C-v2x small sample radio frequency fingerprint classification method and computer equipment

Abstract

The application discloses a C-V2X small sample radio frequency fingerprint classification method and computer equipment. The method comprises the following steps: acquiring a received signal of a C-V2X physical sidelink broadcast channel, performing channel equalization processing, and obtaining an initial radio frequency fingerprint sequence; a feature extraction network is constructed, a manifold mixing mechanism is introduced in the training stage, the hidden state and label of the training sample are randomly linearly interpolated in the deep feature space, and a virtual feature sample is synthesized; in the metric inference stage, a direct push soft clustering strategy based on confidence weighting is introduced, the class prototype is initialized by using the query set topology structure, the soft assignment probability is calculated, the confidence weight is determined based on the uncertainty, the class prototype is adaptively weighted and updated, the high-confidence sample is dominated to drift, and the classification result is output. Under the premise of not increasing additional inference delay, the application realizes high-robustness identification of a radiation source device in a complex dynamic channel and a data-limited scene.

Description

Technical Field

[0001] This invention relates to the field of radio frequency fingerprint recognition technology, and in particular to a C-V2X small sample radio frequency fingerprint classification method and computer equipment. Background Technology

[0002] With the rapid expansion of the Internet of Vehicles (IoV) market, the number of wireless devices has increased dramatically, and they are widely used in fields such as intelligent transportation. However, this also brings serious security challenges such as malicious attacks and unauthorized access. Traditional wireless device authentication methods mainly rely on upper-layer cryptographic technologies, but they have significant limitations, including vulnerability to spoofing and forgery attacks, complex cross-protocol key management, and difficulty in deploying on a massive number of low-power, low-cost IoT devices. In contrast, Radio Frequency Fingerprint (RFF) identification technology, as a lightweight physical layer authentication method, utilizes inherent, non-clonable hardware defects (such as I / Q imbalance and power amplifier nonlinearity) generated during the manufacturing process of wireless devices as characteristics. It naturally possesses the ability to resist physical layer spoofing attacks and requires no additional key distribution computation overhead.

[0003] In recent years, deep learning has been widely applied in radio frequency fingerprint recognition due to its powerful feature extraction capabilities. However, in real-world non-cooperative or industrial IoV scenarios, prolonged signal acquisition may disrupt the normal operation of critical tasks, making it impractical to obtain a large number of high-quality labeled samples. To address the problem of few-shot learning with scarce data, existing technologies often employ traditional data augmentation methods (such as adding noise, flipping, or input-level data mixing). However, traditional input-level mixing only performs linear interpolation on samples in the input space, which is equivalent to simulating simple signal superposition in physical layer radio frequency signal processing. This easily disrupts the strict phase continuity and underlying physical semantics of the radio frequency signal, leading to a decrease in the model's recognition accuracy.

[0004] Furthermore, the extraction of features from small samples in RF fingerprinting is extremely sensitive to wireless channel distortion. Under small sample conditions, deep learning models cannot compensate for random channel fading through statistical averaging of massive samples, as they can in large-scale data scenarios. Existing techniques typically employ traditional least squares algorithms for channel estimation, which rely solely on pilot signals for isolated point estimations, ignoring the channel's autocorrelation in the frequency domain and exhibiting poor noise resistance. This directly results in the extremely small number of fingerprint samples being heavily contaminated with instantaneous channel noise and multipath interference. This feature contamination causes representational bias in small-sample classification models, leading to overfitting to the current instantaneous channel state rather than the inherent RF hardware characteristics of the radiating device. This not only drastically amplifies the intra-class variance of similar samples but also completely disrupts the topological structure of device features in the hidden space.

[0005] Furthermore, when completing feature extraction and entering the metric inference stage of small-sample classification, traditional inductive methods construct class prototypes based on only a very small support set, which easily leads to severe distribution shifts. While existing inductive inference strategies attempt to transform classification into an unsupervised clustering problem to utilize the distribution structure of the query set, traditional inductive soft clustering algorithms typically assign indiscriminate probability weights to all query samples when updating class prototypes. This approach ignores the fact that "fuzzy samples" at the decision boundary (i.e., samples that output similar probabilities for multiple categories) introduce significant noise interference to class prototype updates, causing the final calculated cluster centers to deviate from the true physical manifold, making it difficult to achieve high-precision inter-class separation in complex transmission scenarios such as non-line-of-sight.

[0006] In summary, in order to overcome the limitations of existing technologies in practical vehicle networking scenarios, seeking a novel radio frequency fingerprint authentication scheme that can overcome dynamic multipath channel distortion, extract robust features and achieve high-precision classification under extremely limited data conditions is an important technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a radio frequency fingerprint classification method, the detailed technical solution of which is as follows: A C-V2X small-sample radio frequency fingerprint classification method, comprising: S100: Obtain the received signal of the C-V2X physical side link broadcast channel, perform channel equalization processing on the received signal, and obtain the initial radio frequency fingerprint sequence after eliminating channel interference. S200. Construct a feature extraction network and introduce a manifold mixing mechanism during the training phase: In the deep feature space of the feature extraction network, perform random linear interpolation on the hidden states and corresponding class labels of the training samples to synthesize virtual feature samples for network training. S300. In the metric inference stage, a direct soft clustering strategy based on confidence weighting is introduced: the overall manifold topology of the query set is used to support the initialization of various prototype centers by the set samples, the soft assignment probability from the query sample to each prototype center is calculated, and the confidence weight is determined based on the uncertainty of the soft assignment probability distribution of each query sample. When iteratively updating the class prototype, the assignment probability of each query sample is adaptively weighted with the confidence weight, so that high-confidence samples dominate the class prototype drift until the class prototype converges. The classification result is output according to the distance from the query sample to each prototype center after convergence.

