A 5G radio frequency fingerprint extraction and identification method for multi-user access scenarios

Abstract

The present application relates to the technical field of radio frequency fingerprint extraction and identification, and particularly relates to a 5G radio frequency fingerprint extraction and identification method for a multi-user access scene, comprising the following steps: S1: using multiple 5G devices to simulate a multi-user access scene to send uplink data frames and collect over-the-air signals; S2: after the over-the-air signals are sequentially preprocessed by signal blind synchronization, resource grid resolution and data demodulation, resource blocks and corresponding resource block features are obtained; S3: using the over-the-air signals and the resource block features, a device-channel feature extractor for intra-frame device resource block clustering is trained by using hyperspherical metric learning; S4: using the over-the-air signals and the resource block features, a single-resource block device classifier is trained; and S5: based on the trained device-channel feature extractor and the single-resource block device classifier, a resource block cluster level device identification system in a multi-user scene is constructed, and a device identification result corresponding to the over-the-air signals is output by a voting mechanism.

Description

Technical Field

[0001] This invention relates to the field of radio frequency fingerprint extraction and recognition technology, specifically to a 5G radio frequency fingerprint extraction and recognition method for multi-user access scenarios. Background Technology

[0002] Radio frequency (RF) fingerprinting is a technology that identifies devices based on their hardware characteristics. During the transmission of radio frequency signals, the slight differences in the manufacturing processes of hardware components such as power amplifiers, filters, and oscillators create unique signal characteristics known as "RF fingerprints." As a latent feature, RF fingerprints are difficult to forge and possess uniqueness, thus offering broad application prospects in areas such as IoT security, user authentication, device tracking, and detection of unauthorized devices.

[0003] In 5G communication systems, the application of radio frequency fingerprinting technology faces challenges due to its high frequency, high speed, and multi-user access characteristics. Existing research focuses on single-user access scenarios, primarily addressing the communication process between a single user and the base station. This approach eliminates the need to consider resource block allocation, allowing for RF fingerprint extraction and identification simply by analyzing the transmitted signals of a single user. However, in practical applications, multi-user access is the norm in 5G networks. In multi-user scenarios, the highly dynamic resource block allocation and complex, time-varying channel conditions in 5G systems make it difficult for existing RF fingerprint extraction methods, primarily targeting single-user or static scenarios, to extract stable hardware features. This hinders effective identification of different devices across resource block allocations and various channel environments. Summary of the Invention

[0004] To address the technical problem that existing radio frequency fingerprint extraction methods struggle to achieve effective cross-resource block identification, the present invention aims to provide a 5G radio frequency fingerprint extraction and identification method for multi-user access scenarios. The specific technical solution adopted is as follows:

[0005] Step S1: Use multiple 5G devices to simulate a multi-user access scenario to send uplink data frames and collect air signals;

[0006] Step S2: After performing preprocessing on the air signals, including blind signal synchronization, de-resource lattice processing, and data demodulation, resource blocks and their corresponding characteristics are obtained.

[0007] Step S3: Using airborne signal and resource block features, train a device-channel feature extractor for intra-frame device resource block clustering using hyperspherical metric learning;

[0008] Step S4: Train a single resource block device classifier using air signals and resource block features;

[0009] Step S5: Based on the trained device-channel feature extractor and single resource block device classifier, construct a resource block cluster-level device identification system for multi-user scenarios, and output the device identification results corresponding to the air signals through a voting mechanism.

[0010] Preferably, step S2 includes:

[0011] Step S21: By utilizing the correlation between the cyclic prefix and the symbol tail, blind synchronization of the air signal is completed, and multiple subframes are captured;

[0012] Step S22: Restore each subframe to a frequency domain resource grid, and decompose it into multiple resource blocks;

[0013] Step S23: Demodulate and extract features from the resource blocks to obtain the corresponding resource block features.

[0014] Preferably, step S21 includes:

[0015] Step S211: Determine the symbol length and cyclic prefix length based on the air signal to obtain the correlation function;

[0016] Step S212: Traverse each subcarrier interval, filter the correlation value corresponding to each possible starting point, and obtain the starting point corresponding to the maximum correlation value;

[0017] Step S213: Combine the maximum correlation values ​​corresponding to all subcarrier intervals, select the subcarrier interval corresponding to the maximum value as the actual subcarrier interval, determine the starting point of the air signal, obtain the cyclic prefix and symbol length, and extract multiple subframes.

