A cross-receiver radar emitter fingerprinting method and system based on receiver model and feature degradation isolation

Abstract

A cross-receiver radiation source fingerprinting method and system based on receiver model and feature degradation isolation belongs to the field of radiation source identification technology. It addresses the problems of current methods, such as fuzzy physical model establishment, inadequate domain adaptation methods, and frequent negative transfer and label mismatch in some cross-receiver scenarios. The method includes: establishing a receiver mathematical model and a cross-receiver data transfer model; deriving a feature extraction scheme resistant to cross-receiver distortion; extracting cross-receiver-insensitive features; constructing a domain-adaptive neural network based on the feature extraction scheme; using a pseudo-label method to isolate feature degradation and generate high-confidence pseudo-labels; constructing a main recognition network; using high-confidence pseudo-labels to guide the initialization of the main recognition network; and alternately training the main recognition network using CDAN and LMMD to achieve cross-receiver radiation source fingerprinting. This invention is applicable to cross-receiver radiation source fingerprinting scenarios.

Description

Technical Field

[0001] This invention belongs to the field of specific radiation source identification technology, specifically relating to a cross-receiver radiation source fingerprint identification technology based on receiver model and feature degradation isolation. Background Technology

[0002] With the rapid development of information technology, the types and numbers of wireless devices have increased dramatically, leading to a wider range of application demands for radiation source identification and authentication, while also bringing significant technical challenges. Specific Emitter Identification (SEI), also known as individual radiation source identification or radio frequency fingerprinting, is a technology that identifies different radiation sources by analyzing minute features in signals that are difficult to imitate and clone. It can effectively enhance the security of wireless networks, protect data privacy, and improve the anti-spoofing capabilities of key civilian fields such as mobile communication networks, the Internet of Things (IoT), and Automatic Dependent Surveillance-Broadcast (ADS-B) systems, while also enhancing signal and electromagnetic spectrum situational awareness capabilities. Therefore, individual radiation source identification is considered one of the key technologies for maintaining electromagnetic space security and has shown broad application prospects in related fields.

[0003] Radiation source fingerprinting has made significant progress, with greatly improved recognition rates and speeds, and reduced requirements for training sample sizes. However, these achievements are almost entirely rooted in laboratory environments. In real-world applications, radiation source signal characteristics are affected by various factors such as operating parameters, transmission and reception channels, environmental interference and noise, and receiver inconsistencies. Therefore, it is still some distance from being fully operational outside the laboratory. These problems can be categorized into three types: variations in the radiation source; variations in the channel; and receiver inconsistencies. Among these, receiver inconsistencies, or cross-receiver problems, are among the most impactful on practical applications. Essentially, the contamination features generated by the receiver and the required fingerprint features generated by the transmitter are the opposite of each other, resulting in a high degree of coupling that is difficult to remove using conventional methods. In practical applications, differences between the receiver used for training and the receiver used in actual deployment are unavoidable. Traditional SEI schemes cannot adequately adapt to varying radiation source characteristics and changing usage scenarios, thus severely limiting their reliability and practicality.

[0004] In the early stages of the cross-receiver radiation fingerprinting problem, methods existed for cross-receiver distortion modeling and traditional radiation source fingerprinting. However, with the development of neural network technology in the field of radiation source fingerprinting, current methods for specific cross-receiver problems almost entirely employ domain-adaptive neural networks. The application of domain-adaptive networks has evolved from basic Maximum Mean Discrepancy (MMD) models and Domain-Adversarial Neural Networks (DANNs) to neural networks based on Local Maximum Mean Discrepancy (LMMD), Conditional Adversarial Domain Adaptation (CDAN), and pseudo-labeling methods, achieving better recognition results in some cross-receiver environments.

[0005] In the field of domain adaptation, pseudo-labeling is often used to leverage the model's own predictions to generate reliable supervisory signals for unlabeled target domain data, thereby reducing the distributional differences between the source and target domains. However, the cumulative errors it causes make it difficult to directly apply to cross-receiver source fingerprinting. Specifically, if the initial model performs poorly in the target domain, the generated pseudo-labels will contain numerous errors. Training the model with these erroneous labels further degrades its performance, leading to even worse pseudo-labels and creating a vicious cycle. Poor initial models and frequent domain classification errors are quite common in cross-receiver source fingerprinting, primarily because cross-receiver source fingerprinting deviates from the conditional distribution assumptions of traditional domain adaptation.

[0006] Current cross-receiver radiation source fingerprinting methods, while employing the overall receiver domain adaptation theory, lack specific analysis of the cross-receiver problem, treating it merely as a domain adaptation problem. However, the cross-receiver problem differs from traditional domain adaptation problems. The receiver and transmitter features in a cross-receiver problem are essentially opposite processes originating from the same source. Therefore, the signal distortion caused by switching receivers not only involves the distribution differences of features between different domains assumed by domain adaptation but also feature contamination introduced by the cross-receiver problem. This means that transmitter features are covered by similar receiver features or suffer irreversible distortion. Although domain adaptation methods can achieve good target recognition results on the target receiver when the differences between some receivers are small, this approach always relies on receiver selection and cannot maintain model fit even after arbitrarily switching receivers. In experimental reproduction, this manifests as frequent tag mismatches and negative migrations during the mutual migration between certain source and target receivers. This phenomenon can be attributed to the problem not fully satisfying the domain adaptation assumption in the context of domain adaptation. Therefore, the cross-receiver source fingerprinting problem should be considered a problem situated at the boundary between domain adaptation and traditional model compensation. Neither traditional model methods nor domain adaptation methods can solve this problem. Consequently, current methods still suffer from ambiguity in physical model establishment, and domain adaptation methods are not well-suited to the problem. Negative migrations and tag mismatches frequently occur in some cross-receiver scenarios, severely hindering the development and application of cross-receiver source fingerprinting technology. Summary of the Invention

[0007] This invention provides a cross-receiver radiation source fingerprinting method and system based on receiver model and feature degradation isolation. The purpose is to solve the technical problems in cross-receiver radiation source fingerprinting, such as fuzzy physical model establishment, incompatibility between domain adaptation methods and cross-receiver problems, high coupling between receiver distortion features and transmitter fingerprint features leading to feature contamination and irreversible distortion, frequent negative migration and tag mismatch in cross-receiver scenarios, and feature extraction processing to resist receiver distortion causing feature degradation that is difficult to effectively isolate, ultimately resulting in insufficient reliability, practicality and accuracy of identification.