[0008] In some embodiments, the manifold mixing mechanism in step S200 specifically includes: Decompose the feature network into steps from input to the first... The first mapping of the layer and from the first Second mapping from layer to output ,in For mixed-layer indexes; Randomly select a blending layer index from a set containing an input layer and at least one intermediate hidden layer. ; A virtual feature map is generated by linearly weighting and fusing the first mapping outputs of two training samples at a selected layer using the mixing coefficient λ. Virtual labels are generated by linear interpolation with the same coefficients on the class labels of the two samples. ; The virtual feature map is then further processed through the second mapping. Once the predicted output is obtained, the loss between the predicted output and the virtual label is calculated, and the network parameters are updated via backpropagation.

[0009] In some embodiments, the mixing coefficient λ follows a Beta distribution. , where α is a preset hyperparameter.

[0010] In some embodiments, the confidence-weighted pushover soft clustering strategy described in step S300 specifically includes: S310. Initialize various prototype centers using the mean of the support set samples. For query samples The Softmax function is used to calculate its belonging to the th order. The soft assignment probability of the class; , in, Here, N is the scale parameter, and N is the total number of categories. S320. Calculate the information entropy of the soft assignment probability for each query sample. The confidence weight is determined based on the information entropy, where the smaller the information entropy, the larger the confidence weight. S330. Iteratively update various prototype centers according to the following update rules: , in, Indicates the first The support set of the class is a sample set. For its corresponding sample size, For query set, Confidence weights; S340. Repeat the iteration until the class prototype center converges, and assign the query sample to the category corresponding to the class prototype center that is closest to it.

[0011] In some embodiments, the confidence weight is taken as an exponential decay function. .

[0012] In some embodiments, step S100 involves performing channel equalization processing on the received signal using a linear minimum mean square error algorithm, specifically including: The channel power delay spectrum is obtained by using the frequency domain least squares response of multiple reference symbols, and the root mean square delay spread is calculated. Construct the frequency domain channel autocorrelation matrix based on root mean square delay spread; By combining the prior signal-to-noise ratio and the aforementioned frequency domain channel autocorrelation matrix, the LMMSE weights are calculated to obtain the channel estimate. The received signal is equalized using the channel estimate, and the radio frequency fingerprint sequence is extracted.

[0013] In some embodiments, the feature extraction network uses a multi-layer cascaded residual network as its backbone framework. The residual network contains multiple cascaded residual modules, each of which consists of a convolutional layer, a batch normalization layer, and an activation function, and is used to extract deep features of the radio frequency fingerprint layer by layer.

[0014] In some embodiments, in step S100, before performing channel equalization processing on the received signal, the following preprocessing is performed sequentially: A block energy gradient and threshold determination method is used for signal detection to locate the signal starting point; Coarse synchronization is performed using the periodicity of the master synchronization signal, and fine synchronization is performed using the correlation of the cyclic prefix to determine the frame start position. Carrier frequency offset estimation and compensation are performed using the phase difference of the conjugate cross-correlation of consecutive synchronization symbols.

[0015] In some embodiments, the radio frequency fingerprint classification method further includes: Using the trained feature extraction network and the direct soft clustering strategy, steps S100, S200 and S300 are executed sequentially on the physical side link broadcast channel received signal of the C-V2X device to be identified in an unknown scenario, and the radio frequency fingerprint classification result of the device is output to complete the physical layer identity authentication of the device.

[0016] The C-V2X small-sample radio frequency fingerprint classification method provided in this application has the following technical advantages: 1. Improve the noise and multipath resistance of channel estimation: By performing channel equalization on the received signal, multipath fading and channel distortion are effectively removed, resulting in a cleaner initial RF fingerprint sequence, providing high-quality input for subsequent feature extraction.

[0017] 2. Avoid the destruction of signal physical structure by traditional input-level data augmentation: Instead of mixing the original signal at the input layer, the hidden state is randomly linearly interpolated in the deep feature space of the feature extraction network. This allows for the synthesis of virtual training samples without destroying the original phase continuity and underlying physical semantics of the radio frequency signal, thus improving the robustness of feature representation under small sample conditions.

[0018] 3. Suppress the interference of ambiguous samples at the boundary on the cluster center: In the direct-effect soft clustering inference process, the confidence weight is determined based on the uncertainty of the soft assignment probability distribution of each query sample, and adaptive weighting is performed when iteratively updating the class prototype, so that high confidence samples dominate the class prototype drift, thereby effectively suppressing the interference of ambiguous samples at the decision boundary on the class center and improving the classification accuracy.

[0019] 4. Improve the accuracy and robustness of small sample recognition under dynamic multipath channels: Achieve high-precision and high-robustness device RF fingerprint classification in vehicle networking scenarios with scarce data and complex channel environments (such as NLOS and Doppler shift).

