A multi-modal radio frequency authentication method based on multi-scale signal representation

Abstract

The application discloses a multi-modal radio frequency authentication method based on a multi-scale signal representation, which comprises the following steps: firstly, pre-processing the original IQ signal of a target device received to obtain an instantaneous envelope signal; carrying out multi-scale decomposition and denoising processing on the instantaneous envelope signal to obtain a denoised envelope signal; constructing a multi-modal data set according to the original IQ signal and the denoised envelope signal; carrying out feature extraction and fusion on the multi-modal data set through a pre-trained target feature extraction network to obtain a fused feature representation; and carrying out classification on the fused feature representation through a pre-trained target classification network to output a classification result corresponding to the target device. The amplitude, frequency and time-frequency energy distribution information can be complementarily fused by constructing the multi-modal data set; the multi-modal features are extracted and fused through the target feature extraction network, so that the information loss is effectively avoided; and the target classification network is used for rapid classification, thereby reducing the calculation cost and ensuring the accuracy of identification.

Description

A multimodal radio frequency authentication method based on multi-scale signal representation Technical Field

[0001] This invention belongs to the field of wireless communication and information security technology, and specifically relates to a multi-mode radio frequency authentication method based on multi-scale signal representation. Background Technology

[0002] WiFi technology supports large-scale terminal interconnection in next-generation wireless networks and is an important component of the nation's new infrastructure construction. Benefiting from its advantages such as high bandwidth, low cost, ease of deployment, and wide coverage, WiFi technology has been widely applied in key areas such as the Industrial Internet, smart cities, connected vehicles, and public service networks. However, due to the openness and broadcast nature of wireless networks, WiFi signals face serious security challenges during transmission, including data leakage, spoofed access, and man-in-the-middle attacks.

[0003] Physical layer authentication (PLA) of WiFi signals is an effective means of ensuring wireless network security. PLA utilizes unique and unforgeable fingerprint features (such as amplitude, phase response, IQ imbalance, and power amplifier nonlinearity) formed in radio frequency signals due to hardware non-ideal characteristics to distinguish legitimate from illegitimate devices, thus improving wireless network security performance from the source. Traditionally, WiFi signal PLA methods directly use the raw received IQ signal or its spectral characteristics for modeling and classification. Because the raw IQ signal simultaneously carries strongly coupled device and channel characteristics, the extracted hardware fingerprint lacks stability under different times or scenarios, severely affecting the reliability and generalization performance of the authentication.

[0004] In recent years, physical layer authentication methods based on IQ signal processing have received widespread attention. Researchers have further obtained time-frequency features through IQ signal processing techniques such as Fourier transform, short-time Fourier transform, or wavelet transform, which characterize the hardware non-idealities of devices in both time and frequency domains, effectively suppressing noise interference and enhancing the robustness of features to channel changes. At the same time, by combining deep learning models such as convolutional neural networks for recognition, high-dimensional discriminative features can be automatically learned and complex nonlinear mapping relationships can be captured, which has stronger feature adaptation capabilities and recognition accuracy than traditional manual features or shallow classification models. These studies have demonstrated the effectiveness of physical layer authentication based on IQ signal processing in improving wireless security performance, but the following challenges still exist: (1) The signal representation capability of a single mode is limited, making it difficult to fully characterize the hardware differences of devices in different domains; (2) Feature distribution is still affected by channel changes and environmental disturbances, resulting in unstable performance across scenarios; (3) Traditional deep learning models mostly rely on cross-entropy loss for supervised training, making it difficult to simultaneously optimize intra-class feature compactness and inter-class feature separability, which limits generalization performance. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides a multimodal radio frequency authentication method based on multi-scale signal representation.

[0006] The technical problem to be solved by this invention is achieved through the following technical solution:

[0007] This invention provides a multimodal radio frequency authentication method based on multi-scale signal representation, comprising:

[0008] The received raw IQ signal from the target device is preprocessed to obtain the instantaneous envelope signal;

[0009] The instantaneous envelope signal is decomposed into multiple scales and denoised to obtain a denoised envelope signal.

[0010] A multimodal dataset is constructed based on the original IQ signal and the denoised envelope signal;

[0011] A pre-trained target feature extraction network is used to extract and fuse features from a multimodal dataset to obtain a fused feature representation.

