Radio frequency fingerprinting method and system based on spectrum leakage attention enhancement

Abstract

The application discloses a radio frequency fingerprint identification method and system based on spectrum leakage attention enhancement, first, the received wireless baseband or radio frequency signal is pretreated and converted to a resource grid in the time-frequency domain; then, the resource grid is denoised, and a hardware damage leakage attention enhancement step is added, according to the radio frequency fingerprint theoretical model, the key spectrum leakage subcarriers related to the hardware non-ideal are taken as the key positions of the attention enhancement, and the attention weight is distributed according to the correlation between the source subcarriers and the target leakage subcarriers; finally, the radio frequency fingerprint features are extracted from the denoised signal, and the device classification identification is completed. The method can make the denoising process face the leakage position for directional enhancement and fidelity, while suppressing noise, the radio frequency fingerprint distinguishing features are maximally retained, and the robustness and reliability of the device identification under the low SNR condition are improved.

Description

Technical Field

[0001] This invention belongs to the technical field of wireless communication and information security, and relates to a radio frequency fingerprinting method that is robust to noise in low signal-to-noise ratio scenarios. Specifically, it relates to a radio frequency fingerprinting method and system based on spectrum leakage attention enhancement. Background Technology

[0002] With the widespread adoption of 5G New Radio (NR) in diverse scenarios such as the Internet of Things (IoT), drone communications, satellite internet, and private networks, air interface access faces increasingly complex open wireless environments. These scenarios involve a massive number of terminals, frequent access, and significant differences in link conditions, and are susceptible to security threats such as spoofed terminals, unauthorized access, replay attacks, and interference. Traditional authentication mechanisms based on upper-layer keys or certificates typically require long interaction latency and additional protocol overhead, often proving difficult to balance efficiency and reliability in low-power IoT terminals, low-latency control links, and weakly covered cell edge scenarios. Therefore, technologies capable of rapidly identifying and continuously monitoring device identities at the physical layer have become a crucial direction for improving air interface access security.

[0003] Radio frequency fingerprinting (RFF) leverages the unavoidable non-ideal characteristics resulting from transmitter hardware manufacturing errors and component aging, causing the transmitted signal to exhibit device-specific slight distortions in the frequency, time, or time-frequency domains. Compared to logical identifiers that rely solely on MAC addresses or IMSIs, which can be tampered with, RFF is a "hardware-inherent" feature, theoretically more difficult to completely replicate. Furthermore, it can be extracted directly from the received signal during initial access or service transmission, enabling seamless or low-interaction device identification and authentication. In communication systems such as 5G NR, these hardware non-idealities can lead to spectral leakage phenomena with certain structural patterns.

[0004] However, the discriminative information of RFF (Resonant Frequency Filter) is usually weak and easily overwhelmed by noise, fading, multipath interference, and co-channel interference. Especially under conditions such as cell edges, satellite links, high-speed movement of drones, or low-power IoT terminals, the received signal is often in a low signal-to-noise ratio (SNR) state, leading to a decrease in the separability of fingerprint features and a significant deterioration in recognition accuracy. To improve recognition performance under low SNR, existing solutions often adopt two approaches: one is to design a stronger feature extraction and classification network, hoping to learn discriminative features directly from noisy observations; the other is to use denoising as a preprocessing step, recovering the signal before recognition. However, the former is prone to training instability and insufficient generalization when the SNR is extremely low; the latter, if using general filtering or a denoising network that only pursues reconstruction quality, may smooth out the already subtle hardware distortions while suppressing noise, resulting in over-denoising and fingerprint blurring, making it even more difficult for the features obtained by the subsequent classifier to distinguish device differences.