[0020] This application also 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 C-V2X small sample radio frequency fingerprint classification method described in any of the preceding claims. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the execution of the C-V2X small-sample radio frequency fingerprint classification method in this application embodiment; Figure 2 This is a schematic diagram of the PSBCH substructure frame in an embodiment of this application; Figure 3 This is the overall framework of the C-V2X small sample radio frequency fingerprint classification method in the embodiments of this application.

[0022] Figure 4 This is a visualization of the t-SNE feature space in the embodiments of this application; Figure 5 This is a diagram showing the recognition accuracy under different signal-to-noise ratios in the embodiments of this application; Figure 6 This is a comparison chart of the classification performance of different measurement methods in the embodiments of this application. Detailed Implementation

[0023] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be described in detail below with reference to the accompanying drawings. The description in this part is only exemplary and explanatory, and should not be used to limit the scope of protection of the present invention in any way.

[0024] Before describing the embodiments of this application in detail, the relevant technical terms used in this application will be explained in a unified manner as follows: C-V2X: Cellular Vehicle-to-Everything, refers to a communication system between a vehicle and all external entities based on cellular communication technology, and is the application scenario of the method in this application.

[0025] PSBCH: Physical Sidelink Broadcast Channel, is a channel used for broadcast transmission in C-V2X. The received signals processed in this application originate from this channel.

[0026] LMMSE: Linear Minimum Mean Square Error, is a channel estimation algorithm that optimizes channel response estimation by utilizing channel frequency domain autocorrelation and noise statistics. This application uses this algorithm for channel equalization processing.

[0027] Radio Frequency Fingerprint (RFF): Inherent, non-clonable hardware defects (such as I / Q imbalance, power amplifier nonlinearity, etc.) generated during the manufacturing process of wireless devices can be used as a unique identifier for physical layer authentication.

[0028] Support Set: In few-shot learning tasks, each class contains a small set of labeled samples used to initialize class prototypes or train classifiers.

[0029] Query Set: In few-shot learning tasks, it is the set of unlabeled samples to be classified, used for evaluation during the inference phase.

[0030] Class Prototype: The representative center point of each class in the feature space, usually initialized by the mean of the support set samples and iteratively updated during the transductive clustering process.

[0031] Soft Assignment Probability: The probability distribution of a query sample belonging to each category, with values ​​ranging from 0 to 1, and the sum of the probabilities of each category is 1.

[0032] Information Entropy: Used to measure the degree of uncertainty in a soft-assignment probability distribution. The smaller the entropy value, the higher the classification confidence; the larger the entropy value, the closer the sample is to the decision boundary.

[0033] Confidence Weight: A dynamic weighting coefficient constructed based on information entropy, used to adjust the contribution of each query sample when updating the class prototype. High confidence samples receive a larger weight, while the weight of low confidence (fuzzy boundary) samples is suppressed.

[0034] Manifold Mixup: A data augmentation mechanism that performs random linear interpolation of the hidden states and labels of samples in the deep feature space of a neural network to improve the model's generalization ability under small sample conditions.

[0035] Example 1: Reference Figure 1 and Figure 3 As shown, the C-V2X small-sample radio frequency fingerprint classification method provided in this application includes the following steps: S100: Obtain the received signal of the C-V2X Physical Sidelink Broadcast Channel (PSBCH), perform channel equalization processing on the received signal, and obtain the initial radio frequency fingerprint sequence after removing channel interference.

[0036] Optionally, the Linear Minimum Mean Square Error (LMMSE) algorithm can be used to perform channel equalization on the received signal. This algorithm can fully utilize the channel's frequency domain autocorrelation and noise statistics, effectively suppress Gaussian white noise and multipath interference, and maintain high estimation accuracy even at low signal-to-noise ratios (e.g., 0 dB). Furthermore, it can accurately remove channel distortion and extract RF fingerprints that more closely approximate inherent hardware defects, providing high-quality input for subsequent few-sample learning and preventing the model from overfitting to the instantaneous channel state.

[0037] Of course, other algorithms such as Wiener filter and Kalman filter can also be used to perform channel equalization processing on the received signal.

[0038] S200. Construct a feature extraction network and introduce a manifold mixing mechanism during the training phase: In the deep feature space of the feature extraction network, perform random linear interpolation on the hidden states and corresponding class labels of the training samples to synthesize virtual feature samples for network training.

[0039] S300. In the metric inference stage, a confidence-weighted inductive soft clustering (EW-SKM) strategy is introduced: The overall manifold topology of the query set is utilized to support the initialization of various prototype centers by set samples; the soft assignment probability from the query sample to each prototype center is calculated; the confidence weight is determined based on the uncertainty of the soft assignment probability distribution of each query sample; during iterative updates of class prototypes, the assignment probability of each query sample is adaptively weighted with the aforementioned confidence weight, so that high-confidence samples dominate class prototype drift; until the class prototypes converge, the classification result is output based on the distance from the query sample to each converged prototype center.