[0018] Preferably, step S22 specifically includes:

[0019] Based on the actual subcarrier spacing and the starting point of the air signal, the main body of each symbol is extracted, the data corresponding to the symbol length of the air signal is obtained, the received symbol is restored to a frequency resource grid, and decomposed into multiple resource blocks.

[0020] Preferably, step S23 includes:

[0021] Step S231: Demodulate each resource block to obtain the ideal signal, and calculate the channel state information of the ideal signal;

[0022] Step S232: Average the different symbols at the same resource block position within a single frame to obtain the mean value of the subcarrier channel state information;

[0023] Step S233: Divide the mean of the subcarrier channel state information and extract the envelope features of the CSI vector as resource block features.

[0024] Preferably, step S3 includes:

[0025] Step S31: Generate device-channel labels for each resource block, construct a source label set, and divide it into a training set and a validation set respectively;

[0026] Step S32: Construct a device-channel classifier containing a feature extraction network. Input the training set, map the training set to the feature space through the feature extraction network to obtain features, perform hyperspherical projection on the features to obtain optimized features; use a hyperspherical projection classifier to obtain the classification probability of the optimized features, and define the loss function of the resource block features with respect to the feature extractor.

[0027] Step S33: Repeat step S32 to determine the loss function of the feature extractor;

[0028] Step S34: Use the gradient descent algorithm to select and update the feature extractor and the hyperspherical projection classifier to complete the training of the device-channel feature extractor.

[0029] Preferably, step S4 includes:

[0030] Step S41: Generate resource block device tags for each resource block, construct a user tag set, and divide it into training set and validation set respectively;

[0031] Step S42: Input the training set into the single resource block device classifier to obtain the probability of each category, and define the loss function of the single resource block device classifier obtained from the resource block features;

[0032] Step S43: Repeat step S42 to determine the loss function of the single resource block device classifier;

[0033] Step S44: Update the learning parameters of the single resource block device classifier using the gradient descent algorithm to complete the training of the single resource block device classifier.

[0034] Preferably, step S5 includes:

[0035] Step S51: Use the trained device-channel feature extractor to extract device-channel features from intra-frame resource blocks of air signals, and perform clustering to establish a resource block cluster set;

[0036] Step S52: Use the trained single resource block device classifier to classify each device-channel feature in the resource block cluster set to obtain the device label of the resource block feature. Based on the device label, use a voting mechanism to determine the device identification result corresponding to the resource block cluster set.

[0037] Preferably, step S51 specifically includes:

[0038] Based on the device-channel characteristics corresponding to each pair of resource blocks, the cosine distance is determined as the feature distance. A preset partitioning threshold is used to compare the feature distance and the partitioning threshold, and the resource blocks are divided into multiple clusters. The resource blocks in each cluster are then integrated to establish a resource block cluster set.

[0039] To address the aforementioned problems, the present invention also provides an electronic device comprising a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus, and the processor invokes logical instructions in the memory to execute the 5G radio frequency fingerprint extraction and identification method for multi-user access scenarios described in any of the preceding claims.

[0040] The present invention has the following beneficial effects:

[0041] 1. For the proposed 5G radio frequency fingerprint extraction and recognition method for multi-user access scenarios, the feature distribution in the feature space is optimized by hyperspherical metric learning, which makes the resource block features from the same user cluster in the high-dimensional space and the features of different users are separated, thereby improving the separability between users and improving the accuracy of user identification in multi-user environments; and the device-channel feature extractor and single resource block device classifier are combined to effectively improve the device radio frequency fingerprint recognition capability in multi-user scenarios through resource block identification and voting mechanisms.

[0042] 2. The electronic device provided by this invention has the same beneficial effects as the 5G radio frequency fingerprint extraction and recognition method for multi-user access scenarios provided by this invention, and will not be described in detail here. Attached Figure Description

[0043] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 The implementation process of a 5G radio frequency fingerprint extraction and identification method for multi-user access scenarios provided in one embodiment of the present invention. Figure 1 ;

[0045] Figure 2 The implementation process of a 5G radio frequency fingerprint extraction and identification method for multi-user access scenarios provided in one embodiment of the present invention. Figure 2 ;

[0046] Figure 3This is a flowchart illustrating the steps of a 5G radio frequency fingerprint extraction and identification method for multi-user access scenarios, provided in one embodiment of the present invention. Detailed Implementation

[0047] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a 5G radio frequency fingerprint extraction and identification method for multi-user access scenarios proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0048] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0049] The following description, in conjunction with the accompanying drawings, details a specific scheme for a 5G radio frequency fingerprint extraction and recognition method for multi-user access scenarios provided by the present invention.