[0005] In recent years, generative denoising methods such as diffusion models have demonstrated outstanding performance in image and signal restoration. Through probabilistic modeling of progressive noise addition and denoising, they can achieve high-quality reconstruction under strong noise conditions. Introducing diffusion models into wireless signal denoising is expected to significantly improve recovery capabilities at low SNR. However, existing diffusion model denoising networks typically rely on local correlation and overall reconstruction metrics, failing to explicitly protect and enhance the structured leakage locations of RFF (Radio Frequency Fingerprint). They may still treat device-related fine-grained waveform distortion differences as noise suppression, thus weakening recognition and discrimination capabilities. Therefore, there is an urgent need for a noise-robust device identification method that addresses the physical mechanism of RF fingerprints and preserves hardware damage leakage characteristics during denoising. This method should balance noise suppression and fingerprint separability under low SNR conditions to meet the requirements for fast and reliable device identification in 5G NR and its extended scenarios. In addition, real-world systems also suffer from factors such as channel frequency selectivity and receiver hardware non-ideal conditions, which cause the observation distribution of the same device to drift at different times and in different scenarios, placing higher demands on the robustness of the model. From an engineering implementation perspective, the method should also have deployability and adaptability to waveform parameters of different physical layers. Summary of the Invention

[0006] This invention addresses the problem of reduced device identification accuracy in low signal-to-noise ratio (SNR) wireless access scenarios, where radio frequency (RF) fingerprints, as weak discriminative features introduced by transmitter hardware non-ideals, are easily interfered with by noise. It also addresses the issue that existing denoising methods often prioritize overall reconstruction quality, failing to explicitly utilize the structured spectral leakage patterns caused by hardware impairments. This leads to excessive suppression of key leakage components such as image leakage, center leakage, and adjacent subcarrier leakage during denoising, resulting in blurred RF fingerprint features and loss of discriminative information. The invention proposes an RF fingerprint identification method and system based on spectral leakage attention enhancement. First, the received wireless baseband or RF signal is preprocessed and converted to a time-frequency domain resource grid. Then, denoising is performed on the resource grid, and a hardware impairment leakage attention enhancement step is added. Based on the RF fingerprint theoretical model, key spectral leakage subcarriers related to hardware non-ideals are designated as key locations for attention enhancement, and attention weights are allocated according to the correlation between the source and target leakage subcarriers. Finally, RF fingerprint features are extracted from the denoised signal to complete device classification and identification. The method of this invention enables the denoising process to perform targeted enhancement and fidelity preservation at the leakage location, thereby suppressing noise while preserving the radio frequency fingerprint discrimination features to the maximum extent, and improving the robustness and reliability of device identification under low SNR conditions.

[0007] To achieve the above objectives, the technical solution adopted by this invention is: a radio frequency fingerprinting method based on spectral leakage attention enhancement, comprising the following steps:

[0008] S1. Signal preprocessing and time-frequency representation construction: Receive the wireless baseband or radio frequency signal sent by the device to be identified, preprocess the signal, and transform it into a resource grid in the time-frequency domain; the preprocessing includes at least one of frame synchronization, frequency offset coarse compensation, and cyclic prefix removal;

[0009] S2. Denoising Processing: Denoising processing is performed on the resource raster obtained in step S1. The denoising processing is based on a denoising network model. The noise sample and its corresponding denoising control parameters are used as inputs to train the denoising network model to output noise estimation and / or clean signal estimation. The model is then optimized by the supervised learning objective of reconstruction error or noise estimation error to obtain the optimal denoising network model. The number of denoising steps is then adaptively selected according to the estimated SNR, peak signal-to-noise ratio index or validation set identification index of the input signal until the denoised output is obtained.

[0010] S3. Hardware Impairment Leakage Attention Enhancement: Determine an attention location set based on the spectral leakage subcarrier positions corresponding to hardware non-idealities, and calculate attention weights at the attention location set; the attention weights are allocated according to the correlation between the source subcarrier and the target leakage subcarrier, guiding the denoising process to perform targeted enhancement and fidelity preservation at the leakage locations;

[0011] S4. Radio Frequency Fingerprint Feature Extraction and Recognition Output: Extract and classify radio frequency fingerprint features from the denoised output, and output device identity tags or authentication results.