[0040] The C-V2X few-sample RF fingerprint classification method provided in this application firstly uses LMMSE channel equalization to effectively remove multipath fading and channel distortion, obtaining a pure initial RF fingerprint sequence, providing high-quality input for subsequent feature extraction. Secondly, a manifold mixing mechanism is introduced during the training phase to interpolate the hidden states in the deep feature space, avoiding the problem of traditional input-level data augmentation destroying the phase continuity of RF signals, and significantly improving the feature generalization ability under few-sample conditions. Finally, in the metric inference phase, a confidence-weighted direct soft clustering strategy is adopted, using information entropy to measure the uncertainty of each query sample, reducing the weight of samples with blurred boundaries, and allowing high-confidence samples to dominate the class prototype drift, effectively suppressing the pulling interference of blurred samples on the cluster center. The synergistic effect of these three steps enables this application to achieve high-precision and robust device RF fingerprint classification in vehicular network scenarios with scarce data and complex channels (such as non-line-of-sight NLOS and Doppler shift).

[0041] Example 2: Based on Example 1, this embodiment provides a detailed explanation of the specific implementation of the preprocessing of the received signal and LMMSE channel equalization in step S100.

[0042] The PSBCH standard subframe structure in C-V2X consists of two time slots, and its internal symbols are as follows: Figure 2 As shown. This embodiment performs preprocessing operations (including signal detection, frame synchronization, carrier frequency offset compensation, and channel equalization) on the received PSBCH signal. The specific steps are as follows: Before channel equalization is performed, the received signal undergoes the following preprocessing steps in sequence: Signal detection: The starting point of the signal is located using a block energy gradient and threshold determination method. The length of the signal block is defined as... traverse and receive signals According to block length Divide the data into non-overlapping segments and calculate the energy block by block. When the energy ratio of adjacent blocks exceeds a preset threshold. At that time, determine the starting point of the signal.

[0043] Frame synchronization: Coarse synchronization is performed using the periodicity of the Primary Sidelink Synchronization Signal (PSSS). The received signal is cross-correlated with a locally generated normalized sequence, and correlation peaks are searched to verify that their intervals conform to the standard subframe length (e.g., 2048 sampling points). Subsequently, fine synchronization is performed near the coarse synchronization position using the physical characteristic that the cyclic prefix (CP) is identical to the tail data of its corresponding OFDM symbol. The correlation of CPs of multiple symbols within a frame is accumulated, and the offset of the maximum value is selected as the final fine synchronization compensation point.

[0044] Carrier Frequency Offset (CFO) compensation: Utilizing the characteristic that two consecutive primary synchronization symbols are identical, and two consecutive secondary synchronization symbols are identical, a specific time-domain synchronization segment is extracted. The phase deflection of the complex conjugate cross product of the subsequent symbol and the preceding symbol is calculated. Reverse phase rotation compensation is then applied to the original signal to obtain the frequency offset-calibrated signal. .

[0045] LMMSE Channel Estimation and Equalization: Channel estimation is performed using the linear minimum mean square error algorithm. The specific steps are as follows: (1) Obtain the channel power delay spectrum using the frequency domain least square response of multiple reference symbols (such as demodulation reference signal DMRS) and calculate the root mean square (RMS) delay spread.

[0046] (2) Construct the frequency domain channel autocorrelation matrix RHH based on the root mean square delay spread. Specifically, the elements of the channel autocorrelation matrix are constructed based on the estimated channel power and the delay-related phase offset, which is obtained based on the frequency difference between the subcarriers.

[0047] (3) Calculate the LMMSE estimation matrix by combining the prior signal-to-noise ratio and the frequency domain channel autocorrelation matrix to obtain the channel estimate.

[0048] (4) Use the channel estimation value to equalize the received signal, that is, remove the channel characteristic interference through channel equalization and extract the radio frequency fingerprint sequence of the inherent hardware damage characteristics of the carrying device.

[0049] Optionally, in low signal-to-noise ratio scenarios, if the root mean square delay spread calculation result is invalid, an empirical preset value can be set by default to ensure estimation stability. The initial frequency domain channel response is averaged using multiple symbols to smooth out noise effects, and the channel autocorrelation matrix is ​​Hermitian-symmetricized to eliminate numerical errors. Finally, the frequency domain estimation is transformed to the time domain and a windowing operation is performed to retain only the main multipath clusters with concentrated energy, suppressing broadcast noise and redundant features.

[0050] This embodiment replaces traditional least-squares estimation with LMMSE channel estimation, making full use of the channel's autocorrelation in the frequency domain and significantly improving estimation accuracy under low signal-to-noise ratio conditions. Simultaneously, through operations such as cyclic prefix-assisted fine synchronization and multi-symbol averaging, multipath interference and burst noise are effectively suppressed, providing a high-purity RF fingerprint feature base for subsequent small-sample feature extraction.

[0051] Example 3: This embodiment, based on Embodiment 1, provides a detailed explanation of the manifold mixing mechanism introduced in step S200.

[0052] The feature extraction network constructed in this application adopts a multi-layer cascaded residual network (ResNet) as the backbone framework. The residual network contains multiple (e.g., 4) cascaded residual blocks. Each residual block consists of a convolutional layer, a batch normalization layer, and an activation function (e.g., LeakyReLU). At the end of the block, a max pooling layer is used to reduce the feature dimension, which is used to extract deep features of RF fingerprints layer by layer.

[0053] Decompose the feature network into steps from input to the first... The first mapping of the layer and from the first Second mapping from layer to output ,Right now ,in For hybrid layer indexes, These are the trainable parameters of the network.