[0012] The fused feature representation is classified by a pre-trained target classification network, and the classification result corresponding to the target device is output.

[0013] This invention provides a multimodal radio frequency authentication method based on multi-scale signal representation. By preprocessing the original IQ signal, the saliency of hardware defect features can be enhanced while reducing data dimensionality. By constructing a multimodal dataset, complementary fusion of amplitude, frequency, and time-frequency energy distribution information can be achieved. On this basis, multimodal features are extracted and fused through a target feature extraction network, effectively avoiding information loss. Furthermore, a target classification network is used for rapid classification, which reduces computational overhead while ensuring recognition accuracy, long-term stability, and environmental robustness.

[0014] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0015] Figure 1 is a flowchart illustrating a multi-modal radio frequency authentication method based on multi-scale signal representation provided in an embodiment of the present invention;

[0016] Figures 2A to 2C are schematic diagrams of simulation experimental results of a multi-modal radio frequency authentication method based on multi-scale signal representation provided by an embodiment of the present invention. Detailed Implementation

[0017] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0018] This invention provides a multimodal radio frequency authentication method based on multi-scale signal representation. Referring to Figure 1, the method includes the following steps:

[0019] S10. Preprocess the received raw IQ signal from the target device to obtain the instantaneous envelope signal.

[0020] For example, the target device can be any device capable of wireless network communication, and this embodiment does not limit this. IQ signal (In-phase and Quadrature Signal) is a core signal form in radio frequency (RF), communication, radar and other fields. It is essentially a pair of signals composed of in-phase component (I-path) and quadrature component (Q-path), with a phase difference of 90° between them, and together they carry the amplitude, frequency and phase information of the signal.

[0021] Optionally, step S10 may specifically include:

[0022] S101. Perform Hilbert transform on the received original IQ signal from the target device to obtain the initial transformed signal.

[0023] For example, the original IQ signal can be represented as:

[0024] ,

[0025] in, This is the starting index of the original IQ signal. The length of the original IQ signal.

[0026] For each discrete sample point of the original IQ signal The corresponding Hilbert transform result It can be represented as:

[0027]

[0028] in, Represents Cauchy's principal value integral. Let Hilbert transform function, For the index of cumulative calculation, .

[0029] S102. Construct a complex analytic signal based on the original IQ signal and the initial transformed signal.

[0030] For example, complex analytic signals Represented as:

[0031]

[0032] In the formula, It is the imaginary unit.

[0033] S103. Perform modulo operation on the complex analytic signal to obtain the instantaneous envelope signal.

[0034] For example, instantaneous envelope signal Represented as:

[0035] .

[0036] This embodiment maps the original two-dimensional IQ signal onto a one-dimensional envelope signal. This envelope signal reduces the data dimensionality while making the features introduced by hardware defects explicit, laying the foundation for subsequent feature extraction and analysis.

[0037] S20. Perform multi-scale decomposition and denoising on the instantaneous envelope signal to obtain a denoised envelope signal.

[0038] Optionally, step S20 may specifically include:

[0039] S201. The instantaneous envelope signal is decomposed into multiple scales using discrete wavelet transform to obtain the approximation coefficients and detail coefficients of the instantaneous envelope signal.

[0040] For example, applying discrete wavelet decomposition to the instantaneous envelope signal yields the instantaneous envelope signal at the 1st... Layer and time location The approximation coefficients and detail coefficients of each layer.

[0041] Optionally, the multi-scale decomposition of the instantaneous envelope signal can be expressed as:

[0042]

[0043] in, Indicates the instantaneous envelope signal. Let J be the scaling function of the J-th layer, i.e., the low-frequency basis function. These are the corresponding approximation coefficients; Let j be the wavelet function of the j-th layer, i.e., the high-frequency basis function. Here, J represents the corresponding detail coefficients; J is the maximum number of decomposition levels. This represents the discrete translation position of the wavelet function on the time axis, controlling the position of the analysis window, and... Together they determine the position of the wavelet function in the time domain.

[0044] J can be set This approach aims to fully capture the dynamic characteristics of WiFi signals at different time scales. Furthermore, the Daubechies-4 wavelet, with its favorable time-frequency localization properties, is preferentially used as the wavelet basis function.

[0045] S202. Threshold the detail coefficients to obtain the target detail coefficients.

