method according to an embodiment of this application.

[0013] Figure 2 This is a flowchart illustrating a radio frequency fingerprint recognition method provided in an embodiment of this application.

[0014] Figure 3 This is a schematic diagram of a similarity-guided variational information bottleneck neural network in one embodiment of this application.

[0015] Figure 4 This is a schematic diagram illustrating the generation process of similarity pseudo-tags in one embodiment of this application.

[0016] Figure 5 This is a schematic diagram of the loss function of a similarity-guided variational information bottleneck neural network in one embodiment of this application.

[0017] Figure 6 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] The radio frequency fingerprint recognition method provided in this application embodiment can be applied to, for example... Figure 1 In the application environment shown, terminal 101 communicates with server 102 via a network. A data storage system can store the data that server 102 needs to process. The data storage system can be set up independently, integrated into server 102, or placed in the cloud or on another server. Terminal 101 can send a registration dataset and IQ signal samples of the device to be identified to server 102. Server 102 determines the device identity of the device to be identified based on the received data. Server 102 can then return the device identity of the device to be identified to terminal 101.

[0021] The terminal 101 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle devices. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. The server 102 can be implemented using a standalone server or a server cluster composed of multiple servers, or it can be a cloud server.

[0022] In one exemplary embodiment, such as Figure 2 As shown, a radio frequency fingerprint recognition method is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 Taking server 102 as an example, the explanation includes the following steps 201 to 205.

[0023] Step 201, Obtain the registration dataset The registration dataset includes IQ signal samples and device identification (such as device ID) from multiple registered devices.

[0024] Step 202: A similarity-guided variational information bottleneck neural network is used to perform feature mapping on the IQ signal samples of each registered device in the registered dataset to obtain the feature embedding representation of each registered device. .

[0025] The similarity-guided variational information bottleneck neural network is pre-trained using a training dataset. The training dataset is obtained through training, and each training sample includes an IQ signal sample of a known device, a feature embedding representation, and a similarity pseudo-label.

[0026] like Figure 3 As shown, the similarity-guided variational information bottleneck neural network includes an input layer, a feature extraction backbone, a fully connected layer, and a reparameterized sampling layer connected in sequence. The feature extraction backbone is composed of multiple stacked residual blocks used to extract deep features from the input. In this embodiment, the feature extraction backbone consists of convolutional layers, batch normalization layers, and activation functions.

[0027] In a specific application example, step 202 includes steps 21 to 25.

[0028] Step 21: For the IQ signal sample of any registered device in the registered dataset, generate a two-dimensional channel-independent spectrum based on the IQ signal sample.

[0029] Step 22: Receive the two-dimensional channel-independent spectrum through the input layer.

[0030] Specifically, additive white Gaussian noise is first added to the IQ signal samples with a preset probability to obtain an IQ sequence. Then, the IQ sequence is normalized by root mean square power to obtain a normalized signal. Next, a short-time Fourier transform is used to convert the normalized signal into a time-spectrum graph, which is then center-shifted. Afterward, the amplitude ratio of adjacent elements in the time-spectrum graph is calculated and converted to a logarithmic dB scale to obtain a channel-independent spectrum. Finally, the spectral data located within a predetermined region at the center of the frequency axis in the channel-independent spectrum is used as a two-dimensional channel-independent spectrum.

[0031] Step 23: Extract features from the two-dimensional channel-independent spectrum using the feature extraction backbone to obtain a feature spectrum.

[0032] Step 24: The fully connected layer maps the feature spectrum into a mean vector and a variance vector based on a preset feature embedding dimension. Specifically, the fully connected layer maps the feature spectrum into an N-dimensional mean vector and an N-dimensional variance vector. N is the preset feature embedding dimension, here N=512.

[0033] Step 25: Sample random noise from the standard normal distribution through the reparameterized sampling layer. The latent variables are calculated based on the random noise, the mean vector, and the variance vector to obtain the feature embedding representation of the registered device. Here, I is the identity matrix.

[0034] Specifically, using the formula Calculate the latent variables; where, These are latent variables, i.e., feature embedding representations. It is the mean vector. It is the variance vector. , For characteristic spectra, It is random noise.