[0054] The specific operation of the manifold mixing mechanism is as follows: (1) Randomly from the set Select a hybrid layer index ,in Represents input layer mixing. Representing the Feature space mixing after each residual module.

[0055] (2) For a randomly selected pair of training samples and Using Beta to generate mixing coefficients ,in For preset hyperparameters (e.g.) ), Control the mixing ratio of the two samples.

[0056] (3) In the selected mixing layer At this point, the first mapping outputs of the two samples are linearly weighted and fused to generate a virtual feature map: .

[0057] (4) Perform linear interpolation with the same coefficients on the class labels of the two samples to generate virtual labels: .

[0058] (5) Virtual feature map Continue through the rest of the network Obtain the predicted output, and calculate the predicted output and the virtual label. The loss (e.g., cross-entropy loss) is calculated and backpropagated to update the network parameters.

[0059] It should be noted that the mixing operation is performed not only in the input layer, but also with a certain probability in the intermediate hidden layers of the network. This deep feature space interpolation avoids directly mixing the original I / Q signals in the input layer, thereby protecting the tight phase continuity of the RF signal and the underlying physical semantics.

[0060] This embodiment synthesizes virtual samples in a deep, high-dimensional feature space through a manifold mixing mechanism, which is equivalent to data augmentation at the feature level. Compared to traditional input-level mixups, this method does not destroy the original physical structure of the RF signal, while enabling the model to learn smoother decision boundaries, making it better tolerant of RF fingerprint feature distribution shifts caused by non-stationary channel noise. Under small sample conditions, this mechanism significantly improves the intra-class compactness and inter-class separability of feature clusters.

[0061] Example 4: Based on Example 1, this embodiment provides a detailed explanation of the pushover soft clustering strategy introduced in step S300.

[0062] In the metric inference phase, this application transforms the small-sample classification task into a transductive unsupervised clustering problem involving the entire batch of query samples, and uses the overall manifold topology of the query set to adaptively iteratively update the class prototypes. The specific steps are as follows: S310. Initialize class prototype centers: Initialize the prototype centers of each class with the mean of the support set samples. ,Right now , in For the first The support set of the class is a sample set. Its sample size.

[0063] S320. Calculate the soft assignment probability: For the query sample The Softmax function is used to calculate its belonging to the th The soft assignment probability of a class is measured by calculating the Euclidean distance between a sample and the prototype centers of each class, and a scale parameter is introduced. Adjusting the smoothness of the probability distribution: , in, For scale parameters (e.g.) ), This represents the total number of categories.

[0064] S330. Calculating Information Entropy Confidence Weights: To quantify the model's classification confidence in the query samples, Shannon information entropy is introduced to measure the topological uncertainty of the soft assignment probability. , Information entropy The size of the cluster center directly reflects the distribution of samples in the feature space: when samples are close to a cluster center, the probability of their distribution is highly concentrated. Approaching a minimum value indicates extremely high classification confidence; conversely, if a sample falls into the inter-class aliasing region due to multipath distortion, its allocation probability tends to be more uniform. Significantly increased.

[0065] Based on the above physical intuition, this application constructs an exponential decay term. As dynamic attention weights (i.e. confidence weights), the smaller the information entropy, the larger the confidence weight.

[0066] S340, Weighted Update Class Prototype: In the... The second iteration update Class Prototype Center At that time, the update rule is defined as: , in, Indicates the first The support set of the class is a sample set. Its sample size, This indicates an unlabeled query set.

[0067] This weighting mechanism ensures that only high-confidence query samples can dominate the spatial drift of class prototypes, thereby effectively cutting off the pulling effect of ambiguous boundary samples on the class center.

[0068] S350, Iterative Convergence and Classification: Repeat the above steps (S320 to S340) until the class prototype center positions converge (e.g., the sum of the class prototype changes in two adjacent iterations is less than a preset threshold, or the preset maximum number of iterations is reached). The final query sample category is determined by the shortest Euclidean distance to each converged category center.

[0069] Figure 4 The t-distribution random neighborhood embedding clustering distribution of four radio frequency devices (CX1~CX4) in the feature space is presented. Figure 4 As shown in (a), in a static scene, the feature clusters of each device are extremely compact and the inter-class boundaries are clear; however, in a mobile scene facing severe interference from multipath and Doppler frequency shift (such as...), the feature clusters of each device are extremely compact and the inter-class boundaries are clear. Figure 4 (b) shows that although the intra-class variance of the features increased, the bias caused by blurred samples was effectively suppressed by the introduced information entropy confidence weighting mechanism. The model was still able to robustly classify the features of the four devices into their respective correct physical manifold regions, achieving high-precision inter-class separation.

[0070] This embodiment overcomes the shortcomings of traditional direct-means methods that assign indiscriminate weights to all query samples by introducing an information entropy-weighted pushover soft clustering strategy. In non-line-of-sight transmission or dynamic movement scenarios, severe multipath effects and Doppler shifts can lead to overlapping of deep feature boundaries, making traditional methods prone to being pulled off-center by samples with blurred boundaries. This application utilizes information entropy to adaptively evaluate the confidence level of each query sample, assigning greater weight to samples with high confidence and significantly reducing the weight of samples with blurred boundaries, thereby ensuring that the class prototypes drift towards the true physical manifold and achieving high-precision inter-class separation in data-constrained scenarios.