[0008] Firstly, the present invention aims to provide a cross-receiver radiation source fingerprint recognition method based on receiver model and feature degradation isolation, comprising the following steps: S1: Establish a receiver mathematical model and a cross-receiver data migration model. Based on the receiver mathematical model and the cross-receiver data migration model, derive a feature extraction scheme that resists cross-receiver distortion and extract cross-receiver insensitive features. S2: Construct a domain-adaptive neural network based on the feature extraction scheme of S1, and use the pseudo-label method to isolate feature degradation and generate high-confidence pseudo-labels; S3: Construct a master recognition network with alternating pseudo-label, CDAN, and LMMD guidance. Use the high-confidence pseudo-label generated in S2 to guide the initialization of the master recognition network. Train the master recognition network by alternating guidance of CDAN and LMMD. Use the trained network to achieve fingerprint recognition across receiver radiation sources.

[0009] Furthermore, a preferred solution is provided: the process of establishing the receiver mathematical model is as follows: for the signal reception distortion of the zero intermediate frequency receiver, the non-ideal characteristics of the filter and the nonlinear distortion of the amplifier are selected as distortion factors, and the receiver mathematical model is simplified into a Volterra simplified model that includes nonlinear distortion and memory effect based on the distortion factors.

[0010] Furthermore, a preferred solution is provided: the process of establishing the cross-receiver data migration model is as follows: the memory effect caused by the filter and other components is equivalent to a lower triangular Toeplitz matrix, the mutual conversion relationship of the signals acquired by receivers with different parameters is derived, and the cross-receiver migration model is obtained based on the conversion relationship, thereby realizing the mathematical description of the data migration process between different receivers.

[0011] Furthermore, a preferred solution is provided: the derivation process of the feature extraction scheme against cross-receiver distortion is as follows: the memory effect in the time domain of the cross-receiver migration model is transformed to the frequency domain, the influence of the memory effect is equivalent to frequency domain weighting, and the influence of frequency domain weighting is eliminated by time-frequency domain normalization; random noise is introduced in the normalization process to erase the memory effect-related features of the cross-receiver migration model, thereby obtaining the feature extraction scheme against cross-receiver distortion.

[0012] Furthermore, a preferred solution is provided: the construction process of the domain adaptive neural network includes: connecting the feature extraction scheme of S1 to the feature extraction layer of the CDAN domain adaptive neural network.

[0013] Furthermore, a preferred solution is provided: In S2, the process of generating high-confidence pseudo-labels using the pseudo-label method is as follows: the domain-adaptive neural network is used to predict unlabeled target domain data to generate initial pseudo-labels, and the initial pseudo-labels are filtered by a combination of threshold screening and multi-model voting to obtain high-confidence pseudo-labels.

[0014] Furthermore, a preferred embodiment is provided: In S3, the step of alternately guiding the training of the main recognition network includes: Three independent pseudo-label generation networks were trained using the feature extraction scheme described in S1; Three pseudo-labels are used to generate network voting and filter, resulting in high-confidence pseudo-labels for initializing the main recognition network. Using source domain labeled data and high-confidence pseudo-label data, the main recognition network is guided to complete initial training by initializing the total loss function; The weights of the CDAN and LMMD loss functions are alternately adjusted according to the iterative process. The main recognition network is trained until convergence through a fusion-type total loss function, thereby achieving fingerprint recognition across receiver radiation sources.

[0015] Secondly, the purpose of this invention is to propose a cross-receiver radiation source fingerprinting system based on receiver model and feature degradation isolation. The system is implemented based on a cross-receiver radiation source fingerprinting method based on receiver model and feature degradation isolation as described in any one or more of the above-mentioned schemes. The system includes: Feature extraction module: used to establish receiver mathematical model and cross-receiver data migration model, derive feature extraction scheme against cross-receiver distortion based on the receiver mathematical model and cross-receiver data migration model, and extract cross-receiver insensitive features; The pseudo-label generation module is used to construct a domain-adaptive neural network based on the feature extraction scheme, and uses the pseudo-label method to isolate feature degradation and generate high-confidence pseudo-labels. Network training module: used to construct a master recognition network guided by a triple alternation of pseudo-label, CDAN, and LMMD. The master recognition network is initialized using high-confidence pseudo-labels, trained by alternating CDAN and LMMD, and then used to achieve fingerprint recognition across receiver radiation sources.

[0016] Thirdly, the present invention aims to provide a computer device, the computer device including a memory and a processor, the memory storing a computer program, and when the processor runs the computer program stored in the memory, the processor executes a cross-receiver radiation source fingerprint recognition method based on receiver model and feature degradation isolation according to any one or more of the above-described schemes.

[0017] Fourthly, the present invention aims to provide a computer-readable storage medium for storing a computer program that executes the cross-receiver radiation source fingerprinting method based on receiver model and feature degradation isolation described in any one or more of the above-described schemes.

[0018] Compared with the prior art, the advantages of the present invention are: (1) This invention proposes a data migration model across receivers and a corresponding feature extraction method based thereon. The model analyzes data distortion across receivers from a perspective closer to the physical intrinsics, and the proposed feature extraction method can suppress the distortion.

[0019] (2) This invention proposes a degradation isolation method based on pseudo-labels, which can obtain high-confidence pseudo-labels that suppress negative migration and class mismatch phenomena. It can solve the problem that the common cross-receiver radiation source fingerprint recognition method based on domain adaptation cannot be applied to more receivers.

[0020] (3) This invention proposes a master recognition network based on pseudo-label method - CDAN-LMMD alternating guidance. Under the guidance of the generated high-confidence pseudo-labels that suppress negative migration and class mismatch, it can improve the generalization ability of the neural network by utilizing the strong adversarial nature of CDAN and the high stability of LMMD, thereby obtaining a better solution to the cross-receiver radiation source fingerprint recognition problem.