[0046] For example, the detail factor for each layer Soft thresholding is used instead of hard thresholding because soft thresholding has a smooth transition at the coefficients, which can reduce artifacts after reconstruction and improve the continuity of the waveform after denoising, which is beneficial for downstream feature extraction and learning.

[0047] Optionally, the target detail coefficient is expressed as:

[0048]

[0049] in, For symbolic functions, For the threshold;

[0050]

[0051]

[0052] in, The total length of the instantaneous envelope signal sample. This is an estimate of the noise level. This represents the detail coefficients corresponding to the first layer.

[0053] S203. Perform inverse transformation and reconstruction based on the target detail coefficients and approximation coefficients to obtain the denoised envelope signal.

[0054] Optionally, the denoised envelope signal is represented as:

[0055]

[0056] in, This represents the denoised envelope signal.

[0057] This embodiment effectively improves the stability and robustness of device fingerprint information by suppressing high-frequency random noise and retaining low-frequency hardware features.

[0058] S30. Construct a multimodal dataset based on the original IQ signal and the denoised envelope signal.

[0059] Optionally, step S30 may specifically include: performing a short-time Fourier transform on the original IQ signal to obtain a spectrogram; performing a continuous wavelet transform on the original IQ signal to obtain a wavelet graph; and constructing a multimodal dataset from the original IQ signal, the denoised envelope signal, the spectrogram, and the wavelet graph.

[0060] For example, for each original IQ signal By deriving and supplementing, multiple complementary signal representations are generated, thus forming a multimodal dataset:

[0061] Wavelet-denoised envelope signal (denoted as...) The signal consists of the denoised envelope signal obtained in step 20 after Hilbert transform and wavelet denoising, and mainly reflects the amplitude modulation characteristics caused by carrier frequency offset and hardware defects.

[0062] Spectrogram (denoted as) The two-dimensional time-frequency distribution map is obtained by performing a short-time Fourier transform on the original IQ signal. It can effectively capture the dynamic change characteristics of the signal in the frequency domain.

[0063] Scalogram (notation) ): Obtained by performing continuous wavelet transform on the original IQ signal, it characterizes the energy distribution of the signal in the time-scale domain and is particularly suitable for reflecting the difference between instantaneous features and stationarity.

[0064] For the i-th raw IQ signal from the same target device, the following multimodal dataset is constructed. :

[0065] .

[0066] S40. The multimodal dataset is subjected to feature extraction and fusion through a pre-trained target feature extraction network to obtain a fused feature representation.

[0067] Optionally, the target feature extraction network includes a one-dimensional convolutional neural network, a two-dimensional convolutional neural network, a linear projection layer, and a feature fusion layer.

[0068] Step S40 may specifically include:

[0069] S401. Features are extracted from the original IQ signal and the denoised envelope signal in the multimodal dataset using a one-dimensional convolutional neural network, and then processed by a linear projection layer to obtain the corresponding first feature vector and second feature vector.

[0070] S402. Features are extracted from the spectrogram and wavelet graph in the multimodal dataset using a two-dimensional convolutional neural network, and then processed by a linear projection layer to obtain the corresponding third and fourth feature vectors.

[0071] S403. The first feature vector, the second feature vector, the third feature vector and the fourth feature vector are fused through the feature fusion layer to obtain the fused feature representation.

[0072] For example, independent feature encoding networks are designed for each modality to extract high-level semantic features from input data (multimodal datasets) of different modalities. Given the differences in data structure across modalities, convolutional neural networks (CNNs) are used as the primary feature extraction framework, differentiated according to the data's dimensionality: for time-series data (including original IQ signals and denoised envelope signals), one-dimensional convolutional neural networks (1D-CNNs) are used for feature extraction; for time-frequency image data (including spectrograms and wavelet graphs), two-dimensional convolutional neural networks (2D-CNNs) are used.

[0073] Each modality's feature encoding network consists of standard convolutional modules (convolutional layers, batch normalization layers, and non-linear activation functions) and adaptive average pooling layers. It can effectively extract hierarchical features from the input data and output a fixed-length (target length) feature vector. ,in .

[0074] To ensure that feature representations from different modalities have consistent dimensionality and are comparable, a linear projection layer is introduced at the end of each modal branch, and its calculation form is as follows:

[0075]

[0076] in, Indicates the first CNN feature extraction function corresponding to each modality and These are the weights and bias parameters of the linear projection layer corresponding to this mode.