[0071] Example 5: This embodiment describes the detailed structure of the feature extraction network based on Embodiment 1.

[0072] This application employs a multi-layered cascaded residual network (ResNet) as the backbone framework to construct a fingerprint feature mapping model. The network input is preprocessed radio frequency fingerprint data (i.e., the initial radio frequency fingerprint sequence obtained after LMMSE channel equalization). The data is separated into real and imaginary parts and reshaped into a two-dimensional feature matrix for input into the network.

[0073] For example, this network structure mainly consists of four cascaded residual blocks, each containing a convolutional layer, a batch normalization layer, and an activation function (such as ReLU). The specific configuration is as follows: First residual module: 64 output channels, feature map size 56×56.

[0074] Second residual module: 128 output channels, feature map size 28×28.

[0075] The third residual module has 256 output channels and a feature map size of 14×14.

[0076] Fourth residual module: 512 output channels, feature map size 7×7.

[0077] Each residual module employs skip connections to mitigate the vanishing gradient problem. During training, the feature extraction network and manifold hybrid mechanism are jointly optimized, and network parameters are updated through backpropagation.

[0078] It should be noted that the specific number of layers, channels, and convolutional kernel size of the feature extraction network can be adjusted according to the actual dataset size and computing resources. The configuration provided in this embodiment is a preferred implementation and does not constitute a limitation on the scope of protection of this application.

[0079] The residual network structure used in this embodiment can effectively extract multi-level features of RF fingerprints, outputting comprehensive features at shallow levels and detailed features at deeper levels. Through cascaded residual modules, the network can clearly map the fingerprint information of the RF signal to the feature space, providing a high-quality embedding representation for subsequent transductive soft clustering.

[0080] Example 6: This embodiment, based on embodiment 1, further includes application steps after the model training is completed.

[0081] Using the trained feature extraction network and the direct soft clustering strategy, steps S100, S200 and S300 are executed sequentially on the physical side link broadcast channel received signal of the C-V2X device to be identified in an unknown scenario, and the radio frequency fingerprint classification result of the device is output to complete the physical layer identity authentication of the device.

[0082] Specifically, during the data acquisition phase, a software-defined radio receiver (e.g., USRP B205mini) is used to collect C-V2XPSBCH signals. Operating parameters can be set as follows: center frequency 5.9 GHz, sampling rate 30.72 Msps, and receiving bandwidth 20 MHz. The collected data is divided into a training set and a test set (e.g., 7:3). After learning the parameters of the feature extraction network and the transductive soft clustering strategy on the training set, for each device signal to be identified in the test set, the trained model is directly applied for inference to output the device category.

[0083] This embodiment applies the proposed small-sample radio frequency fingerprint classification method to a real C-V2X communication scenario. During the training phase, only a very small number of labeled samples (e.g., 1 or 5 samples per class) are needed to achieve high-precision device identity authentication under dynamic channel conditions. This avoids the cost of large-scale data collection and labeling required by traditional methods and has high engineering practical value.

[0084] Example 7: A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method as described in any one of Embodiments 1 to 6.

[0085] Experiment and Results Analysis: To verify the effectiveness of the C-V2X radio frequency fingerprint classification method proposed in this application, 12 identical LTE-V2X modules were used to transmit PSBCH subframes, with a total bandwidth of 20MHz, a carrier frequency of 5.9GHz, and a sampling rate of 30.72Msps. A general-purpose software radio peripheral B205 was used to receive the signals. The data acquisition environment was divided into stationary scenarios (LOS and NLOS) and mobile scenarios (MOV1 (LOS) and MOV2 (NLOS)). All transmitted PSBCH subframes were randomized, and approximately 1000 subframes were obtained per device in each scenario. After preprocessing and fingerprint extraction, training and test datasets were constructed. All models were implemented using Python 3.12, and each model was trained for approximately 100 epochs. The model with the best results was selected for testing.

[0086] Overall accuracy is used as the primary evaluation metric, and its calculation formula is as follows: , in, To actually belong to the category And it was correctly classified into categories by the model. The sample, For those that do not actually belong to the category However, it was incorrectly classified into a category by the model. The sample, To actually belong to the category However, samples that were incorrectly classified into other categories by the model, For those that do not actually belong to the category And the sample was correctly predicted as another category.

[0087] I. Cross-experiments in different scenarios: This application designed a cross-scenario classification experiment, which included four typical physical channel scenarios. The classification accuracy of 4-way 1-shot and 5-shot is shown in Table 1.

[0088] Table 1. Classification under different scenarios: When the training and test sets are in the same channel scenario, the model exhibits extremely high recognition accuracy. For example, in the MOV2 scenario, the 1-shot and 5-shot accuracies of the same distribution test are as high as 94.25% and 99.31%, respectively, indicating that the backbone network combined with the manifold hybrid feature extraction strategy of this application can fully and accurately uncover the inherent features of RF hardware hidden behind complex distortions.