[0050] To better explain, 5G radio frequency fingerprint refers to an identifier formed by analyzing the characteristics of radio frequency signals such as signal strength, phase, frequency offset, and multipath fading characteristics generated by terminal devices during wireless communication. It has a high degree of uniqueness and stability, similar to human fingerprints, hence the name radio frequency fingerprint.

[0051] In 5G networks, especially in scenarios with multiple concurrent users such as dense urban areas and large venues, base stations need to serve a large number of terminal devices simultaneously. In multi-user scenarios, the dynamic nature of resource block allocation and the complex and ever-changing channel conditions make it difficult for existing radio frequency fingerprinting methods to achieve effective identification across resource blocks. For example, in dense urban areas, a single base station may need to provide services to hundreds or even thousands of mobile terminals simultaneously. As these terminals move rapidly, the channel state between them and the base station will change drastically due to factors such as obstruction and reflection, resulting in significant differences in the signal characteristics carried by resource blocks allocated at different times and locations. Existing radio frequency fingerprinting methods are mostly based on static or semi-static signal characteristics, which have poor stability in environments with dynamic resource block allocation and rapid channel changes, making it difficult to achieve unified identification and association across resource blocks.

[0052] Please combine Figures 1-3 It illustrates an implementation flowchart and step flowchart of a 5G radio frequency fingerprint extraction and identification method for multi-user access scenarios provided in the first embodiment of the present invention, the method comprising:

[0053] Step S1: Use multiple 5G devices to simulate a multi-user access scenario to send uplink data frames and collect air signals;

[0054] Step S2: After performing preprocessing on the air signals, including blind signal synchronization, de-resource lattice processing, and data demodulation, resource blocks and their corresponding characteristics are obtained.

[0055] Step S3: Using airborne signal and resource block features, train a device-channel feature extractor for intra-frame device resource block clustering using hyperspherical metric learning;

[0056] Step S4: Train a single resource block device classifier using air signals and resource block features;

[0057] Step S5: Based on the trained device-channel feature extractor and single resource block device classifier, construct a resource block cluster-level device identification system for multi-user scenarios, and output the device identification results corresponding to the air signals through a voting mechanism.

[0058] To clarify, multi-user access scenarios refer to situations where multiple mobile devices simultaneously connect and transmit data within the same wireless communication system, such as cellular mobile communication networks, Wi-Fi networks, or IoT systems. In this scenario, each user device communicates with the base station or access point via wireless channels such as radio frequency signals. Radio frequency fingerprinting utilizes the unique radio frequency characteristics of these devices during communication, such as amplitude, phase, frequency offset, delay spread, and multipath effects, to uniquely identify and authenticate the devices.

[0059] As explained, step S1 specifically includes:

[0060] Multiple 5G devices are used to simulate a multi-user access scenario to send uplink data frames, and USRP devices are used to collect air signals.

[0061] As an alternative implementation, 5G devices include smartphones, tablets, IoT sensors, and the like.

[0062] Specifically, in the deployment scenario, multiple 5G devices are configured, each simulating an independent user and sending uplink data frames containing IP packets of different sizes to simulate diverse traffic such as voice, video, and web browsing in actual communication scenarios. Then, a USRP (Universal Software Radio Peripheral) device is used to capture the 5G uplink signal propagating in the air in real time, i.e., to collect the air signal.

[0063] It can be explained that in 5G NR (New Radio) signals, a resource block (RB) is the smallest unit of frequency domain resource allocation, consisting of 12 consecutive subcarriers. Therefore, during a single data collection, multiple 5G devices transmit data on different resource blocks. The device and channel scenario to which each resource block belongs within the collected data frame are known. Multiple resource block allocation ranges and signals in multiple scenarios are collected for each device. That is, one device can allocate multiple resource blocks, and one resource block can contain multiple allocation ranges, which are used for the construction of subsequent datasets, i.e., as prior data.

[0064] Further, step S2 includes:

[0065] Step S21: By utilizing the correlation between the cyclic prefix and the symbol tail, blind synchronization of the air signal is completed, and multiple subframes are captured.

[0066] To clarify, blind synchronization refers to the process of achieving time and frequency synchronization by analyzing the received signal without requiring the sending end to send additional synchronization information.

[0067] Further, step S21 includes:

[0068] Step S211: Determine the symbol length and cyclic prefix length based on the air signal to obtain the correlation function.