[0077] After obtaining the feature vectors of each modality, they are fused into a unified representation to comprehensively characterize the device's feature information. This embodiment employs a multimodal feature fusion method based on vector concatenation. Finally, the feature vectors corresponding to the four modalities are concatenated as follows:

[0078]

[0079] in, This represents the fusion feature representation.

[0080] This embodiment can retain the independent information of each modality without loss, avoiding the information loss that may be caused by complex weighting or interaction, and has high performance.

[0081] Optionally, the training process of the target feature extraction network may specifically include:

[0082] The fused feature representation samples output by the target feature extraction network are input into a preset linear mapping network to obtain contrastive learning representation vector samples;

[0083] The target feature extraction network is trained based on the contrastive learning representation vector samples, the corresponding device category labels, and the target loss function.

[0084] For example, this embodiment introduces a supervised contrastive learning mechanism. Specifically, the fused feature representations of the samples... The samples are first fed into a pre-defined linear mapping network, which consists of several fully connected layers and non-linear activation functions, i.e., a multilayer perceptron (MLP). Through this pre-defined linear mapping network, feature representations of the samples are fused. The vectors are projected into a new embedding space to obtain the final contrastive learning representation vector samples used for contrastive learning. :

[0085]

[0086] in, This represents the mapping function of a predefined linear mapping network. express 3D real space.

[0087] In this embedding space, this embodiment uses a supervised contrastive loss function (SupConLoss) to optimize the network. Unlike unsupervised contrastive learning methods that only treat augmentations as positive samples, this embodiment utilizes known device category labels to treat all samples from the same category as a set of positive samples.

[0088] Optionally, for a single contrastive learning representation vector sample in a single training batch The corresponding loss function Defined as:

[0089]

[0090] in, Indicates that in a single training batch, excluding All outside and The set of positive samples with the same device category label. express The samples in This indicates that in a single training batch, excluding All other sample sets, express The samples in; Indicates the cosine similarity between samples; This is a temperature coefficient used to adjust the sharpness of the distribution. This indicates the sample index.

[0091] in, Represented as:

[0092]

[0093] in, This represents two different contrastive learning representation vector samples. The calculation method is the same, so it will not be repeated here.

[0094] The target loss function can be expressed as:

[0095]

[0096] in, Represents the target loss function. This represents the number of samples in a single training batch.

[0097] Once the training in the contrastive learning phase has converged, remove the projection head used in the contrastive learning step. The parameters of the target feature extraction network are frozen. The fused feature representation output by the target feature extraction network is directly used as the final representation of the device fingerprint.

[0098] This embodiment employs a supervised contrastive learning method to optimize intra-class feature compactness and inter-class feature separation. By minimizing the objective loss function, the model is motivated to maximize the similarity between the anchor point and its positive samples, while minimizing its similarity with all negative samples.

[0099] S50. Classify the fused feature representation using a pre-trained target classification network and output the classification result corresponding to the target device.

[0100] Optionally, the target classification network adopts a multilayer perceptron structure, consisting of a hidden layer and an output layer; and a batch normalization layer and a modified linear unit activation function are set between the hidden layer and the output layer.

[0101] The loss function used during the training of the object classification network , is represented as:

[0102]

[0103] in, Indicates the total number of equipment categories; The unique hot-coded representation for the device category label; For target classification networks, classify devices by category. The predicted probability.

[0104] For example, the target classification network is a lightweight classification network employing a multilayer perceptron (MLP) structure, consisting of one hidden layer and one output layer. The input fused feature representation is first mapped to a 256-dimensional hidden representation, and then further reduced to 128 dimensions to reduce computational complexity and enhance model generalization. To improve the stability of the model training process, a batch normalization layer and a rectified linear unit (ReLU) activation function are introduced between the hidden and output layers. Finally, the output layer maps the 128-dimensional feature vector to a logits representation corresponding to the number of device categories and calculates the predicted probability for each device category. Therefore, based on the classification results of the target device, the identity of legitimate and illegitimate devices can be distinguished, achieving device authentication.

[0105] Since the feature extraction capability of the target feature extraction network has been fully converged and frozen during the pre-training stage, the fine-tuning process of this target classification network can be completed quickly with low computational overhead, while ensuring high recognition accuracy and robustness.