[0089] For cross-scenario testing, when the model is trained in a clean LOS scenario and tested in a MOV2 scenario containing dynamic multipath interference and Doppler shift, performance degrades to some extent. This is because static line-of-sight channels lack complex multipath interference, and the model fails to fully learn robust boundaries against severe waveform superposition. Conversely, when the model is trained in complex NLOS or MOV2 scenarios and tested in a simple LOS scenario, it still maintains extremely high accuracy. This indicates that training in complex distorted channels forces the network to abandon overfitting to the instantaneous channel state and instead delve into the inherent hardware nonlinear impairment characteristics of the radiation source, thus enabling it to perform the classification task more stably when the test environment is transformed into an ideal channel.

[0090] II. Classification experiments under different signal-to-noise ratios: The radio frequency signal at the receiving end is inevitably subject to severe noise interference. This application conducted small-sample classification experiments within a signal-to-noise ratio range of 0dB to 30dB to evaluate the system's noise robustness. The results are as follows: Figure 5 As shown in the figure (horizontal axis represents signal-to-noise ratio, and vertical axis represents classification accuracy), experimental results show that the classification accuracy of the model generally increases steadily with the increase of SNR, and approaches 99% at 30dB, confirming that the backbone network of this application has an extremely high upper limit of feature representation under ideal conditions.

[0091] In the harsh low signal-to-noise ratio range of 0dB to 10dB, strong noise energy can overwhelm the microscopic physical distortion characteristics of the radio frequency signal. At this point, a 1-shot task relying solely on a single sample is prone to initial prototype shifts due to random noise contamination, resulting in an accuracy of approximately 82%. In contrast, the accuracy of a 5-shot task remains above 94%, indicating that adding a small number of support samples can effectively neutralize Gaussian noise fluctuations and provides strong anti-interference capabilities.

[0092] III. Comparison of Different Measurement Methods: This application tested the classification accuracy in 1-shot and 5-shot scenarios under two typical conditions: moving and stationary. It compared the accuracy with inductive reasoning methods (Euclidean mean, cosine mean), hard K-Means (Hard-KM), soft K-Means (Soft-KM), and the information entropy-weighted soft K-Means (EW-SKM) method proposed in this application. The results are as follows: Figure 6 As shown.

[0093] Experiments show that the transductive clustering algorithm outperforms other algorithms overall. Especially in mobile 1-shot scenarios lacking prior information, the NCM method achieves an accuracy of only around 85%; however, after introducing the transductive mechanism, the accuracy reaches over 90%. This indicates that in complex physical channels, class prototypes constructed using only a very small number of support set samples will suffer severe distribution shifts. The transductive method, by iterating through the unlabeled query set of the target scenario, can effectively utilize the manifold topology of the overall features, thereby significantly smoothing and correcting the initial decision boundary.

[0094] In the internal comparison of the direct-push methods, the hard allocation strategy adopted by Hard-KM is prone to misallocation of ambiguous samples at the boundary, which in turn biases the iteration direction of the prototype. Soft-KM adopts probabilistic allocation, which better preserves the soft transition information between categories, making the smooth drift of the prototype more consistent with the distribution law of real physical signals.

[0095] The proposed EW-SKM method achieved state-of-the-art classification performance in all four test scenarios. EW-SKM introduces information entropy as a confidence penalty term, actively weakening the weight of high-entropy (low-confidence) samples during prototype updates and amplifying the guiding role of low-entropy (high-confidence) samples. This ensures that the model maintains high classification reliability and inference robustness even when facing the dual challenges of severe channel distortion and extreme sample scarcity.

[0096] IV. Enhancement Strategy Analysis: This application compares the accuracy of the baseline model, input-level mixup, and manifold mixup on 1-shot and 5-shot tasks with low signal-to-noise ratio, and calculates the average inference time for a single task. The specific results are shown in Table 2.

[0097] Table 2 Evaluation under different enhancement strategies: The baseline model achieves a 1-shot accuracy of 94.14% without enhancement strategies. Traditional input-level mixing offers a slight performance improvement. In contrast, the feature manifold mixing introduced in this application demonstrates significant advantages, with 1-shot and 5-shot accuracies jumping to 95.74% and 96.76%, respectively. By shifting the interpolation operation to a deep, high-dimensional feature space, physical distortions at the input can be avoided, manifold gaps between device feature clusters can be filled, and the network can learn smoother, more robust decision boundaries, mitigating overfitting issues with small sample sizes. Furthermore, the manifold mixing strategy optimizes the feature space only during the training phase, incurring no additional computational overhead during inference deployment, thus balancing the recognition accuracy and low latency requirements of C-V2X systems.

[0098] In summary, the C-V2X small-sample RF fingerprint classification method proposed in this application based on manifold mixing and transductive soft clustering has significant advantages in improving classification accuracy in small sample cases and coping with complex environments, and can still maintain high recognition accuracy when facing low signal-to-noise ratio conditions.

[0099] This article uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only for the purpose of helping to understand the method and core ideas of the present invention. The above descriptions are only preferred embodiments of the present invention. It should be noted that due to the limitations of textual expression, there are objectively infinite specific structures. For those skilled in the art, several improvements, modifications, or changes can be made without departing from the principles of the present invention, and the above technical features can also be combined in an appropriate manner. These improvements, modifications, changes, or combinations, or the direct application of the inventive concept and technical solution to other situations without modification, should all be considered within the scope of protection of the present invention.