[0012] As an improvement to the present invention, the frame synchronization in step S1 preprocessing specifically includes:

[0013]

[0014]

[0015]

[0016] in, To receive discrete baseband signals, Indicates conjugate. To repeat the length of the leading half, For the starting index of the sliding related points, For relevant quantities, For energy terms, For the estimated frame start or symbol boundary position, This is the sequence index number;

[0017] The frequency offset coarse compensation specifically refers to:

[0018]

[0019] in, Indicates phase, To normalize frequency bias, For FFT points, This is the signal after frequency offset compensation; The imaginary unit, This is the sequence index number;

[0020] The removal of the cyclic prefix specifically refers to:

[0021]

[0022] in, For FFT points, The length of the cyclic prefix. For orthogonal frequency division multiplexing modulation symbol index, The first one after removing the cycle prefix One OFDM symbol valid sampled sequence, This is the sequence index number.

[0023] As an improvement of the present invention, after obtaining the resource grid in the time-frequency domain representation in step S1, at least one of amplitude normalization, energy normalization, or logarithmic power spectrum transformation is performed on the resource grid.

[0024] As another improvement to the present invention, in the denoising network model of step S2, samples with different noise intensities are obtained by gradually adding noise in the diffusion implementation, so as to... Input to train the noisy network model By minimizing the reconstruction error through supervision, the loss achieved through diffusion is specifically as follows:

[0025]

[0026] in, For loss function, For mathematical expectation, For the number of diffusion steps, For clean samples, The actual noise added. For parameters The noise in the network estimation.

[0027] As another improvement of the present invention, the adaptive selection method for the number of denoising steps in step S2 is specifically as follows:

[0028]

[0029] in, To adaptively select the number of denoising steps, The SNR estimate of the input signal. and These represent the minimum and maximum allowed number of denoising steps, respectively. and This represents the upper and lower bounds of the SNR mapping interval. This is for rounding up.

[0030] As another improvement of the present invention, step S3 specifically includes the following steps:

[0031] S31. Determination of the set of attention locations: Based on the radio frequency fingerprint theoretical model, a set of leakage locations related to hardware non-ideals is determined, and the set of leakage locations is used as the attention application locations; the leakage locations include one or more of the following: mirror leakage locations, spectrum center leakage locations, and adjacent subcarrier leakage locations;

[0032] S32. Valid attention region limitation: The valid region for attention calculation is limited by an indicator function; the indicator function includes a mirror leakage region indicator function, a central leakage region indicator function, and / or an adjacent leakage region indicator function;

[0033] S33. Correlation weight allocation: Within the effective attention area obtained in step S32, attention weights are allocated based on the similarity or statistical correlation between the feature vectors of the source subcarrier and the target leaked subcarrier. The weights are then normalized to achieve targeted enhancement of the radio frequency fingerprint features at the leak location.

[0034] S34. Multi-head attention and fusion: Multi-head structures are used to model leakage features of different leakage types or different frequency domain scales, and the outputs of each attention head are weighted and fused using learnable fusion coefficients to obtain an enhanced representation.

[0035] As another improvement of the present invention, in step S31, the mirror leakage specifically refers to:

[0036] ,

[0037] in, Indicates the leaked component in the image. This represents the frequency domain symbol of an ideal baseband signal on a discrete subcarrier. This is the IQ imbalance coefficient. As the IQ amplitude imbalance factor, IQ phase imbalance factor; The number of FFT points; The sequence number for activating subcarriers; It is the imaginary unit.

[0038] The specific details of the spectrum center leakage are:

[0039]

[0040] in, Indicates the center leakage component of the spectrum. This is the DC bias coefficient. The Dirac impulse function;

[0041] Neighboring subcarrier leakage specifically refers to:

[0042]

[0043] in, Indicates the first The component formed by leakage from adjacent subcarriers at each subcarrier. For the source subcarrier index that participates in the leakage overlay, To activate the subcarrier index set; This refers to the leakage convolution kernel or interference coefficient caused by carrier frequency offset; The sequence number for activating the subcarrier.