[0106] This embodiment of the multimodal RF authentication method based on multi-scale signal representation preprocesses the original IQ signal using Hilbert transform and wavelet denoising, which enhances the saliency of hardware defect features while reducing data dimensionality. A multimodal dataset constructed by combining the original IQ signal, denoised envelope signal, spectrogram, and wavelet graph achieves complementary fusion of amplitude, frequency, and time-frequency energy distribution information. Furthermore, a differential convolutional neural network extracts multimodal features and performs linear projection stitching, effectively avoiding information loss. A supervised contrastive learning mechanism is further introduced, significantly improving the compactness of intra-class features and the separability of inter-class features, enabling the model to learn stable fingerprint features that are insensitive to time changes and robust across timeframes. Finally, a frozen feature extraction network and a lightweight classification network are used for rapid classification, reducing computational overhead while ensuring recognition accuracy, long-term stability, and environmental robustness.

[0107] The following simulation experiment further illustrates the multimodal radio frequency authentication method based on multi-scale signal representation provided by this invention.

[0108] This embodiment uses the large-scale public WiFi radio frequency fingerprint data proposed by Elmaghbub and Hamdaoui as the dataset, and selects a subset of data collected over three consecutive days (Day 1, Day 2, Day 3) in the "indoor wireless fixed location" scenario for training and evaluating the model. The experiments were conducted using Python 3.8 and the PyTorch 1.10.0 deep learning framework on an NVIDIA Tesla V100 GPU. To rigorously evaluate the model's generalization ability in time-varying environments, Day 1, Day 2, and Day 3 of the dataset were used, employing a "single-day training, two-day testing" cross-validation strategy. For example, the data from Day 1 was used as the training set, and testing was performed on the data from Day 2 and Day 3, and so on for other cases.

[0109] Table 1 details the classification accuracy of the multimodal RF authentication method based on multi-scale signal representation under three training-test combinations. The summarized data shows that when the training and test data come from the same day, the method achieves nearly 100% classification accuracy over all three days, indicating that these features contain sufficient distinguishable information under relatively stable channel conditions. In the more challenging cross-day testing scenario, the proposed method achieves high performance while using a low data dimensionality, with an input tensor size of (2, 1678). Through effective front-end signal processing, the method successfully refines and reduces the dimensionality of features, improving recognition robustness while reducing the complexity requirements of subsequent models, resulting in higher efficiency.

[0110] Table 1. Accuracy (%) of the method of the present invention in cross-day testing

[0111]

[0112] Figures 2A to 2C show the corresponding confusion matrices. Specifically, Figures 2A to 2C represent the confusion matrices of the denoised envelope signal of this invention. Figure 2A shows the training using the day1 dataset and the testing using the day2 and day3 datasets; Figure 2B shows the training using the day2 dataset and the testing using the day1 and day3 datasets; and Figure 2C shows the training using the day3 dataset and the testing using the day1 and day2 datasets.

[0113] As can be seen, the denoised envelope signal exhibits optimal performance and stability in most cross-day test scenarios. When trained using Day1 data, its test accuracy on Day2 and Day3 reached 97.32% and 99.31%, respectively, fully demonstrating that the method proposed in this invention can effectively suppress noise interference and extract hardware fingerprint features with greater time invariance.

[0114] In summary, the experimental results show that the method of the present invention can effectively overcome the challenges brought about by cross-day environmental changes while reducing data dimensionality and computational overhead, enhance the saliency of hardware fingerprint features, and achieve higher accuracy and stronger robustness in device identification.

[0115] It should be noted that the terms "first," "second," etc., are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention.

[0116] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0117] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings and the disclosure in carrying out the claimed invention. In the description of the invention, the word "comprising" does not exclude other components or steps, "a" or "an" does not exclude a plurality, and "a plurality" means two or more, unless otherwise explicitly specified. Furthermore, while different embodiments may describe certain measures, this does not mean that these measures cannot be combined to produce good results.