[0021] This invention is applicable to fingerprint recognition scenarios involving multiple receiver radiation sources. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram illustrating the relationship between the receiver model and the migration model according to a specific embodiment of the present invention; Figure 2 This is a flowchart of the feature extraction method based on cross-receiver data migration model data correction according to a specific embodiment of the present invention; Figure 3 This is a flowchart illustrating the pseudo-tag generation process according to a specific embodiment of the present invention; Figure 4 This is a schematic diagram of the main recognition network structure combining the pseudo-label method and the domain adaptation neural network as described in a specific embodiment of the present invention; Figure 5 This is a comparison chart showing the impact of the feature extraction method described in a specific embodiment of the present invention on the recognition accuracy. Figure 6 This is a comparison chart showing the impact of the feature extraction method on recognition accuracy when the receiver differences are small, as described in a specific embodiment of the present invention. Figure 7 This is a TSNE diagram for correct identification and classification as described in a specific embodiment of the present invention; Figure 8 The TSNE diagram for complete identification and classification as described in a specific embodiment of the present invention; Figure 9This is the TSNE graph after pseudo-label screening using TSNE dimensionality reduction analysis as described in a specific embodiment of the present invention; Figure 10 This is a line graph showing the average correct number of pseudo-tags and the fluctuation range under different signal-to-noise ratios, as described in a specific embodiment of the present invention. Figure 11 This is a line graph showing the average accuracy and fluctuation range under different signal-to-noise ratios as described in a specific embodiment of the present invention. Figure 12 This is a comparison chart of the recognition accuracy of the method under different signal-to-noise ratios according to a specific embodiment of the present invention; Figure 13 This is a comparison chart of the ablation experiment effects described in a specific embodiment of the present invention; Figure 14 This is a graph showing the accuracy variation during the iteration of the method described in this invention, as illustrated in a specific embodiment of the invention. Figure 15 This is a graph showing the accuracy change during iterations after removing LMMD, as described in a specific embodiment of the present invention. Figure 16 This is a graph showing the accuracy change during iterations after removing CDAN, as described in a specific embodiment of the present invention. Detailed Implementation

[0024] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application can also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0025] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0027] Implementation Method 1 This embodiment proposes a cross-receiver radiation source fingerprinting method based on receiver model and feature degradation isolation, including the following steps: S1: Establish a receiver mathematical model and a cross-receiver data migration model. Based on the receiver mathematical model and the cross-receiver data migration model, derive a feature extraction scheme that resists cross-receiver distortion, extract cross-receiver insensitive features, and avoid significant differences in feature extraction effect that are difficult to compensate for due to the replacement of receiver. S2: Construct a domain-adaptive neural network based on the feature extraction scheme of S1, and use the pseudo-label method to isolate feature degradation and generate high-confidence pseudo-labels; S3: Construct a master recognition network with alternating pseudo-label, CDAN, and LMMD guidance. Use the high-confidence pseudo-label generated in S2 to guide the initialization of the master recognition network. Train the master recognition network by alternating guidance of CDAN and LMMD. Use the trained network to achieve fingerprint recognition across receiver radiation sources.

[0028] Implementation Method 2 Reference Figure 1 , Figure 2 This implementation method is described below.

[0029] This embodiment is a further illustrative example of S1 described in Embodiment 1.

[0030] In S1, the process of establishing the receiver mathematical model is as follows: for the signal reception distortion of the zero intermediate frequency receiver, the non-ideal characteristics of the filter and the nonlinear distortion of the amplifier are selected as distortion factors. Based on the distortion factors, the receiver mathematical model is simplified into a Volterra simplified model that includes nonlinear distortion and memory effect.

[0031] Specifically: This model describes the process of a receiver receiving a signal and the potential distortions it may cause. For a common zero-IF receiver, the deterministic distortions in the signal reception process include non-ideal characteristics of the filter, nonlinear distortion of the amplifier, frequency source distortion, even harmonics, IQ imbalance distortion, and ADC distortion. However, considering the complexity of the model, it needs to be simplified. The simplified model focuses on the following two types of distortion: (1) Filter non-ideal characteristics. This mainly refers to the distortion of multiple filters in the receiving path. Due to design tolerances and manufacturing errors, errors inevitably occur in the passband of analog filters, mainly manifested as tilting or jittering of the frequency response amplitude within the band and non-ideal group delay characteristics. Different receivers have very different filter settings, so filter distortion is common among receivers. In zero-IF receivers, the analog filter is mainly a low-pass baseband filter. Compared with the intermediate frequency bandpass filter of a superheterodyne receiver, its hardware structure is simple and the distortion order is limited. However, the digital filter at the back end will cause its distortion order to deepen, specifically manifested as enhanced memory effect.

[0032] (2) Nonlinear distortion of amplifier. This mainly refers to the nonlinear characteristics of the amplifier in the entire receiving path. In the receiver design process, a low-noise amplifier is used to adjust the signal power. In this process, the characteristics of the amplifiers used by different receivers, the gain adjusted by the automatic gain control, and even the number and model of the amplifiers are different, resulting in nonlinear differences in the received signal.

[0033] Therefore, it mainly has two factors that may cause signal distortion: nonlinear distortion and memory effect.

[0034] Therefore, its actual model can be regarded as a simplification of the Volterra model.

[0035] The Volterra model incorporates all types of nonlinear distortion and memory effects, thus it can unify the nonlinear distortion of the equivalent amplifier and the memory effect of the equivalent filter. Its expression is: , (1) in, For the output of the model, The maximum nonlinear order of the model. to For the maximum length of the memory effect under different nonlinear orders, The kernel coefficients of the Volterra model under different nonlinear orders and memory effect lengths are given. The signal input is processed by the memory effect.

[0036] The memory polynomial model is a simplified Volterra model. Most terms in the Volterra series are redundant; redundant terms do not improve model accuracy but rather make the model unstable. This implementation removes several cores, resulting in the MP model, represented as: , (2) in, For the length of the model memory effect, This represents the model parameters under different nonlinear orders and memory effect lengths.

[0037] The model retains the most important nonlinear effects and memory effects, but does not consider the cross-effects between multiple nonlinear effects.

[0038] Furthermore, the cross-receiver data migration model describes a possible mathematical process, namely, the possible data migration between two receivers with different model parameters, primarily considering the receiver models. The relationship between the two models is as follows: Figure 1 As shown.

[0039] Based on the nonlinear effects of the two equivalent amplifiers, the signal after passing through the equivalent filter can be expressed as follows: The effect of the memory effect caused by the filter, its distortion, and other components on the overall signal can be considered as a lower triangular Toeplitz matrix. This means that each descending diagonal line from left to right is constant, thus satisfying the linear time-invariant filtering operation. Therefore, the actually received signal can be represented as: , (3) in, Toeplitz matrix representing the memory effect, and These represent the ideal signal and the actual signal sampled by the receiver, respectively.