[0035] The role of similarity-guided variational information bottleneck neural networks is to compress high-dimensional input signals into low-dimensional probability distribution representations, and to introduce Gaussian noise to make feature embedding have controllable randomness, thereby providing differentiable random sampling paths for subsequent information compression optimization.

[0036] In a specific application example, the training process of the similarity-guided variational information bottleneck neural network includes the following steps 301 to 307.

[0037] Step 301: Obtain the raw IQ signal dataset. The raw IQ signal dataset includes IQ signal samples from various known devices. Specifically, the raw IQ signal dataset is collected using a LoRa device.

[0038] Step 302: For any IQ signal sample in the original IQ signal dataset, generate a sample two-dimensional channel-independent spectrum based on the IQ signal sample.

[0039] Specifically, firstly, additive white Gaussian noise (randomly sampled within a signal-to-noise ratio range of 20dB to 80dB) is added to the I / Q signal samples with a preset probability to improve model robustness. Secondly, the root mean square power of each I / Q sequence is normalized to a unit value, resulting in a normalized signal. Then, a short-time Fourier transform with a window length of 256 and a hop count of 128 is used to convert the normalized signal into a time-spectrum graph, which is then center-shifted. Next, the amplitude ratio of adjacent columns in the time-spectrum graph is calculated and converted to a logarithmic dB scale to obtain a channel-independent spectrum. Finally, only the spectral data in the central 40% interval of the frequency axis is retained to reduce computational redundancy. After the above preprocessing, a uniformly sized two-dimensional channel-independent spectrum graph is obtained, which serves as the input to the subsequent feature extraction network.

[0040] Step 303: Input the sample two-dimensional channel-independent spectrogram into the similarity-guided variational information bottleneck neural network to obtain the feature embedding representation corresponding to the IQ signal sample.

[0041] Step 304, based on the index of the current training batch and The distribution is used to randomly map each feature embedding representation in the current training batch according to a preset probability, so as to obtain the perturbation embedding corresponding to each feature embedding representation in the current training batch.

[0042] like Figure 4 As shown, during the training phase, in order to ensure that the feature embedding representations retain both label discrimination information and sample similarity structure, a random mixup is first performed on the B feature embedding representations of the current training batch with a preset probability p (0.7 is used in the example of this application): (1) From The mixing coefficient λ is obtained by sampling from the distribution, where, (2) Shuffle the indices of the current training batch to obtain the shuffled indices. (3) Generate perturbation embedding ,otherwise The mixup mechanism described above breaks the deterministic mapping between features and pseudo-labels, enabling the model to learn neighborhood structure. This random mapping mechanism ensures the accuracy of subsequently generated similarity pseudo-labels. With original feature embedding representation Maintaining a non-deterministic relationship between them (i.e.) This avoids optimization stagnation caused by premature saturation of mutual information (i.e., prevents...). become (a deterministic function).

[0043] Step 305: Perform DBSCAN clustering on the perturbation embeddings corresponding to each feature embedding representation in the current training batch to obtain the similarity pseudo-labels corresponding to each IQ signal sample in the current training batch. Y s This allows the similarity relationship between samples to be quantified as a supervisory signal.

[0044] Step 306: Generate a training dataset based on each IQ signal sample in the current training batch, the feature embedding representation corresponding to each IQ signal sample, and the similarity pseudo-label.

[0045] Step 307: Based on the training dataset, determine the overall loss function based on conditional mutual information compression loss, label discrimination loss, and similarity separation loss, and optimize the similarity-oriented variational information bottleneck neural network according to the overall loss function until the overall loss function converges to obtain the trained similarity-oriented variational information bottleneck neural network.

[0046] like Figure 5 As shown, the overall loss function is: ;in, For the overall loss, To compensate for the loss of conditional mutual information compression, To determine loss for the label, For similarity separation loss, The weighting coefficients are for the conditional mutual information compression loss. The weighting coefficients for label discrimination loss. These are the weighting coefficients for the similarity separation loss.