[0069] It can be explained that in OFDM (Orthogonal Frequency-Division Multiplexing) systems, the cyclic prefix (CP) is a copy of the useful symbol tail. In other words, if the received air signal is perfect, then two parts of the signal separated by one useful symbol length are highly correlated, that is, the prefix part and the data part of the symbol tail have the same waveform.

[0070] Specifically, the received air signals are defined as Its symbol length is The length of the cyclic prefix is The correlation function is calculated using the following formula:

[0071]

[0072] in, Indicates possible starting points The correlation function; Indicates a signal point in the air; This indicates the complex conjugate operation.

[0073] Step S212: Traverse each subcarrier interval, filter the correlation value corresponding to each possible starting point, and obtain the starting point corresponding to the maximum correlation value.

[0074] Specifically, based on the possible starting points in step S211 Similarly, to obtain the corresponding correlation function, we iterate through each subcarrier interval, i.e., the interval between the center frequencies of two adjacent subcarriers, and calculate the correlation function for possible starting points. In this embodiment, the starting point refers to the initial position of the symbol on the time axis, i.e., the first sampled signal point after the cyclic prefix of the symbol in the air signal ends. We then filter the starting points corresponding to the correlation functions and select the starting point corresponding to the maximum correlation value to determine the maximum correlation value for each subcarrier interval. The corresponding calculation formula is as follows:

[0075]

[0076]

[0077] in, Indicates the starting point corresponding to the maximum correlation value; Indicates possible starting points The correlation function; This represents the maximum correlation value.

[0078] Step S213: Combine the maximum correlation values ​​corresponding to all subcarrier intervals, select the subcarrier interval corresponding to the maximum value as the actual subcarrier interval, determine the starting point of the air signal, obtain the cyclic prefix and symbol length, and extract multiple subframes.

[0079] Specifically, the maximum correlation value of all subcarrier intervals is obtained according to step S212. The subcarrier spacing corresponding to the maximum value is selected as the actual subcarrier spacing, denoted as . And the corresponding starting point is taken as the signal starting point, denoted as . The cyclic prefix and symbol length are calculated, and multiple subframes are extracted. Then, after blind synchronization is completed, the cyclic prefix of the received air signals is removed, and the symbol content is restored.

[0080] Step S22: Restore each subframe to a frequency domain resource grid, and decompose it into multiple resource blocks.

[0081] To clarify, a demodulated resource grid refers to the set of original data or information symbols represented in the frequency domain resource grid after the receiver performs a series of demodulation, decoding, and resource mapping operations on the radio frequency signal after synchronization processing in a wireless communication system. To better illustrate, the Fourier Transform (FFT) converts a signal from the time domain to the frequency domain by decomposing a complex periodic signal into a superposition of sinusoidal components of different frequencies.

[0082] Furthermore, in step S22, specifically:

[0083] Based on the actual subcarrier spacing and the starting point of the air signal, the main body of each symbol is extracted, the data corresponding to the symbol length of the air signal is obtained, the received symbol is restored to a frequency resource grid, and decomposed into multiple resource blocks.

[0084] Specifically, based on the actual subcarrier spacing determined in the relevant steps of step S21 and signal start point The subframes are processed, that is, the main body of each symbol is extracted from... To begin, obtain the symbol length as: The data is used to restore the received symbols into frequency resource grids, which are then decomposed into multiple resource blocks. The corresponding calculation formula is as follows:

[0085]

[0086] in, Indicates the first The symbol of the first Data on each subcarrier; This represents the received symbol after removing the cyclic prefix, i.e., the recovered signal corresponding to the subframe; This represents the Fourier transform.

[0087] Step S23: Demodulate and extract features from the resource blocks to obtain the corresponding resource block features.

[0088] To clarify, Channel State Information (CSI) refers to a set of parameters in a wireless communication system that describes the channel transmission characteristics between the transmitter and receiver. It reflects the specific effects of fading, multipath effects, and interference on the wireless signal during propagation and can be used to measure channel quality.

[0089] Step S231: Demodulate each resource block to obtain the ideal signal, and calculate the channel state information of the ideal signal.

[0090] Specifically, the calculation formula corresponding to step S231 is:

[0091]

[0092] in, Indicates the first The symbol of the first CSI of each subcarrier; Indicates the first The symbol of the first Data on each subcarrier; Represents resource block The corresponding ideal signal.

[0093] Step S232: Average the different symbols at the same resource block position within a single frame to obtain the mean value of the subcarrier channel state information.