[0040] If there is As If the inverse matrix exists, it represents the difference between signals acquired by different receivers when only the memory effect is considered. and They can be converted to each other, that is: , (4) Based on the properties of the Toeplitz matrix and the properties of the filter, we can conclude that... The conclusion is that there exists a matrix such that any of them is a Toeplitz matrix, therefore... It can be represented as: , (5) in, A cross-receiver migration model that exists for a certain theoretical purpose.

[0041] Furthermore, the derivation process of the feature extraction scheme against cross-receiver distortion is as follows: The memory effect in the time domain of the cross-receiver migration model is transformed to the frequency domain; the influence of the memory effect is equivalent to frequency domain weighting; time-frequency domain normalization is used to eliminate the influence of frequency domain weighting; random noise is introduced during the normalization process to erase the memory effect-related features of the cross-receiver migration model, thus obtaining the feature extraction scheme against cross-receiver distortion. The feature extraction steps are as follows: Figure 2 As shown.

[0042] Specifically: due to Not for The finite-length filter structure makes it unstable for correction from the perspective of model extraction. Even slight changes in parameters can lead to The dramatic changes in parameters, a characteristic that to some extent illustrates the instability of the cross-receiver data migration model, also contributes to this. However, the properties of its lower triangular Toeplitz matrix mean that it can still be considered a time-domain convolution, and after converting it to the frequency domain, it can be transformed into the following form: , (6) in, and Two different receivers and The frequency domain of the signal To be It can be regarded as a frequency domain representation after temporal convolution.

[0043] Therefore, the memory effect across receivers can be viewed as a frequency-domain weighting, with the weights of different frequencies changing as the receiver changes. Thus, it can be eliminated using a special normalization method. Taking STFT as an example of a time-frequency variation method, the specific implementation involves, for any signal from the same receiver, calculating the average value of all frequencies along the time axis after the STFT, and then dividing by this average value to achieve the desired normalization. The removal of.

[0044] To eliminate the memory effect, the following conditions must be met: 1. Non-time domain: The cross-receiver model actually includes the inverse process of a filter. If we consider it as a change in the time domain, it can be proven that there is no filter of a fixed length that can completely eliminate the memory effect between receivers. Therefore, it cannot be removed by a time-domain-only method.

[0045] 2. Randomness: While erasing memory effects, additional noise related to memory effects should be randomly generated to prevent the neural network from learning incorrectly during the erasure process.

[0046] 3. Controllability of feature degradation: Such erasure will cause feature degradation, and this degradation effect should be kept within a controllable range. For example, memory effect erasure cannot be performed by randomly generating filters with arbitrary parameters, because such a brutal erasure will severely affect other features, or even render them completely unusable for recognition.

[0047] In other words, if the memory effect is to be erased, erasure should be considered in the time-frequency domain, and the erasure method may change as the time-frequency changes.

[0048] Taking the most common STFT time-frequency variation as an example, the impact of the cross-receiver memory effect on different receivers can be regarded as a kind of frequency domain weighting, with the weight of different frequency points changing with the receiver. Specifically, for any signal from the same receiver, the average value at that frequency is calculated along the time axis after the STFT. This value is then used to normalize each frequency axis on the time-frequency graph. To introduce randomness, the average value is generated primarily from the signal itself, and some neighboring signals are introduced as secondary components or random noise, rather than from statistics derived from the entire long-time signal collected by the receiver. Therefore, this method can be used to achieve... Removal at the time-frequency plot level.

[0049] The same method as STFT can be used for time-frequency plots generated by other time-frequency transforms such as CWT. However, taking CWT as an example, using this method will destroy the properties of the wavelet basis, so its adaptability to other time-frequency plots needs further analysis.

[0050] This feature erasure method based on time-frequency maps also introduces new problems. While extracting time-frequency domain features, this method inevitably introduces spectral distortion and weakens the correlation between the time-domain axis and the frequency-domain axis in the time-frequency map. In other words, this will damage the features themselves, making it difficult for the neural network using this method to achieve the ideal recognition accuracy in non-domain adaptive environments or when the feature distribution changes little due to changes in the receiver.

[0051] Completely eliminating signal variations caused by memory effects is a difficult ideal to achieve. Moreover, during targeted elimination, new memory effects are inevitably introduced due to differences in the accuracy of parameter extraction or errors in parameter estimation. This analysis can also be verified in past signal processing steps. In high-precision collaborative scenarios such as communication and speech enhancement, an adaptive filtering framework is used to solve this problem. This allows the filter coefficients to be dynamically adjusted according to the statistical characteristics of the signal, thereby "forgetting" past irrelevant information and tracking the currently useful signal components. When dealing with strong interference signals such as earthquakes and radar signals with known characteristics, time-varying filters based on time-frequency analysis are usually used. This method determines the time-frequency location of the interference and signal through time-frequency analysis and then designs a corresponding time-varying filter. For complex signals with abundant data, such as audio restoration and data cleaning, various data-driven methods are usually used to separate memory effects.

[0052] However, in the scenario of cross-receiver radiation source fingerprinting, the problem is that the data is complex and the attributes are insufficient (the target domain lacks labels). It is impossible to remove the memory effect through the two directional filtering methods mentioned above, and it is also difficult to rely entirely on the neural network-based approach (because the features of the source domain also exist on top of the memory effect).

[0053] Considering these issues, this embodiment employs the aforementioned method to remove the memory effect, namely, by using a targeted erasure and blurring of cross-receiver sensitive features, to prevent the subsequent neural network from learning any features related to the memory effect.

[0054] Implementation Method 3 Reference Figure 3 This implementation method is described below.

[0055] This embodiment is a further illustrative example of S2 described in Embodiment 1.

[0056] Step S2 can isolate the feature extraction methods that have special functions (high stability and resistance to receiver distortion) but degrade the features themselves, thus preserving their high stability and resistance to distortion while avoiding the impact of degraded features on the final recognition accuracy.

[0057] This embodiment takes into account that feature erasure in embodiment two only erases the feature degradation caused by the memory effect, while still retaining the feature degradation caused by nonlinear effects and other possible random factors. Therefore, it is necessary to use a domain-adaptive neural network for further model correction to improve its recognition effect.

[0058] In S2, the domain-adaptive neural network is constructed as follows: The network employs a feature extraction scheme that resists cross-receiver distortion, which suppresses the negative migration phenomenon caused by the incomplete matching between the cross-receiver problem and the domain adaptation problem. It is then connected to a domain adaptation neural network based on CDAN, which uses domain adaptation to make the neural network robust for cross-domain recognition, thereby realizing the ability to identify radiation source fingerprints across receivers.