[0047] The conditional mutual information compression loss is based on the derived upper bound of conditional mutual information. Construct, which includes a variance regularization term Alignment with mean ,Right now Where X is the set of characteristic spectra, In a given Sample characteristics under the condition The true distribution For approximation The prior distribution is set to a Gaussian distribution. , These are the weighting coefficients for the variance regularization term. The weighting coefficients for the mean alignment term. It is the variance vector. For characteristic spectra, The first in the current training batch i The similarity pseudo-labels corresponding to each IQ signal sample The first in the current training batch j The similarity pseudo-labels corresponding to each IQ signal sample The first in the current training batch i Feature embedding representation corresponding to each IQ signal sample The first in the current training batch j Feature embedding representation corresponding to each IQ signal sample.

[0048] Label discrimination loss is based on The construction loss, in actual deployment, is equivalent to minimizing the standard cross-entropy loss function: ;in, The first in the current training batch i The feature embedding representation of each IQ signal sample represents the true device category label. For a parameterized classifier, it represents the representation based on feature embeddings. Its tags Classify, For the true label random variable of the sample, For the true labels of the samples, To parameterize The classifier represents the representation based on the sample embedding. Its tags Make predictions. For encoder (parameters are) The conditional distribution of ) represents the conditional distribution of a given input. Time concealment The distribution Generate a distribution for the data, representing the input. and tags The true joint distribution, For mutual information, measurement and The degree of dependency between them, maximizing this value can ensure The loss function aims to retain sufficient discriminative information for classification. Its purpose is to ensure that the feature representation retains discriminative information relevant to device identity.

[0049] Similarity separation loss can be described as increasing the similarity by separating outliers. In actual deployment, the project is implemented as follows: ;in, This is the preset interval parameter. The weights of the sample separation terms.

[0050] In summary, the training process of the similarity-guided variational information bottleneck neural network can be summarized as follows: In each training epoch, the computer device randomly divides the training dataset into multiple batches, each batch containing B feature embedding representations (32 in this example). For each batch, the computer device sequentially executes steps 301 to 303: first, the original IQ signal is converted into a channel-independent spectrogram and input into the feature extraction network to obtain the mean, log-variance, and feature embedding representation. Then, the random mixup and clustering operations in steps 304 and 305 are performed to obtain similarity pseudo-labels. Finally, the overall loss is obtained according to step 307. The gradient of the overall loss with respect to all trainable parameters of the feature extraction network and the classification head is calculated using the backpropagation algorithm, and the parameters are updated using an optimizer (such as RMSprop). After each batch, the computer device uses a validation set to evaluate the recognition performance of the current model. If the performance of the current batch is better than the historical best performance, the current model parameters are saved; otherwise, the next batch continues. When the training reaches the maximum preset number of epochs (1000 epochs) or the performance does not improve after a consecutive preset number of epochs (20 epochs), the computer device stops training. After training, the computer device fixes the parameters of the feature extraction network and discards the classification head (which is only used to provide supervision signals during the training phase). At this point, the trained similarity-guided variational information bottleneck neural network has the ability to map the input signal into a highly generalized feature embedding representation, which can be used for subsequent device registration and open set recognition.

[0051] This application proposes a Similarity-Oriented Open-Set Recognition Framework (SOVIB), which combines the advantages of Variational Information Bottleneck (VIB) and similarity constraints to improve open-set generalization performance. During training, similarity information between samples is introduced as supplementary supervision signals, enabling feature representations to not only retain label discriminative information but also characterize the similarity structure between samples. Specifically, this application extends the traditional Markov chain structure of the Variational Information Bottleneck by introducing similarity variables and random perturbations, and designs a conditional prior distribution to adapt to distance-based classifiers. Through these improvements, more generalizable features can be extracted from data, effectively distinguishing known and unknown devices, and, combined with a three-stage framework, enabling dynamic access and identification of IoT devices.

[0052] In summary, the SOVIB model enhances traditional open-set identification methods in three ways: it effectively compresses redundant information in the input, retains discriminative information related to device categories, and actively learns the similarity structure between samples. This enhances its ability to capture the intrinsic correlations between features of different devices, enabling it to more effectively extract more generalizable fingerprint features from the original IQ signal to distinguish between known devices and unknown attack devices.

[0053] Step 203: Associate and store the feature embedding representation of each registered device in the registration dataset with the device identity identifier of the corresponding registered device to generate a registration template library. The registration template library includes multiple templates, and each template includes a registered device, the feature embedding representation of the registered device, and the device identity identifier.