[0044] As another improvement of the present invention, the indicator function of step S32 is specifically as follows:

[0045] ;

[0046] The specific areas where the mirror image was leaked are:

[0047]

[0048] in, For mirror position, For window width, This indicates the current subcarrier position. The location of the target subcarrier. It is the FFT point count;

[0049] The central leakage area is specifically as follows:

[0050]

[0051] in, Central Index For window width, This indicates the current subcarrier position.

[0052] The adjacent leakage areas are specifically:

[0053]

[0054] in, For window width, This indicates the current subcarrier position. The location of the target subcarrier. As a further improvement of the present invention, in step S34, an attention head is configured for each type of leakage mode to obtain the first... The height in Output at each subcarrier :

[0055]

[0056] in, For value vectors, For attention head The weight, Let the activation matrix corresponding to attention head k satisfy the indicator function. Current subcarrier With target subcarrier The set that constitutes;

[0057] The outputs of each head are fused according to the learnable fusion coefficient. Fusion, resulting in attention-enhanced representation :

[0058] .

[0059] To achieve the above objectives, the present invention also adopts the following technical solution: a radio frequency fingerprint recognition system based on spectrum leakage attention enhancement, comprising a computer program, wherein the computer program, when executed by a processor, implements the steps of any of the methods described above.

[0060] Compared with existing technologies, the present invention offers the following advantages: It provides a radio frequency fingerprinting method based on spectrum leakage attention enhancement. This method uses an iterative learning denoising network to denoise resource grids, effectively suppressing noise and improving signal availability under low signal-to-noise ratio (SNR) conditions. To avoid overly smooth fingerprints caused by general denoising, the present invention embeds a hardware impairment leakage attention step into the denoising network. It determines the attention location set based on the spectrum leakage locations corresponding to hardware non-ideals and calculates and assigns attention weights at these leakage locations. This allows the denoising update process to directionally enhance device-related structured leakage components such as mirror leakage, center leakage, and adjacent leakage, thereby improving the separability and recognition accuracy of radio frequency fingerprint features. Compared to existing methods that only emphasize general denoising and reconstruction quality, this invention places greater emphasis on the protection and enhancement of discriminative features for frequency domain leakage modes. This significantly improves recognition accuracy and robustness under low SNR conditions and is compatible with various communication standards and classifiers, facilitating engineering deployment. It is suitable for physical layer device identification and access security for various standards such as 5G NR, LTE, and WiFi. Attached Figure Description

[0061] Figure 1This is a schematic diagram of the method steps of the present invention. Detailed Implementation

[0062] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0063] Example 1

[0064] A radio frequency fingerprinting method based on spectral leakage attention enhancement, such as Figure 1 As shown, it includes the following stages:

[0065] Step S1, Signal Preprocessing and Time-Frequency Representation Construction: Receive the wireless baseband or radio frequency signal sent by the device to be identified, and transform the signal into a resource grid in the time-frequency domain after preprocessing.

[0066] S11. Wireless signal reception: Receives wireless baseband signals or radio frequency sampling signals sent by the device to be identified. In 5GNR scenarios, it can be uplink PUSCH or other multi-carrier signal samples containing pilots or known structures.

[0067] S12. Synchronization and Preprocessing: Perform at least one or more of the following on the signal received in step S12: frame synchronization, coarse frequency offset compensation, and cyclic prefix removal, to obtain the subcarrier domain signal and reduce the influence of non-target factors; wherein, frame synchronization specifically includes:

[0068]

[0069] in, To receive discrete baseband signals, Indicates conjugate. To repeat the length of the leading half, For the starting index of the sliding related points, For relevant quantities, For energy terms, The estimated frame start or symbol boundary location;

[0070] The coarse frequency offset compensation is as follows:

[0071]

[0072] in, Indicates phase, To normalize frequency bias, For FFT points, The signal after frequency offset compensation

[0073] Removing the cyclic prefix specifically involves:

[0074]

[0075] in, The length of the cyclic prefix. For Orthogonal Frequency Division Multiplexing (OFDM) symbol index, The first one after removing the cycle prefix One OFDM symbol valid sampling sequence.