[0059] In the field of cross-receiver source fingerprinting, this problem differs slightly from common domain adaptation problems. In this case, the physical nature of transmitter features and receiver distortions is the same, leading to additional feature degradation or distortion at the feature level, rather than simply the mismatch between the source and target domain probability distributions assumed in common domain adaptation methods. The impact of cross-receiver distortion on source fingerprint features is significant, potentially completely masking or altering some of the original features of the source. Under these conditions, negative transfer caused by not meeting domain adaptation conditions becomes the main reason for reduced stability and accuracy of neural network cross-receiver transfer. Therefore, this problem lies on the boundary of domain adaptation, and domain adaptation methods alone cannot completely solve it.

[0060] This implementation uses CDAN, a representative method in unsupervised domain adaptation. By introducing a conditional adversarial mechanism, it significantly improves the model's adaptability to the target domain. The core of CDAN lies in conditional adversarial training. Unlike traditional methods that only align the global feature distribution, CDAN combines category information (the classifier's predicted output) with the feature output of the feature extractor, both of which are used simultaneously for domain identification by the domain discriminator. Adversarial training using gradient inversion layers aligns the distribution of source and target domain features, effectively avoiding confusion and mismatch between different categories of features during domain adaptation. This implementation uses the CDAN network primarily because its strong adversarial properties suppress negative transfer caused by initial classification errors in this application scenario.

[0061] Furthermore, in S2, the process of generating high-confidence pseudo-labels using the pseudo-label method is as follows: the domain-adaptive neural network is used to predict the unlabeled target domain data to generate initial pseudo-labels, and the initial pseudo-labels are filtered by a combination of threshold screening and multi-model voting to obtain high-confidence pseudo-labels.

[0062] Specifically: First, a pseudo-label generation network is used to predict unlabeled target domain data and generate initial pseudo-labels. The entropy regularization mechanism of the pseudo-label method can encourage the model to make low-entropy, high-confidence predictions on unlabeled data, reducing the dependence on manually labeled data. The pseudo-label information in the prediction results is extracted, and the degraded feature data generated after feature extraction is discarded. The pseudo-label is used as the carrier of the anti-distortion feature extraction method to avoid degraded features from entering the subsequent training process of the main recognition network, thereby fundamentally isolating the impact of feature degradation on recognition accuracy.

[0063] To ensure the accuracy of pseudo-labels, this implementation adopts a fixed threshold screening + proportional screening + three-model voting screening strategy to screen the initial pseudo-labels layer by layer and generate high-confidence pseudo-labels. Through this method, the label stability of the features that resist receiver distortion can be retained as a pseudo-label form generated by the pseudo-label method, while degraded features are first screened by the pseudo-label threshold and then excluded from the influence in the pseudo-label generation step.

[0064] The specific operating procedure is as follows: 1. Set a fixed confidence threshold to remove initial pseudo-labels with confidence levels below the threshold; then set a pseudo-label ratio threshold to filter the number of pseudo-labels in each category of the target domain, thus mitigating the impact of category imbalance. 2. Train three pseudo-label generation networks with identical structures independently. Use the three networks to independently predict and vote on the initial pseudo-labels after double threshold screening. Only retain pseudo-labels with "average confidence greater than the threshold + pseudo-labels labeled by the three networks are completely consistent" and identify them as high-confidence pseudo-labels. 3. Combine the high-confidence pseudo-labels obtained from the final screening with the corresponding target domain data to construct a pseudo-label dataset, which will be used for the initial training of the main recognition network.

[0065] The method proposed in this implementation uses an entropy regularization mechanism to encourage the model to make low-entropy (i.e., high-confidence) predictions on unlabeled data. This helps the decision boundary shift towards low-density regions between classes, aligning with the basic assumptions of semi-supervised learning and effectively improving the model's generalization performance on the test set. The pseudo-label method has the advantages of simple implementation, compatibility with various neural network models, and no dependence on specific data augmentation methods, making it widely applicable. It provides an effective approach to addressing the key challenge of scarce labeled data in real-world tasks.

[0066] Since Implementation Method 2 proposes a method that can partially suppress cross-receiver feature distortion but causes feature degradation itself, and the negative migration phenomenon caused by the incomplete conformity with the domain adaptation assumption in cross-receiver radiation source fingerprint recognition is the key issue, this implementation method combines the anti-receiver distortion feature extraction method based on the cross-receiver migration model with the pseudo-label method. It only retains the pseudo-labels assigned to the target domain data by the pseudo-label method, thereby retaining its relatively stable class classification recognition ability while avoiding the decline in recognition accuracy caused by feature degradation affecting the overall recognition accuracy. Furthermore, it uses the fixed and relative dual thresholds of pseudo-label confidence to filter pseudo-labels, thereby achieving the goal of generating pseudo-labels with high confidence and stable class classification accuracy.

[0067] The reason for using feature isolation instead of the commonly used feature fusion method is that the anti-receiver distortion feature extraction method based on the cross-receiver migration model proposed in Implementation Method 2 does not propose new features. The features used are essentially erasers of existing features with high cross-receiver sensitivity ranges, rather than creating or extracting new features. Moreover, the characteristics of this feature are quite obvious: it has a high class classification accuracy and can suppress negative migration, but the degradation of the feature itself will lead to a decrease in classification accuracy when negative migration does not occur. Therefore, it is not suitable to introduce this method into the main recognition network using a feature fusion-based approach.

[0068] Implementation Method 4 This embodiment is a further illustrative example of S3 described in Embodiment 1.

[0069] The S3 described in this embodiment proposes a main recognition network that combines pseudo-labeling method with domain-adaptive neural network to address feature degradation problem. Through degradation isolation, the main recognition network can simultaneously achieve high stability from data correction and high recognition accuracy of the original data.

[0070] The main identification network structure described in this embodiment is shown in Figure 4.

[0071] The high-confidence pseudo-labels are used to initialize and guide the main classification model. This improves the stability of the main recognition network and enhances the accuracy of the initial classification of the neural network, thus avoiding class mismatch in the initial recognition and suppressing the occurrence of negative transfer.