[0054] This registration template library supports dynamically adding new devices or deleting existing devices without retraining the similarity-guided variational information bottleneck neural network. The purpose of this step is to establish a reference set of known device features for subsequent open-set recognition.

[0055] Step 204: The similarity-guided variational information bottleneck neural network is used to identify the device. Feature mapping is performed on the IQ signal samples to obtain the embedded representation of the features to be identified. .

[0056] Step 205: Determine the device identity identifier of the device to be identified based on the embedded representation of the feature to be identified and the registration template library.

[0057] In a specific application example, the K-nearest neighbor classifier is used for device identification: The embedding representation of the features to be identified is calculated. The Euclidean distance between the template and all templates in the registered template library is used to select the closest one. Templates ( The example in this application is a preset positive integer. The process involves statistically analyzing the device identifiers corresponding to these templates and outputting the most frequent identifier as the identification result for the device to be identified. This step utilizes a trained similarity-guided variational information bottleneck neural network and a registered template library to determine the device identity of unknown devices, thereby achieving open-set identification of new devices even with a limited number of trained device categories.

[0058] In this application, the number of registered device categories in the registration dataset is greater than the number of known device categories in the training dataset. For example, the training dataset contains 7 device categories, while the registration dataset may contain 30 device categories, of which 23 are new devices that did not appear in the training. The device to be identified belongs to the same category as the registered devices in the registration dataset, but its IQ signal samples are different.

[0059] The training phase uses a limited number of device categories (e.g., 7 categories), with each category containing sufficient IQ signal samples (e.g., 500 samples per category). The testing phase includes more device categories (e.g., 30 categories), most of which did not appear in the training phase.

[0060] This application uses pseudo-labels generated by clustering to approximate the similarity relationships between samples and introduces the learning of stable similarity information through mixup random transformation in the latent space. Furthermore, the device recognition accuracy and mean precision of the LoRa dataset are calculated based on the K-nearest neighbor distance between the registered device and the device to be identified. Finally, comprehensive experiments under multiple openness conditions verify that the model outperforms other schemes in open set recognition capabilities.

[0061] In summary, this application has at least the following advantages over the prior art: (1) High accuracy and strong generalization ability in open set recognition. The Similarity-Guided Variational Information Bottleneck (SOVIB) method proposed in this application can significantly improve the recognition accuracy in open set scenarios where the number of training device categories is limited (e.g., only 7 categories) while the number of test device categories is large (e.g., 30 categories, most of which are new devices not seen in training). Experimental results show that when the openness is 0.517, the classification accuracy of this application reaches 0.867 and the macro-average F1 reaches 0.866, which is significantly higher than the method based on Triplet Loss (accuracy 0.671) and the method based on Contrastive Loss (accuracy 0.426). This advantage comes from the fact that this application introduces the similarity information between samples as a supplementary supervision signal during the training process, so that the feature representation not only depends on the limited category labels, but can also actively learn the similarity structure between samples, thus having a stronger generalization ability when facing unseen devices (corresponding to similarity pseudo-label generation and similarity separation loss). ).

[0062] (2) The feature representation has a stable ability to distinguish unknown devices, and the performance gap between known and unknown devices is small. Existing metric learning-based methods force similar samples to cluster tightly and dissimilar samples to be strictly separated, which can easily lead to overfitting of the model to the training devices, resulting in a large performance gap between known and unknown devices (high Novelty_gap, such as 0.0689 for CNN-Triplet and 0.0953 for CNN-Contrastive). This application avoids gradient saturation caused by deterministic mapping by introducing random mixup operations and cluster pseudo-labels. At the same time, it uses conditional prior distribution to replace the standard Gaussian prior in traditional VIB, preventing samples from being forcibly compressed to the origin and destroying the relative geometric relationship. Experimental results show that the Novelty_gap of this application is only 0.0272, which is much smaller than the comparison methods, indicating that the recognition performance of this application for known and unknown devices is more balanced (corresponding to conditional prior design, random mixup mechanism, and conditional mutual information compression loss). ).

[0063] (3) Supports dynamic device registration without retraining the model. This application adopts a three-stage process of training-registration-recognition. After the similarity-guided variational information bottleneck neural network is trained, newly added legitimate devices only need to input a small number of their signal samples into the network to obtain feature embeddings and store them in the registration template library to be recognized without retraining the model. This feature gives this application good scalability and ease of deployment in actual IoT device dynamic access scenarios.