[0118] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A multimodal radio frequency authentication method based on multi-scale signal representation, characterized in that, include: The received raw IQ signal from the target device is preprocessed to obtain an instantaneous envelope signal; the instantaneous envelope signal is decomposed into approximation coefficients and detail coefficients using discrete wavelet transform to obtain these coefficients; the detail coefficients are thresholded to obtain target detail coefficients; and an inverse transform is performed based on the target detail coefficients and the approximation coefficients to reconstruct the denoised envelope signal. The multi-scale decomposition of the instantaneous envelope signal can be expressed as follows: in, Indicates the instantaneous envelope signal. Let J be the scaling function of the J-th layer, i.e., the low-frequency basis function. These are the corresponding approximation coefficients; Let j be the wavelet function of the j-th layer, i.e., the high-frequency basis function. Here, J represents the corresponding detail coefficients; J is the maximum number of decomposition levels. This represents the discrete translation position of the wavelet function on the time axis, controlling the position of the analysis window, and... Together, they determine the position of the wavelet function in the time domain; the target detail coefficients are expressed as: in, For symbolic functions, For the threshold; in, The total length of the sample of the instantaneous envelope signal is given. This is an estimate of the noise level. The first layer represents the detail coefficients; the denoised envelope signal is represented as: in, The original IQ signal and the denoised envelope signal are used to construct a multimodal dataset. Feature extraction and fusion of the multimodal dataset are performed using a pre-trained target feature extraction network to obtain a fused feature representation. The fused feature representation is then classified using a pre-trained target classification network to output the classification result corresponding to the target device.

2. The multimodal radio frequency authentication method based on multi-scale signal representation according to claim 1, characterized in that, The step of constructing a multimodal dataset based on the original IQ signal and the denoised envelope signal includes: performing a short-time Fourier transform on the original IQ signal to obtain a spectrogram; performing a continuous wavelet transform on the original IQ signal to obtain a wavelet graph; and constructing a multimodal dataset from the original IQ signal, the denoised envelope signal, the spectrogram, and the wavelet graph.

3. The multimodal radio frequency authentication method based on multi-scale signal representation according to claim 2, characterized in that, The target feature extraction network includes a one-dimensional convolutional neural network, a two-dimensional convolutional neural network, a linear projection layer, and a feature fusion layer. The step of using the pre-trained target feature extraction network to extract and fuse features from the multimodal dataset to obtain a fused feature representation includes: extracting features from the original IQ signal and the denoised envelope signal in the multimodal dataset using the one-dimensional convolutional neural network, and processing them through the linear projection layer to obtain corresponding first and second feature vectors; extracting features from the spectrogram and wavelet graph in the multimodal dataset using the two-dimensional convolutional neural network, and processing them through the linear projection layer to obtain corresponding third and fourth feature vectors; and fusing the first, second, third, and fourth feature vectors through the feature fusion layer to obtain a fused feature representation.

4. The multimodal radio frequency authentication method based on multi-scale signal representation according to claim 3, characterized in that, The training process of the target feature extraction network includes: inputting the fused feature representation samples output by the target feature extraction network into a preset linear mapping network to obtain contrastive learning representation vector samples; and training the target feature extraction network based on the contrastive learning representation vector samples, the corresponding device category labels, and the target loss function.

5. The multimodal radio frequency authentication method based on multi-scale signal representation according to claim 4, characterized in that, For a single contrastive learned representation vector sample in a single training batch The corresponding loss function Defined as: in, Indicates that in a single training batch, excluding All outside and The set of positive samples with the same device category label. express The samples in This indicates that in a single training batch, excluding All other sample sets, express The samples in; Indicates the cosine similarity between samples; This is a temperature coefficient used to adjust the sharpness of the distribution. The sample index is represented; the target loss function is expressed as: in, Represents the target loss function. This represents the number of samples in a single training batch.

6. The multimodal radio frequency authentication method based on multi-scale signal representation according to claim 1, characterized in that, The target classification network employs a multilayer perceptron structure, consisting of one hidden layer and one output layer; a batch normalization layer and a modified linear unit activation function are placed between the hidden layer and the output layer; the loss function used by the target classification network during training is... , is represented as: in, Indicates the total number of equipment categories; The unique hot-coded representation for the device category label; For target classification networks, classify devices by category. The predicted probability.

7. The multimodal radio frequency authentication method based on multi-scale signal representation according to claim 1, characterized in that, The step of preprocessing the received raw IQ signal from the target device to obtain an instantaneous envelope signal includes: performing a Hilbert transform on the received raw IQ signal from the target device to obtain an initial transformed signal; constructing a complex analytic signal based on the raw IQ signal and the initial transformed signal; and performing modulo processing on the complex analytic signal to obtain the instantaneous envelope signal.

Images