[0072] S3 specifically includes the following steps: Design an initial total loss function, which includes the source domain loss and the pseudo-label loss, expressed as: , (7) in For the total loss function, For source domain loss, For pseudo-label weight hyperparameters, This is due to the loss from false labels.

[0073] In this process, the network's recognition performance in the target domain depends solely on the generated pseudo-labels and their loss function during training. and For target domain categories with correct and sufficient generated pseudo-labels, the initialized main recognition network can correctly identify them as the corresponding type of the target domain under the guidance of the pseudo-labels. For target domain categories with insufficient pseudo-labels but basically correct labels, the initialized main recognition network can fix this part and similar samples as the correct type under the guidance of the pseudo-labels. Even if negative transfer occurs in the subsequent domain adaptation process, it will be easy to transfer to this class because there are parts of this class that are fixed as the correct identification class. For target domain categories with almost no pseudo-labels, since the initialized network does not have the ability to adapt to the domain, that is, there are only losses based on the source domain and losses based on pseudo-labels, the negative transfer phenomenon will not occur in this process. For a small number of possible but incorrect pseudo-labels in the target domain, they can also be corrected to a certain extent by relying on the strong adversarial nature of the subsequent CDAN neural network.

[0074] CDAN, representing adversarial domain-adaptive neural networks, possesses powerful distribution alignment capabilities. By aligning distributions across different domains, it learns domain-invariant features, thereby enhancing the model's adaptability in the target domain. Compared to domain-adaptive methods like DANN, CDAN fully considers the impact of classification labels on domain adaptation when aligning features. It combines feature information with the soft labels predicted by the classifier through operations such as outer product, performing conditional adversarial training to achieve fine-grained alignment based on class awareness. This effectively captures the multimodal structure behind the data, making cross-domain feature distribution matching more accurate.

[0075] However, due to its intense adversarial nature, CDAN can cause more drastic fluctuations in the feature extraction part of the neural network. Therefore, it is more suitable for use in the middle of the iteration process to avoid the feature extraction network getting trapped in local optima. For the small number of incorrect target domain pseudo-labels that may appear during pseudo-label guidance, CDAN's strong adversarial nature can help it successfully escape the current error type and revert to the correct classification. For target domain categories with few or no pseudo-labels, this strong adversarial characteristic and domain adaptability improve the generalization ability of the neural network, thus converging it towards the correct classification.

[0076] The loss function of CDAN is: , (8) in For domain classifiers, and These are the feature vectors of the source and target domains. and The predicted class label.

[0077] CDAN's strong adversarial nature brings strong generalization ability to the model, which can solve the misclassification caused by class imbalance or a small number of pseudo-label errors in the previous pseudo-label guidance. However, its high adversarial nature also means its instability. Therefore, in the later batches of iteration, the CDAN loss function can be reduced by introducing the LMMD loss function, thereby reducing the instability of the adversarial domain-adaptive neural network and improving the stability of the neural network.

[0078] The LMMD loss function relies on the existing self-generated target domain classification, which, while enhancing the current recognition performance, easily leads to local optima or class mismatches. Therefore, it is more suitable for enhancing the generalization ability and learning effect of the neural network after the network has initially stabilized. Hence, it is strengthened in the later stages of iteration to solve the convergence instability caused by the strong adversarial nature of CDAN.

[0079] The loss function for LMMD is: , (9) in, This represents the output of the last fully connected layer. It is the regenerating nucleus Hilbert space A nonlinear feature mapping, and These represent the number of source and target domains under category c, respectively.

[0080] In summary, the overall loss function is: , (10) in This represents the current iteration number. This represents the total number of iterations. for Weights in the loss function for Weights in the loss function.

[0081] Therefore, the complete training process includes the following steps: 1) Using a receiver distortion-resistant neural network, three pseudo-label generation networks are trained simultaneously with as few iterations as possible (5 iterations).

[0082] 2) Use three pseudo-label networks to vote and generate pseudo-labels for part of the target domain data to guide the main identification network.

[0083] 3) Simultaneously, pseudo-labels from the source domain data and target domain data are used to guide the initialization of the main recognition network (first N iterations).

[0084] 4) By utilizing CDAN-LMMD cross-guidance, the neural network eventually converges.

[0085] Implementation Method 5 This embodiment is a further illustrative example of a cross-receiver radiation source fingerprinting method based on receiver model and feature degradation isolation, as described in Embodiments 1 to 4.

[0086] The dataset used in this embodiment is called the WiSig dataset, which is a large Wi-Fi dataset consisting of 174 Wi-Fi transmitters and 41 USRP receivers. To select a larger data sample, the SingleDay portion was chosen for the experiment, containing 800 data entries per day from 28 transmitters and 10 receivers. This embodiment selected 6-8 different pairs of transmitters and their data on various receivers for subsequent experiments.

[0087] Experiment (1) Considering that this experiment needs to verify the stability and feasibility of the method through alternating migrations of more different receivers, receivers 1 to 6 in the SingleDay part of the WiSig dataset were selected, and they were used as source and target receivers to migrate alternately. The selected transmitter sample consisted of 6 transmitters randomly selected from transmitters 1 to 10 as a group for the six-class classification experiment. Multiple different transmitter combinations were selected to enhance the persuasiveness and representativeness of the experiment.

[0088] In this implementation, both the domain adaptation network and the pseudo-label generation network used in the experiment are based on CDAN domain adaptation. When comparing feature extraction performance, the two networks differ only in the feature extraction network part: one uses the proposed special feature extraction method based on neural networks for cross-receiver radiation source fingerprinting, while the other does not use this method but directly performs time-frequency transformation for feature extraction as a control. The pseudo-label generation network and the domain adaptation network differ only in whether the final fully connected layer is connected to the pseudo-label screening module.

[0089] In cross-receiver radiation source fingerprinting, the most common error is whole-class labeling error caused by negative migration, also known as label mismatch, where a transmitter in the target domain data is completely mistaken for another transmitter in the source domain data. This is the most significant factor affecting the overall recognition rate in cross-receiver radiation source fingerprinting, while this phenomenon is less common in other domain adaptation fields. Therefore, it is necessary to consider label classification errors separately.

[0090] Therefore, a pseudo-label accuracy rate of over 60% is considered a correct pseudo-label classification. This means that using the pseudo-label to guide the subsequent training of the network will reduce the occurrence of class mismatch.