[0064] (4) Insensitive to the number of registered samples, exhibiting strong robustness. The recognition performance of this application continuously improves as the number of registered samples increases from 100 to 500, and stabilizes after 500 samples (accuracy remains above 0.85). Even with a limited number of registered samples (e.g., 100), the accuracy of this application still reaches 0.559, significantly outperforming traditional methods under the same conditions. This advantage stems from the fact that similarity-oriented feature learning enables the feature embedding space to have continuous and structured neighborhood relationships, allowing the K-nearest neighbor classifier to construct a relatively stable decision boundary (corresponding to neighborhood structure learning) even with small sample registrations.

[0065] Based on the same inventive concept, this application also provides an radio frequency fingerprint recognition device for implementing the radio frequency fingerprint recognition method described above. The solution provided by this device is similar to the implementation described in the above method; therefore, the specific limitations in one or more embodiments of the radio frequency fingerprint recognition device provided below can be found in the limitations of the radio frequency fingerprint recognition method described above, and will not be repeated here.

[0066] In one exemplary embodiment, a radio frequency fingerprint recognition device is provided, including the following functional modules: The data acquisition module is used to acquire the registration dataset; the registration dataset includes IQ signal samples and device identification identifiers of multiple registered devices.

[0067] The first feature mapping module is used to perform feature mapping on the IQ signal samples of each registered device in the registration dataset using a similarity-guided variational information bottleneck neural network, thereby obtaining the feature embedding representation of each registered device. The similarity-guided variational information bottleneck neural network is pre-trained using a training dataset, which includes training samples. Each training sample includes an IQ signal sample of a known device, a feature embedding representation, and a similarity pseudo-label.

[0068] The template generation module is used to associate and store the feature embedding representation of each registered device in the registration dataset with the device identity identifier of the corresponding registered device, and generate a registration template library; The second feature mapping module is used to perform feature mapping on the IQ signal samples of the device to be identified using the similarity-guided variational information bottleneck neural network to obtain the embedded representation of the features to be identified. The identity recognition module is used to determine the device identity identifier of the device to be identified based on the embedded representation of the feature to be identified and the registration template library.

[0069] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 6As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores registration datasets and IQ signal samples of the devices to be identified. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When executed by the processor, the computer program implements a radio frequency fingerprinting method.

[0070] Those skilled in the art will understand that Figure 6 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, but are not limited to, the following: Figure 6 The diagram shows more or fewer components, or combinations of certain components, or different component arrangements.

[0071] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0072] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0073] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0074] 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 described above. 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).

[0075] 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, central processing units, graphics processing units, digital signal processors, programmable logic devices, etc., and are not limited to these.

[0076] 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.

[0077] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A radio frequency fingerprint recognition method, characterized in that, include: Obtain the registration dataset; the registration dataset includes IQ signal samples and device identification identifiers from multiple registered devices; A similarity-guided variational information bottleneck neural network is used to perform feature mapping on the IQ signal samples of each registered device in the registration dataset to obtain the feature embedding representation of each registered device. The similarity-guided variational information bottleneck neural network is pre-trained using a training dataset, which includes training samples. Each training sample includes an IQ signal sample of a known device, a feature embedding representation, and a similarity pseudo-label. The feature embedding representation of each registered device in the registration dataset is associated with and stored with the device identity identifier of the corresponding registered device to generate a registration template library; The similarity-guided variational information bottleneck neural network is used to perform feature mapping on the IQ signal samples of the device to be identified, and the embedded representation of the feature to be identified is obtained. The device identity identifier of the device to be identified is determined based on the embedded representation of the feature to be identified and the registration template library.

2. The radio frequency fingerprint recognition method according to claim 1, characterized in that, The similarity-guided variational information bottleneck neural network includes an input layer, a feature extraction backbone, a fully connected layer, and a reparameterized sampling layer connected in sequence. A similarity-guided variational information bottleneck neural network is used to perform feature mapping on the IQ signal samples of each registered device in the registered dataset, resulting in a feature embedding representation for each registered device, including: For the IQ signal samples of any registered device in the registered dataset, a two-dimensional channel-independent spectrum is generated based on the IQ signal samples; The two-dimensional channel-independent spectrum is received through the input layer; The feature extraction backbone is used to extract features from the two-dimensional channel-independent spectrum to obtain the feature spectrum; The fully connected layer maps the feature spectrogram into a mean vector and a variance vector based on a preset feature embedding dimension. Random noise is sampled from a standard normal distribution through the reparameterized sampling layer, and latent variables are calculated based on the random noise, the mean vector, and the variance vector to obtain the feature embedding representation of the registered device.