Claims

1. A C-V2X small-sample radio frequency fingerprint classification method, characterized in that, The C-V2X small-sample radio frequency fingerprint classification method includes: S100: Obtain the received signal of the C-V2X physical side link broadcast channel, perform channel equalization processing on the received signal, and obtain the initial radio frequency fingerprint sequence after eliminating channel interference; S200. Construct a feature extraction network and introduce a manifold mixing mechanism during the training phase: In the deep feature space of the feature extraction network, perform random linear interpolation on the hidden state and corresponding category label of the training sample to synthesize virtual feature samples for network training. S300. In the metric inference stage, a direct soft clustering strategy based on confidence weighting is introduced: the overall manifold topology of the query set is used to support the initialization of various prototype centers by the set samples; the soft assignment probability from the query sample to each prototype center is calculated; the confidence weight is determined based on the uncertainty of the soft assignment probability distribution of each query sample; when iteratively updating the class prototype, the assignment probability of each query sample is adaptively weighted with the confidence weight, so that high-confidence samples dominate the class prototype drift; until the class prototype converges, the classification result is output according to the distance from the query sample to each prototype center after convergence. The confidence-weighted pushover soft clustering strategy described in step S300 specifically includes: S310. Initialize various prototype centers using the mean of the support set samples. For the query sample The Softmax function is used to calculate its belonging to the th The soft assignment probability of the class; ， in, Here, N is the scale parameter, and N is the total number of categories. S320. Calculate the information entropy of the soft assignment probability for each query sample. The confidence weight is determined based on the information entropy, where the smaller the information entropy, the larger the confidence weight. S330. Iteratively update various prototype centers according to the following update rules: ， in, Indicates the first The support set of the class is a sample set. For its corresponding sample size, For query set, Confidence weights; S340. Repeat the iteration until the class prototype center converges, and assign the query sample to the category corresponding to the class prototype center that is closest to it.

2. The C-V2X small-sample radio frequency fingerprint classification method as described in claim 1, characterized in that, The manifold mixing mechanism described in step S200 specifically includes: Decompose the feature network into steps from input to the first... The first mapping of the layer and from the first Second mapping from layer to output ,in For mixed-layer indexes; Randomly select a blending layer index from a set containing an input layer and at least one intermediate hidden layer. ; Using the mixing coefficient λ to analyze two training samples and Linear weighted fusion is performed on the first mapping output of the selected layer to generate a virtual feature map. Virtual labels are generated by linear interpolation with the same coefficients on the class labels of the two samples. ,in, , These represent the input data for the two training samples, respectively. , These represent the class labels of the two training samples, respectively. The virtual feature map is then further processed through the second mapping. Once the predicted output is obtained, the loss between the predicted output and the virtual label is calculated, and the network parameters are updated via backpropagation.

3. The C-V2X small-sample radio frequency fingerprint classification method as described in claim 2, characterized in that, The mixing coefficient λ follows a Beta distribution. , where α is a preset hyperparameter.

4. The C-V2X small-sample radio frequency fingerprint classification method as described in claim 1, characterized in that, The confidence weight is taken as an exponential decay function. .

5. The C-V2X small-sample radio frequency fingerprint classification method as described in claim 1, characterized in that, In step S100, the received signal is subjected to channel equalization processing using a linear minimum mean square error algorithm, specifically including: The channel power delay spectrum is obtained by using the frequency domain least squares response of multiple reference symbols, and the root mean square delay spread is calculated. Construct the frequency domain channel autocorrelation matrix based on root mean square delay spread; By combining the prior signal-to-noise ratio and the aforementioned frequency domain channel autocorrelation matrix, the LMMSE weights are calculated to obtain the channel estimate. The received signal is equalized using the channel estimate, and the radio frequency fingerprint sequence is extracted.

6. The C-V2X small-sample radio frequency fingerprint classification method as described in claim 1, characterized in that, The feature extraction network uses a multi-layer cascaded residual network as its backbone framework. The residual network contains multiple cascaded residual modules, each consisting of a convolutional layer, a batch normalization layer, and an activation function, used to extract deep features of the radio frequency fingerprint layer by layer.

7. The C-V2X small-sample radio frequency fingerprint classification method as described in claim 1, characterized in that, In step S100, before performing channel equalization processing on the received signal, the following preprocessing is performed sequentially: A block energy gradient and threshold determination method is used for signal detection to locate the signal starting point; Coarse synchronization is performed using the periodicity of the master synchronization signal, and fine synchronization is performed using the correlation of the cyclic prefix to determine the frame start position. Carrier frequency offset estimation and compensation are performed using the phase difference of the conjugate cross-correlation of consecutive synchronization symbols.

8. The C-V2X small-sample radio frequency fingerprint classification method as described in claim 1, characterized in that, The radio frequency fingerprint classification method further includes: Using the trained feature extraction network and the direct soft clustering strategy, steps S100, S200 and S300 are executed sequentially on the physical side link broadcast channel received signal of the C-V2X device to be identified in an unknown scenario, and the radio frequency fingerprint classification result of the device is output to complete the physical layer identity authentication of the device.

9. A computer device, characterized in that, It includes 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 C-V2X small sample radio frequency fingerprint classification method as described in any one of claims 1 to 8.

Images