[0076] S13, Time-frequency domain conversion: The preprocessed signal is converted into a two-dimensional resource raster in the time-frequency domain as the input for subsequent denoising.

[0077] The preprocessed signal is then transformed into a two-dimensional resource raster in the frequency domain using an FFT transform:

[0078]

[0079] Will Index by subcarrier With symbol index Stacking forms a two-dimensional resource grid The total number of subcarriers is Select OFDM symbols.

[0080] S14. Normalization processing: Perform amplitude normalization on the resource grid to reduce the impact of path loss and channel gain changes on subsequent processing.

[0081]

[0082] This is the normalized resource raster.

[0083] Step 2, Denoising Processing: Denoising processing is performed on the resource grid. By designing a denoising network, signal denoising is achieved to improve the signal-to-noise ratio.

[0084] First, the denoising strategy parameters are set. This involves setting the number of denoising steps, noise intensity parameters, or noise scheduling parameters, and obtaining the corresponding denoising control parameters. In the diffusion implementation, the diffusion step is set to... Noise scheduling is control.

[0085] Then, training samples are constructed. During the training phase, noise is injected into the clean signal of the training samples according to the denoising control parameters to generate noise samples with different noise intensities, which are used to train the denoising network.

[0086] During the training phase, samples with different noise intensities are constructed. In the diffusion implementation, this is achieved by gradually adding noise.

[0087] , .

[0088] in, For a clean signal carrying an RF fingerprint, i.e., under the condition of high signal-to-noise ratio (SNR) = 40 dB. . The random noise added when constructing the training samples follows a standard normal distribution.

[0089] Subsequently, a denoising network model is trained. The noise samples and their corresponding denoising control parameters are used as inputs to train the denoising network to output noise estimates and / or clean signal estimates. The model is then optimized by a supervised learning objective of reconstruction error or noise estimation error.

[0090] by Input to train the noise prediction network The reconstruction error is minimized by supervising the objective. The loss of the diffusion-based implementation can be:

[0091]

[0092] During inference, iterative denoising is performed on the received noisy resource grid. In each iteration, the noise is estimated and the current signal is updated using a denoising network until a denoised output is obtained. During the inference phase, the noisy resource grid is iteratively updated, and denoising can be performed by inferring the actual added noise from the backpropagation model. Predict noise and calculate :

[0093]

[0094] Repeat the iteration until the denoised output is restored. .

[0095] Finally, the number of denoising steps is adaptively selected based on the estimated SNR of the input signal, the peak signal-to-noise ratio index, or the validation set identification index, in order to balance noise suppression and RF fingerprint feature preservation.

[0096] The number of denoising steps is adaptively selected based on the input estimated SNR or the validation set recognition metric:

[0097]

[0098] in, The SNR estimate of the input signal. and These represent the minimum and maximum allowed number of denoising steps, respectively. and This represents the upper and lower bounds of the SNR mapping interval. This is for rounding up.

[0099] Step 3, Hardware Impairment Leakage Attention Enhancement: A hardware impairment leakage attention module is embedded in the denoising network. An attention location set is determined based on the spectral leakage subcarrier positions corresponding to hardware non-idealities, and attention weights are calculated at the attention location set. The attention weights are allocated according to the correlation between the source subcarrier and the target leakage subcarrier, and are used to guide the denoising process to perform targeted enhancement and fidelity preservation at the leakage locations to retain the radio frequency fingerprint discrimination features.

[0100] S31. Determination of attention location set: Based on the radio frequency fingerprint theoretical model, determine the set of leakage locations related to hardware non-ideality, and use the set of leakage locations as the attention application locations; the leakage locations include one or more of the following: mirror leakage location, spectrum center leakage location, and adjacent subcarrier leakage location.