[0094] Specifically, the calculation formula corresponding to step S232 is:

[0095]

[0096] in, Indicates the first The average CSI vector of each subcarrier; Indicates the first One symbol; Indicates the number of symbols.

[0097] Step S233: Divide the mean of the subcarrier channel state information and extract the envelope features of the CSI vector as resource block features.

[0098] Specifically, the average CSI vector Each group of 12 points is divided into a resource block, denoted as: Perform an FFT on the resource block and extract the envelope features of the average CSI vector corresponding to the resource block, which will be used as the final resource block features.

[0099] Furthermore, step S3 includes:

[0100] Step S31: Generate device-channel labels for each resource block, construct a source label set, and divide it into training set and validation set respectively.

[0101] Understandably, a dataset is built based on the resource block features corresponding to the air signals. Hyperspherical metric learning is used to further optimize the resource block features and train a feature extractor. By projecting the feature vectors onto the hypersphere and classifying them, the separability between features is enhanced, so that the features of the same user are clustered as much as possible in the feature space, while the features of different users are pushed apart.

[0102] Specifically, based on resource blocks Corresponding user identity Constructing open set training samples, i.e. , , Indicate user identity The number of elements; construct an open set training set, denoted as . , Indicate the number of samples in the training set; construct the open set validation set, i.e. , This represents the number of samples in the validation set; where and The user identity tags do not overlap, that is... .

[0103] Step S32: Construct a device-channel classifier containing a feature extraction network. Input the training set, map the training set to the feature space through the feature extraction network to obtain features, perform hyperspherical projection on the features to obtain optimized features; use the hyperspherical projection classifier to obtain the classification probability of the optimized features, and define the loss function of the resource block features with respect to the feature extractor.

[0104] The device-channel classifier comprises several fully connected layers, batch normalization layers, and activation functions. The fully connected layers linearly combine the output features of the previous layer to map the high-dimensional feature space to the category space. The batch normalization layers normalize the input of each layer to alleviate the vanishing or exploding gradient problem and accelerate model convergence. The activation function introduces nonlinear transformation capabilities into the user classifier, enabling it to fit complex user behavior patterns.

[0105] Specifically, the device-channel classifier extracts device-channel features from the training set through a feature extraction network, that is, maps the training set to the feature space to obtain the features, i.e. ,in , Indicates the first The device-channel features corresponding to each training sample; Represents a feature extraction network; for features Perform hyperspherical projection to obtain optimized features, i.e. , Indicates optimization features; Represents the hypersphere radius. In this embodiment, the hypersphere radius is used as a hyperparameter. Determine the learning parameters of the device-channel classifier, i.e. , Output the current analysis of the first... The classification probability corresponding to each resource block is calculated using the following formula:

[0106]

[0107] in, Indicates the first The classification probability corresponding to each resource block; Indicated by An exponential function with base 0; , They represent the first The resource block and the first Transpose of the learning parameters of the device-channel classifier corresponding to each resource block; This indicates an optimized feature.

[0108] With the first Each resource block is expanded and its corresponding resource block characteristics are defined. The obtained information about the feature extractor The loss function, i.e. , Indicates the first The loss function corresponding to each resource block; minimizing the loss function is equivalent to minimizing the learning parameters of the device-channel classifier. and optimization features The cosine distance between them minimizes all the optimization features corresponding to the same user. The cosine distance between them.

[0109] Step S33: Repeat step S32 to determine the loss function of the feature extractor.

[0110] Specifically, step S32 is repeated, that is, the hyperspherical metric learning process is repeated. In this embodiment, the process is repeated. Next, the loss function of the feature extractor is determined, and the corresponding calculation formula is:

[0111]

[0112] in, This represents the loss function of the feature extractor; This indicates the number of training iterations.

[0113] Step S34: Use the gradient descent algorithm to select and update the feature extractor and the hyperspherical projection classifier to complete the training of the device-channel feature extractor.

[0114] Specifically, through the loss function of the feature extractor The gradient descent algorithm is used to select and update the feature extractor and device-channel classifier, that is, to update their learned parameters using gradient descent. The corresponding calculation formula is as follows:

[0115]

[0116]

[0117] in, Indicates a feature extractor; Indicates the learning rate. ; Represents the loss function About feature extractors The gradient of the learning parameters; Represents the user classifier; Represents the loss function About the device - channel classifier The gradient of the learning parameters.