[0091] By traversing the transmitter data collected from receivers 1 to 6 in the dataset, and through multiple experiments, the effectiveness of the anti-receiver distortion feature extraction method with and without cross-receiver transfer models for the recognition of domain-adaptive neural networks was statistically analyzed under different numbers of erroneous labels. Figure 5 As shown in Table 1, the error statistics for pseudo-labels are as follows.

[0092] Table 1. Impact of Feature Extraction Methods on Pseudo-Label Generation Results

[0093] Figure 5 Under the condition that the pseudo-label effect is the same, the recognition network using the special feature extraction method has a higher recognition accuracy than the network without special feature extraction, and its stability is also stronger than the network without special feature extraction. This experiment fully demonstrates that the feature extraction method proposed in this invention has strong recognition accuracy and stability in the cross-receiver radiation source fingerprint recognition problem.

[0094] Table 1 compares the number of pseudo-label errors after generating pseudo-labels with and without special feature extraction. Firstly, a less frequent type of label error arises from the aforementioned assumption: in the cross-receiver radiation source fingerprinting problem, the dominant cause is negative migration and label mismatch due to the incomplete fit between the cross-receiver problem and the domain adaptation assumption. In other words, errors primarily occur through the interchange of a pair of labels in the target domain. Secondly, by statistically analyzing the number of pseudo-label generation errors, it can be concluded that, from the perspective of generating pseudo-labels for the target domain, the feature extraction method proposed in this invention significantly reduces the occurrence of pseudo-label errors.

[0095] Experiment (II) This experiment selected a pair of receivers, X to Y, that showed good transfer performance in Experiment (I) as the source and target receivers for this experiment. Data from transmitters were randomly selected from the transmitter list to construct a dataset for repeated experiments. The average classification accuracy and the accuracy of generated pseudo-labels were statistically analyzed. The recognition results are as follows: Figure 6 As shown in Table 2, the error statistics for pseudo-labels are as follows. Since no odd number of label errors resulting from non-label mismatches occurred in this experiment, the statistics are categorized into three types: no pseudo-label errors, one pair of pseudo-label errors, and more pseudo-label errors.

[0096] Table 2. Impact of Feature Extraction Methods on Pseudo-Tag Generation Effectiveness When Receiver Differences are Small

[0097] Figure 6 In cases where receiver differences are small and there are no pseudo-label errors, the recognition accuracy using the special feature extraction method is slightly lower than that of the neural network without this method. This verifies the assertion that this special feature extraction method will cause feature damage. It is precisely because of this feature degradation phenomenon that the necessity of using the pseudo-label method for feature isolation is also demonstrated.

[0098] Table 2 shows the false label errors in this experiment, which proves that even when the differences in receivers are small, class mismatch often occurs. Class mismatch further becomes the main factor hindering the improvement of overall recognition accuracy. However, using this feature extraction method can significantly suppress the occurrence of class mismatch.

[0099] To verify that the pseudo-label method is effective and reasonable for filtering target domain samples, TSNE dimensionality reduction analysis was used to analyze the pseudo-label filtering effect. Figure 7 , 8 As shown in Figure 9 exist Figure 7-9 The TSNE diagram drawn with all labels shows a noticeable tendency to misidentify some target domains, such as category A, as closely related category B (the brown areas in the diagram are identified as blue). While this isn't a serious case of label mismatch, using all labels indiscriminately for subsequent recognition, or even treating them as the entire recognition process, will negatively impact the overall recognition performance. After filtering out false labels, such as... Figure 9 As shown, in this case, the incorrect labels are eliminated through pseudo-label filtering. Through TSNE dimensionality reduction and analysis, the pseudo-label filtering process significantly and intuitively increases the accuracy of pseudo-labels.

[0100] Considering the need to test the impact of using this feature extraction method at different signal-to-noise ratios on the accuracy of pseudo-label generation and the overall recognition network accuracy, an experiment is required. This experiment compares the presence and absence of the feature extraction method proposed in this invention as the sole difference between the two networks. The data used consists of data collected by six receivers (receivers 1 to 6) in a centralized dataset, transferred between them under random transmitter conditions, and traversed all possible source-target receiver combinations at signal-to-noise ratios ranging from -10 to 28 dB. Figure 10 To illustrate the error between the average number of correct pseudo-labels and the average accuracy rate under different signal-to-noise ratios. Figure 11 To illustrate the variation in accuracy and error under different signal-to-noise ratios.

[0101] pass Figure 10The experimental results demonstrate that at signal-to-noise ratios (SNR) greater than 5 dB, the pseudo-label generation network with special feature extraction exhibits better pseudo-label generation performance at the same SNR. In environments with SNRs of 10 dB or higher, the average accuracy of pseudo-labels generated by the network with special feature extraction exceeds that of the source network at any SNR. Furthermore, the stability and fluctuation range of pseudo-labels generated by the network with special feature extraction are superior to the source method at all SNRs. Although its performance is poor in low SNR environments, pseudo-labels generated by the network without this special feature extraction are also unusable due to their low accuracy under low SNR conditions. Therefore, the superiority of the proposed method in target domain pseudo-label generation can be considered.

[0102] pass Figure 11 The experimental results show that the domain adaptation network with special feature extraction proposed in this invention has poor recognition efficiency at low signal-to-noise ratios due to feature degradation, only exceeding the baseline model at 15dB. However, the pseudo-tag generation process can reduce this relatively poor recognition efficiency by about 10dB. Therefore, the pseudo-tag method can successfully achieve the purpose of degradation isolation, thereby generating pseudo-tags with anti-trans-receiver distortion characteristics.

[0103] Experiment (3) To compare this method with other cross-receiver radiation source fingerprinting methods, this experiment selected one non-domain-adaptive neural network and five domain-adaptive neural networks to identify six transmitters under seven different receiver migration conditions, comparing their accuracy on the target receiver. These included: Empirical Risk Minimization (ERM), Joint Adaptation Network (JAN), Margin Disparity Discrepancy (MDD), and Minimum Class Confusion (MCC).

[0104] The test results are shown in Table 3.

[0105] Table 3. Recognition accuracy compared with other methods

[0106] Experimental results show that, compared with other domain adaptation methods, the proposed method significantly improves the recognition accuracy in the target domain, achieving an average accuracy of 98.47%, which is significantly better than other domain adaptation methods. To further investigate the stability of the model under different signal-to-noise ratio (SNR) conditions, this embodiment evaluated the proposed method in an environment with an SNR ranging from -10 dB to 20 dB. Two high-performing methods selected from Table 1—MDD and CDAN—were used for comparison. In the experiment, data from receiver 8 was used as the source domain, and data from receiver 9 was used as the target domain. Detailed comparison results are as follows: Figure 12 As shown.