[0101] Based on the RF fingerprinting theoretical model, three types of structured leakage modes related to hardware non-ideals are identified, and an attention location set is formed accordingly. The three typical hardware impairments and their leakage modes are: DC bias, IQ imbalance, and carrier offset, corresponding to center leakage, image leakage, and adjacent leakage, respectively.

[0102] Mirror leak:

[0103] ,

[0104] in, Indicates the leaked component in the image. This represents the frequency domain symbol of an ideal baseband signal on a discrete subcarrier. This is the IQ imbalance coefficient. As the IQ amplitude imbalance factor, This is the IQ phase imbalance factor.

[0105] Center leak:

[0106]

[0107] in, Indicates the center leakage component of the spectrum. This is the DC bias coefficient, which reflects the magnitude and phase of the DC components of the transmitter's I / Q channels at zero frequency; is the Dirac impulse function, used to represent the leakage concentration at zero frequency, i.e., the center of the spectrum.

[0108] Adjacent leakage:

[0109]

[0110] in, Indicates the first The component formed by leakage from adjacent subcarriers at each subcarrier. For the source subcarrier index that participates in the leakage overlay, To activate the subcarrier index set; The leakage convolution kernel or interference coefficient caused by carrier frequency offset describes the leakage strength of the source subcarrier to the target subcarrier.

[0111] S32. Effective attention region limitation: The effective region for attention calculation is limited by an indicator function, so that attention is calculated only in the region corresponding to the set of leakage locations or non-leakage locations are suppressed; the indicator function includes a mirror leakage region indicator function, a central leakage region indicator function and / or an adjacent leakage region indicator function.

[0112] To avoid overfitting caused by attention spreading to irrelevant regions, an indicator function is used to limit attention to areas of leakage. The indicator function is defined as:

[0113]

[0114] Three types of leakage region indicator functions are constructed respectively.

[0115] Mirror leaked areas:

[0116]

[0117] in, For mirror position, This is the window width.

[0118] Central leak area:

[0119]

[0120] in, Central Index This is the window width.

[0121] Adjacent leak areas:

[0122]

[0123] in, This is the window width.

[0124] Furthermore, for each type of leakage k, the source subcarrier is defined. Valid target set:

[0125] .

[0126] S33. Correlation weight allocation: Within the effective attention area, attention weights are allocated based on the similarity or statistical correlation between the feature vectors of the source subcarrier and the target leaked subcarrier, and the weights are normalized to achieve targeted enhancement of the radio frequency fingerprint features at the leak location.

[0127] In the valid set Internally, attention weights are assigned based on the correlation between the source subcarrier signal and the target leaked subcarrier signal. This embodiment employs dot product attention, calculating the correlation between the subcarrier signal and the target leaked subcarrier signal through vector projection, and then generating the final attention weights through summation and softmax normalization.

[0128]

[0129] For attention weights, For the first Subcarrier feature vectors The target subcarrier key vector. The dimension is the subcarrier vector.

[0130] S34. Multi-head attention and fusion: Multi-head structures are used to model leakage features of different leakage types or different frequency domain scales, and the outputs of each attention head are weighted and fused using learnable fusion coefficients to obtain an enhanced representation.

[0131] Configure an attention head for each type of leakage pattern to obtain the first... The height in Output at each subcarrier :

[0132]

[0133] in, It is a value vector.

[0134] The outputs of each head are fused according to the learnable fusion coefficient. Fusion, resulting in attention-enhanced representation :

[0135] .

[0136] The contents of steps S31-S34 above are embedded into one or more layers of the denoising network model to directionally preserve the leakage structure associated with the RF fingerprint on a multi-scale representation. For example, using UNet as the backbone of the noise prediction network, the hardware impairment leakage attention enhancement of step S3 is inserted between the convolutional layers.