[0118] It can be noted that, in this embodiment, the learning rate Set to 1e-3, batch size The gradient descent uses the Adam optimizer; the feature extractor model structure is as follows, and the input data dimension is... This indicates that each sample contains 26 features. The structure includes three fully connected layers, each followed by batch normalization and activation functions to improve training stability and accelerate convergence. The first layer, Linear(26, 128), maps the input data (training set) to 128 dimensions. The second layer, Linear(128, 64), maps the 128-dimensional data to 64 dimensions. The third layer, Linear(64, zdim), finally maps the 64-dimensional data to the embedding space, with dimensions of zdim. After passing through the feature extractor model, the resource block features in the input training set are transformed into... Dimensional embedding vector.

[0119] It should be noted that in step S3, the resource block features are the original radio frequency features extracted based on CSI; the device-channel features are the high-dimensional embedded features obtained by mapping the resource block features through the device-channel feature extractor.

[0120] Further, step S4 includes:

[0121] Step S41: Generate resource block device tags for each resource block, construct a user tag set, and divide it into training set and validation set respectively.

[0122] Specifically, based on resource blocks Corresponding resource block device tag Construct closed-set classification training samples, i.e. , , Indicates resource block device label The number of training sets; constructing the training set, i.e. , Indicate the number of samples in the training set; construct the validation set, i.e. , This represents the number of samples in the validation set. Wherein, and All contain the same Equipment label.

[0123] Step S42: Input the training set into the single resource block device classifier to obtain the probability of each category, and define the loss function of the single resource block device classifier obtained from the resource block features.

[0124] Specifically, the single-resource-block device classifier still contains several fully connected layers, batch normalization layers, and activation functions. The training set is input into the single-resource-block device classifier, and the relevant learning parameters of the single-resource-block device classifier are determined. and The classification probability of the resource block is obtained, and the corresponding calculation formula is:

[0125]

[0126] in, Indicates the first The classification probability of each resource block; Indicated by An exponential function with base 0; , They represent the first The resource block and the first Transpose of the learning parameters corresponding to each resource block; Indicates training set Data samples in; , They represent the first The resource block and the first The learning parameters corresponding to each resource block.

[0127] Then, define the first Resource block characteristics of a resource block The resulting loss function for the device classifier, i.e. .

[0128] Step S43: Repeat step S42 to determine the loss function of the single resource block device classifier.

[0129] Specifically, repeat step S42 until... Next, the loss function of the single resource block device classifier is determined, and the corresponding calculation formula is:

[0130]

[0131] in, Single resource block device classifier The loss function.

[0132] Step S44: Update the learning parameters of the single resource block device classifier using the gradient descent algorithm to complete the training of the single resource block device classifier.

[0133] Specifically, the corresponding calculation formula is:

[0134]

[0135]

[0136] in, , All of these represent the learning parameters of the device classifier; Indicates the learning rate. ; , Represent the loss function respectively Regarding single-resource block device classifiers Learning parameters and The gradient.

[0137] It can be explained that the learning parameters and Gradient descent updates are performed; in this embodiment, the learning rate is... Set to 1e-3, batch size The Adam optimizer is used for gradient descent.

[0138] To better illustrate, during the execution of step S5, the device-channel feature extractor is used to learn a general feature representation across devices, and the single resource block device classifier is used to make a distinction within a preset set of devices.

[0139] Furthermore, step S5 includes:

[0140] Step S51: Use the trained device-channel feature extractor to extract device-channel features from intra-frame resource blocks of air signals, and perform clustering to establish a resource block cluster set.

[0141] It is explained that, based on the unknown signals in the air signals, the corresponding resource blocks are obtained according to step S2, and the trained device-channel feature extractor is used to extract device-channel features for subsequent clustering processing.

[0142] Furthermore, in step S51, specifically:

[0143] Based on the device-channel characteristics corresponding to each pair of resource blocks, the cosine distance is determined as the feature distance. A preset partitioning threshold is used to compare the feature distance and the partitioning threshold, and the resource blocks are divided into multiple clusters. The resource blocks in each cluster are then integrated to establish a resource block cluster set.

[0144] Specifically, the first resource block As the initial member of the first cluster, create a new cluster, denoted as... At the same time, Add to this cluster In the middle; then, from the second resource block Begin by progressively calculating the device-channel characteristics of each resource block. Device-channel characteristics of the previous resource block The cosine similarity between features is used to determine the cosine distance as the feature distance between two features. The corresponding calculation formula is:

[0145]

[0146] in, Indicates the first The and the first The characteristic distance of the device-channel characteristics of a resource block.