3. The radio frequency fingerprint recognition method according to claim 2, characterized in that, Generating a two-dimensional channel-independent spectrum based on the IQ signal samples includes: Additive white Gaussian noise is added to the IQ signal samples with a preset probability to obtain the IQ sequence; The root mean square power of the IQ sequence is normalized to obtain a normalized signal; The normalized signal is converted into a time-spectrum graph using short-time Fourier transform and then center-shifted. The amplitude ratio of adjacent elements in the time-frequency spectrum is calculated and converted to a logarithmic dB scale to obtain a channel-independent spectrum. The spectral data located within a defined region at the center of the frequency axis in the channel-independent spectrum is used as a two-dimensional channel-independent spectrum.

4. The radio frequency fingerprint recognition method according to claim 2, characterized in that, The latent variables are calculated using the following formula: ； in, As a latent variable, It is the mean vector. It is the variance vector. For characteristic spectra, It is random noise.

5. The radio frequency fingerprint recognition method according to claim 1, characterized in that, The training process of the similarity-guided variational information bottleneck neural network is as follows: Obtain the raw IQ signal dataset; the raw IQ signal dataset includes IQ signal samples from multiple known devices; For any IQ signal sample in the original IQ signal dataset, generate a sample two-dimensional channel-independent spectrum based on the IQ signal sample; The sample two-dimensional channel-independent spectrogram is input into the similarity-guided variational information bottleneck neural network to obtain the feature embedding representation corresponding to the IQ signal sample; Based on the index of the current training batch and The distribution is randomly mapped to each feature embedding representation in the current training batch according to a preset probability, so as to obtain the perturbation embedding corresponding to each feature embedding representation in the current training batch. DBSCAN clustering is performed on the perturbation embeddings corresponding to each feature embedding in the current training batch to obtain the similarity pseudo-labels corresponding to each IQ signal sample in the current training batch. A training dataset is generated based on each IQ signal sample in the current training batch, the feature embedding representation corresponding to each IQ signal sample, and the similarity pseudo-label. Based on the training dataset, an overall loss function is determined using conditional mutual information compression loss, label discrimination loss, and similarity separation loss. The similarity-oriented variational information bottleneck neural network is then optimized based on the overall loss function until the overall loss function converges, resulting in a trained similarity-oriented variational information bottleneck neural network.

6. The radio frequency fingerprint recognition method according to claim 5, characterized in that, The overall loss function is: ； ； ； ； in, For the overall loss, To compensate for the loss of conditional mutual information compression, To determine loss for the label, For similarity separation loss, The weighting coefficients for the conditional mutual information compression loss are... The weighting coefficients for label discrimination loss. These are the weighting coefficients for the similarity separation loss. It is the variance vector. For characteristic spectra, The first in the current training batch i The similarity pseudo-labels corresponding to each IQ signal sample The first in the current training batch j The similarity pseudo-labels corresponding to each IQ signal sample The first in the current training batch i Feature embedding representation corresponding to each IQ signal sample The first in the current training batch j Feature embedding representation corresponding to each IQ signal sample The first in the current training batch i The feature embedding representation of each IQ signal sample represents the true device category label. This is the preset interval parameter. For a parameterized classifier, it represents the representation based on feature embeddings. Its tags Make predictions. The weights of the variance regularization term. The weights of the mean-aligned terms, The weights of the sample separation terms.

7. The radio frequency fingerprint recognition method according to claim 1, characterized in that, The number of categories of registered devices in the registration dataset is greater than the number of categories of known devices in the training dataset; the device to be identified belongs to the same category as the registered devices in the registration dataset, but the IQ signal samples are different.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the radio frequency fingerprint recognition method according to any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the radio frequency fingerprint recognition method according to any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the radio frequency fingerprint recognition method according to any one of claims 1-7.

Images