[0137] Step 4: Radio Frequency Fingerprint Feature Extraction and Recognition Output: Extract and classify radio frequency fingerprint features from the denoised output, and output the device identity label or authentication result.

[0138] The denoised resource raster is input into the feature extraction network, which automatically learns to distinguish the hardware damage features of different devices. CNN / ResNet / Transformer and other structures can be used as classifiers. The classifier outputs the predicted device identity label to complete device identification.

[0139] In summary, the method of this invention is applicable to multi-carrier wireless access scenarios under low signal-to-noise ratio (SNR) conditions. Addressing the issues of RF fingerprints being easily submerged in strong noise and the tendency of general denoising processing to cause excessive smoothing of fingerprint features, this invention proposes an RF fingerprint recognition method based on attention enhancement of spectrum leakage. Unlike existing general denoising preprocessing that only pursues reconstruction quality, this invention uses an attention-enhanced denoising recognition framework guided by structured spectrum leakage patterns caused by hardware impairments. It performs targeted fidelity preservation and enhancement on the interpretable structural features of frequency domain leakage patterns, significantly improving recognition accuracy and robustness under low SNR conditions. It is applicable to physical layer device identification and access security for various standards such as 5G NR, LTE, and WiFi.

[0140] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.

Claims

1. A radio frequency fingerprinting method based on spectral leakage attention enhancement, characterized in that, Includes the following steps: S1. Signal preprocessing and time-frequency representation construction: Receive the wireless baseband or radio frequency signal sent by the device to be identified, preprocess the signal, and transform it into a resource grid in the time-frequency domain; the preprocessing includes at least one of frame synchronization, frequency offset coarse compensation, and cyclic prefix removal; S2. Denoising Processing: Denoising processing is performed on the resource raster obtained in step S1. The denoising processing is based on a denoising network model. The noise sample and its corresponding denoising control parameters are used as inputs to train the denoising network model to output noise estimation and / or clean signal estimation. The model is then optimized by the supervised learning objective of reconstruction error or noise estimation error to obtain the optimal denoising network model. The number of denoising steps is then adaptively selected according to the estimated SNR, peak signal-to-noise ratio index or validation set identification index of the input signal until the denoised output is obtained. S3. Hardware Impairment Leakage Attention Enhancement: Determine the attention location set based on the spectral leakage subcarrier positions corresponding to hardware non-idealities, and calculate the attention weight at the attention location set; The attention weights are assigned based on the correlation between the source subcarrier and the target leaked subcarrier, guiding the denoising process to perform targeted enhancement and fidelity preservation at the leak location; S4. Radio Frequency Fingerprint Feature Extraction and Recognition Output: Extract and classify radio frequency fingerprint features from the denoised output, and output device identity tags or authentication results.

2. The radio frequency fingerprinting method based on spectral leakage attention enhancement as described in claim 1, characterized in that: The frame synchronization in step S1 preprocessing specifically involves: ； in, To receive discrete baseband signals, Indicates conjugate. To repeat the length of the leading half, For the starting index of the sliding related points, For relevant quantities, For energy terms, For the estimated frame start or symbol boundary position, This is the sequence index number; The frequency offset coarse compensation specifically refers to: ； in, Indicates phase, To normalize frequency bias, For FFT points, This is the signal after frequency offset compensation; The imaginary unit; The removal of the cyclic prefix specifically refers to: ； in, The length of the cyclic prefix. For orthogonal frequency division multiplexing modulation symbol index, The first one after removing the cycle prefix One OFDM symbol valid sampling sequence.

3. The radio frequency fingerprinting method based on spectral leakage attention enhancement as described in claim 2, characterized in that: After obtaining the resource grid in the time-frequency domain representation in step S1, at least one of amplitude normalization, energy normalization, or logarithmic power spectrum transformation is performed on the resource grid.