[0147] It can be explained that, Similarly That is, the first The and the first Each resource block is a pair of adjacent resource blocks. The process of calculating the feature distance between them is as follows: The preset threshold for partitioning is denoted as When the characteristic distance of the two devices-channel features When two resource blocks are sufficiently similar to belong to the same cluster, the device-channel characteristics of the resource blocks are considered. Add to current cluster In the middle; conversely, When two resource blocks are not very similar and do not belong to the same cluster, a new cluster is created. The device-channel characteristics of the resource block As the initial members of this cluster, the device-channel characteristics of all resource blocks are similarly divided according to this process until all processing is completed, ultimately yielding all clustering results, i.e., the set of resource block clusters. , This indicates the number of resource block clusters. It should be noted that the above clustering process can be performed on any set of resource blocks within a frame; the example provided only uses sequential clustering of adjacent resource blocks as an example.

[0148] Step S52: Use the trained single resource block device classifier to classify each device-channel feature in the resource block cluster set to obtain the device label of the resource block feature. Based on the device label, use a voting mechanism to determine the device identification result corresponding to the resource block cluster set.

[0149] Specifically, according to step S51, for any user, that is, for any device, all resource blocks are determined to form its corresponding resource block cluster set. , Indicates the first The set of resource block clusters corresponding to each user; Indicates the first One resource block; Indicates the number of resource blocks; calculates the device identification result of the device-channel characteristics for each resource block, and the corresponding calculation formula is:

[0150]

[0151] in, Indicates the first Category index of device-channel feature prediction for each resource block; Indicates that the choice makes Largest index user ; This represents a feature extractor.

[0152] It can be explained that, based on step S5, the device identification results of the device-channel characteristics of all devices are obtained in the same way. The device identification result with the highest frequency is selected as the user's device tag and determined as the device identification result corresponding to the final unknown signal.

[0153] Understandably, the proposed 5G RF fingerprint extraction and recognition method for multi-user access scenarios optimizes the feature distribution in the feature space through hyperspherical metric learning, enabling resource block features from the same user to cluster in a high-dimensional space, while separating the features of different users, thereby improving the separability between users and enhancing the accuracy of user identification in a multi-user environment. Furthermore, by combining a device-channel feature extractor with a single resource block device classifier, the method effectively improves the device RF fingerprint recognition capability in multi-user scenarios through resource block identification and voting mechanisms.

[0154] To better illustrate this, a 5G radio frequency fingerprint extraction and identification method for multi-user access scenarios proposed in this application is applicable to complex 5G multi-user environments and achieves high-precision identification of devices in multi-user scenarios. It can accurately extract radio frequency fingerprint features and effectively identify device identities in scenarios with complex channel conditions and multi-user access, and has extremely high application value. It is suitable for fields such as IoT security, user authentication and device tracking in 5G network environments.

[0155] The second embodiment of the present invention provides an electronic device, which includes a processor, a communication interface, a memory, and a communication bus. The processor, the communication interface, and the memory communicate with each other through the communication bus. The processor calls logical instructions in the memory to execute a 5G radio frequency fingerprint extraction and identification method for multi-user access scenarios as described in any embodiment of the present invention.

[0156] When it is in operation, it needs to use a 5G radio frequency fingerprint extraction and identification method for multi-user access scenarios. Therefore, whether the device and program data are integrated or different hardware is configured to produce a function with similar effect to that achieved by the present invention, it is within the protection scope of the present invention. The device has the same beneficial effect as the aforementioned 5G radio frequency fingerprint extraction and identification method for multi-user access scenarios, and will not be described in detail here.

[0157] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0158] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A 5G radio frequency fingerprint extraction and recognition method for multi-user access scenarios, characterized in that, The method includes: Step S1: Use multiple 5G devices to simulate a multi-user access scenario to send uplink data frames and collect air signals; Step S2: After performing preprocessing on the air signals, including blind signal synchronization, de-resource lattice processing, and data demodulation, resource blocks and their corresponding characteristics are obtained. Step S3: Using airborne signal and resource block features, train a device-channel feature extractor for intra-frame device resource block clustering using hyperspherical metric learning; Step S4: Train a single resource block device classifier using air signals and resource block features; Step S5: Based on the trained device-channel feature extractor and single resource block device classifier, construct a resource block cluster-level device identification system for multi-user scenarios, and output the device identification results corresponding to the air signals through a voting mechanism.

2. The 5G radio frequency fingerprint extraction and recognition method for multi-user access scenarios according to claim 1, characterized in that, Step S2 includes: Step S21: By utilizing the correlation between the cyclic prefix and the symbol tail, blind synchronization of the air signal is completed, and multiple subframes are captured; Step S22: Restore each subframe to a frequency domain resource grid, and decompose it into multiple resource blocks; Step S23: Demodulate and extract features from the resource blocks to obtain the corresponding resource block features.

3. The 5G radio frequency fingerprint extraction and recognition method for multi-user access scenarios according to claim 2, characterized in that, Step S21 includes: Step S211: Determine the symbol length and cyclic prefix length based on the air signal to obtain the correlation function; Step S212: Traverse each subcarrier interval, filter the correlation value corresponding to each possible starting point, and obtain the starting point corresponding to the maximum correlation value; Step S213: Combine the maximum correlation values ​​corresponding to all subcarrier intervals, select the subcarrier interval corresponding to the maximum value as the actual subcarrier interval, determine the starting point of the air signal, obtain the cyclic prefix and symbol length, and extract multiple subframes.

4. The 5G radio frequency fingerprint extraction and recognition method for multi-user access scenarios according to claim 3, characterized in that, In step S22, specifically: Based on the actual subcarrier spacing and the starting point of the air signal, the main body of each symbol is extracted, the data corresponding to the symbol length of the air signal is obtained, the received symbol is restored to a frequency resource grid, and decomposed into multiple resource blocks.

5. A 5G radio frequency fingerprint extraction and recognition method for multi-user access scenarios according to claim 2, characterized in that, Step S23 includes: Step S231: Demodulate each resource block to obtain the ideal signal, and calculate the channel state information of the ideal signal; Step S232: Average the different symbols at the same resource block position within a single frame to obtain the mean value of the subcarrier channel state information; Step S233: Divide the mean of the subcarrier channel state information and extract the envelope features of the CSI vector as resource block features.

6. The 5G radio frequency fingerprint extraction and recognition method for multi-user access scenarios according to claim 1, characterized in that, Step S3 includes: Step S31: Generate device-channel labels for each resource block, construct a source label set, and divide it into a training set and a validation set respectively; Step S32: Construct a device-channel classifier containing a feature extraction network. Input the training set, map the training set to the feature space through the feature extraction network to obtain features, perform hyperspherical projection on the features to obtain optimized features; use a hyperspherical projection classifier to obtain the classification probability of the optimized features, and define the loss function of the resource block features with respect to the feature extractor. Step S33: Repeat step S32 to determine the loss function of the feature extractor; Step S34: Use the gradient descent algorithm to select and update the feature extractor and the hyperspherical projection classifier to complete the training of the device-channel feature extractor.

7. The 5G radio frequency fingerprint extraction and recognition method for multi-user access scenarios according to claim 1, characterized in that, Step S4 includes: Step S41: Generate resource block device tags for each resource block, construct a user tag set, and divide it into training set and validation set respectively; Step S42: Input the training set into the single resource block device classifier to obtain the probability of each category, and define the loss function of the single resource block device classifier obtained from the resource block features; Step S43: Repeat step S42 to determine the loss function of the single resource block device classifier; Step S44: Update the learning parameters of the single resource block device classifier using the gradient descent algorithm to complete the training of the single resource block device classifier.

8. A 5G radio frequency fingerprint extraction and recognition method for multi-user access scenarios according to claim 1, characterized in that, Step S5 includes: Step S51: Use the trained device-channel feature extractor to extract device-channel features from intra-frame resource blocks of air signals, and perform clustering to establish a resource block cluster set; Step S52: Use the trained single resource block device classifier to classify each device-channel feature in the resource block cluster set to obtain the device label of the resource block feature. Based on the device label, use a voting mechanism to determine the device identification result corresponding to the resource block cluster set.

9. A 5G radio frequency fingerprint extraction and recognition method for multi-user access scenarios according to claim 8, characterized in that, In step S51, specifically: Based on the device-channel characteristics corresponding to each pair of resource blocks, the cosine distance is determined as the feature distance. A preset partitioning threshold is used to compare the feature distance and the partitioning threshold, and the resource blocks are divided into multiple clusters. The resource blocks in each cluster are then integrated to establish a resource block cluster set.

10. An electronic device, characterized in that, The device includes a processor, a communication interface, a memory, and a communication bus. The processor, the communication interface, and the memory communicate with each other through the communication bus. The processor calls logical instructions in the memory to execute the 5G radio frequency fingerprint extraction and identification method for multi-user access scenarios as described in any one of claims 1 to 9.

Images