4. The radio frequency fingerprinting method based on spectral leakage attention enhancement as described in claim 1, characterized in that: In the denoising network model of step S2, samples with different noise intensities are obtained by gradually adding noise in the diffusion implementation. Input to train the noisy network model The loss function is achieved by minimizing the reconstruction error through a supervised objective and its diffusion-based implementation. Specifically: ； in, For mathematical expectation, For the number of diffusion steps, For clean samples, The actual noise added. For parameters The noise in the network estimation.

5. The radio frequency fingerprinting method based on spectral leakage attention enhancement as described in claim 4, characterized in that: In step S2, the adaptive selection method for the number of denoising steps is as follows: ； in, To adaptively select the number of denoising steps, The SNR estimate of the input signal. and These are the minimum and maximum allowed number of denoising steps, respectively. and This represents the upper and lower bounds of the SNR mapping interval. This is for rounding up.

6. The radio frequency fingerprinting method based on spectral leakage attention enhancement as described in claim 1, characterized in that: Step S3 specifically includes the following steps: S31. Determination of the set of attention locations: Based on the radio frequency fingerprint theoretical model, a set of leakage locations related to hardware non-ideals is determined, and the set of leakage locations is used as the attention application locations; the leakage locations include one or more of the following: mirror leakage locations, spectrum center leakage locations, and adjacent subcarrier leakage locations; S32. Valid attention region limitation: The valid region for attention calculation is limited by an indicator function; the indicator function includes a mirror leakage region indicator function, a central leakage region indicator function, and / or an adjacent leakage region indicator function; S33. Correlation weight allocation: Within the effective attention area obtained in step S32, attention weights are allocated based on the similarity or statistical correlation between the feature vectors of the source subcarrier and the target leaked subcarrier. The weights are then normalized to achieve targeted enhancement of the radio frequency fingerprint features at the leak location. S34. Multi-head attention and fusion: Multi-head structures are used to model leakage features of different leakage types or different frequency domain scales, and the outputs of each attention head are weighted and fused using learnable fusion coefficients to obtain an enhanced representation.

7. The radio frequency fingerprinting method based on spectral leakage attention enhancement as described in claim 6, characterized in that: In step S31, the mirror leakage specifically refers to: , ； in, Indicates the leaked component in the image. This represents the frequency domain symbol of an ideal baseband signal on a discrete subcarrier. This is the IQ imbalance coefficient. As the IQ amplitude imbalance factor, IQ phase imbalance factor; For FFT points, To activate the subcarrier sequence number, The imaginary unit; The specific details of the spectrum center leakage are: ； in, Indicates the center leakage component of the spectrum. This is the DC bias coefficient. The Dirac impulse function; Neighboring subcarrier leakage specifically refers to: ； in, Indicates the first The component formed by leakage from adjacent subcarriers at each subcarrier. For the source subcarrier index that participates in the leakage overlay, To activate the subcarrier index set; This refers to the leakage convolution kernel or interference coefficient caused by carrier frequency offset.

8. The radio frequency fingerprinting method based on spectral leakage attention enhancement as described in claim 7, characterized in that: The indicator function in step S32 is specifically as follows: ； The specific areas where the mirror image was leaked are: ； in, For mirror position, For window width, This indicates the current subcarrier position. The location of the target subcarrier. It is the FFT point count; The central leakage area is specifically as follows: ； in, Central Index For window width, This indicates the current subcarrier position. The adjacent leakage areas are specifically: ； in, For window width, This indicates the current subcarrier position. This refers to the location of the target subcarrier.

9. The radio frequency fingerprinting method based on spectral leakage attention enhancement as described in claim 8, characterized in that: In step S34, an attention head is configured for each type of leakage mode to obtain the first... The height in Output at each subcarrier : ； in, For value vectors, For attention head The weight, Let k be the activation matrix corresponding to the attention head. The outputs of each head are fused according to the learnable fusion coefficient. Fusion, resulting in attention-enhanced representation : 。 10. A radio frequency fingerprint recognition system based on spectral leakage attention enhancement, comprising a computer program, characterized in that: When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-9 above.

Images