[0107] This experiment demonstrates that the proposed method significantly outperforms existing methods under various signal-to-noise ratio (SNR) conditions. Specifically, the method achieves a recognition accuracy of 87.5% at a SNR of 0 dB, while maintaining an accuracy above 99% when the SNR increases to 10 dB. The overall performance curve exhibits a smooth upward trend, and the variance is extremely small across multiple experimental runs, indicating that the proposed method possesses high robustness and consistency.

[0108] To evaluate the contribution of each subsystem in the proposed method, this embodiment designed an ablation experiment including four specific ablation conditions: Remove the conditional domain adversarial network loss from the main identification network; Remove the local maximum mean difference loss from the main recognition network; The mechanism for removing fake tags; While retaining the pseudo-label guidance, the anti-receiver distortion structure was removed from the pseudo-label generation model.

[0109] Each ablation configuration was compared to a complete neural network (without any removal) through repeated testing. The results of these comparative experiments are summarized in... Figure 13 .

[0110] In this experiment, the absence of a pseudo-labeling method or the lack of receiver distortion resistance in the pseudo-label generation model both led to increased volatility in recognition accuracy. However, even under these conditions, when the model is ideally initialized, it still has the potential to achieve relatively high accuracy. This observation fully demonstrates that the aforementioned components play a decisive role in improving the model's recognition rate and stability. Nevertheless, Figure 13The results do not clearly reveal the specific contribution of the alternating guidance mechanism composed of CDAN and LMMD. Although the model using this alternating guidance differs from the full model in accuracy volatility, the difference is not significant enough. Therefore, to further explore the effect of this alternating guidance loss function from multiple perspectives, this implementation analyzes the accuracy curves as a function of iterations during multiple training runs. To amplify subtle performance differences, the number of transmitters was increased to eight in this experiment. The test results are as follows: Figure 14-16 As shown.

[0111] Figure 14-16 The solid line in the graph represents the accuracy variation curve in each trial. Figure 16 In the model, after removing the CDAN method, the accuracy slowly increases with the number of iterations, but there is a certain probability that it will converge to a local optimum. Figure 15 In practice, when the LMMD method is absent, accuracy fluctuations during training increase significantly. In contrast, Figure 14 The results show that the proposed combined method improves convergence stability, with the adversarial nature of CDAN helping the model escape local optima. These experimental results collectively validate the superiority of the proposed method.

[0112] It is understood that the present invention has been described through some embodiments 13. Those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A cross-receiver radiation source fingerprint recognition method based on receiver model and feature degradation isolation, characterized in that, Includes the following steps: S1: Establish a receiver mathematical model and a cross-receiver data migration model. Based on the receiver mathematical model and the cross-receiver data migration model, derive a feature extraction scheme that resists cross-receiver distortion and extract cross-receiver insensitive features. S2: Construct a domain-adaptive neural network based on the feature extraction scheme of S1, and use the pseudo-label method to isolate feature degradation and generate high-confidence pseudo-labels; S3: Construct a master recognition network with alternating pseudo-label, CDAN, and LMMD guidance. Use the high-confidence pseudo-label generated in S2 to guide the initialization of the master recognition network. Train the master recognition network by alternating guidance of CDAN and LMMD. Use the trained network to achieve fingerprint recognition across receiver radiation sources.

2. The cross-receiver radiation source fingerprint recognition method based on receiver model and feature degradation isolation according to claim 1, characterized in that, The process of establishing the receiver mathematical model is as follows: for the signal reception distortion of the zero intermediate frequency receiver, the non-ideal characteristics of the filter and the nonlinear distortion of the amplifier are selected as distortion factors. Based on the distortion factors, the receiver mathematical model is simplified into a Volterra simplified model that includes nonlinear distortion and memory effect.

3. The cross-receiver radiation source fingerprint recognition method based on receiver model and feature degradation isolation according to claim 1, characterized in that, The process of establishing the cross-receiver data migration model is as follows: the memory effect caused by the filter and other components is equivalent to a lower triangular Toeplitz matrix, the mutual conversion relationship of the signals acquired by receivers with different parameters is derived, and the cross-receiver migration model is obtained based on the conversion relationship to realize the mathematical description of the data migration process between different receivers.

4. The cross-receiver radiation source fingerprint recognition method based on receiver model and feature degradation isolation according to claim 1, characterized in that, The derivation process of the feature extraction scheme against cross-receiver distortion is as follows: the memory effect in the time domain of the cross-receiver migration model is transformed to the frequency domain, the influence of the memory effect is equivalent to frequency domain weighting, and the influence of frequency domain weighting is eliminated by time-frequency domain normalization. Random noise is introduced in the normalization process to erase the memory effect-related features of the cross-receiver migration model, thus obtaining the feature extraction scheme against cross-receiver distortion.

5. The cross-receiver radiation source fingerprint recognition method based on receiver model and feature degradation isolation according to claim 1, characterized in that, The construction process of the domain-adaptive neural network includes: connecting the feature extraction scheme of S1 to the feature extraction layer of the CDAN domain-adaptive neural network.

6. The cross-receiver radiation source fingerprint recognition method based on receiver model and feature degradation isolation according to claim 1, characterized in that, In S2, the process of generating high-confidence pseudo-labels using the pseudo-label method is as follows: the domain-adaptive neural network is used to predict unlabeled target domain data to generate initial pseudo-labels, and the initial pseudo-labels are filtered by a combination of threshold screening and multi-model voting to obtain high-confidence pseudo-labels.

7. The cross-receiver radiation source fingerprint recognition method based on receiver model and feature degradation isolation according to claim 1, characterized in that, In step S3, the alternating guidance training of the main recognition network includes: Three independent pseudo-label generation networks were trained using the feature extraction scheme described in S1; Three pseudo-labels are used to generate network voting and filter, resulting in high-confidence pseudo-labels for initializing the main recognition network. Using source domain labeled data and high-confidence pseudo-label data, the main recognition network is guided to complete initial training by initializing the total loss function; The weights of the CDAN and LMMD loss functions are alternately adjusted according to the iterative process. The main recognition network is trained until convergence through a fusion-type total loss function, thereby achieving fingerprint recognition across receiver radiation sources.

8. A fingerprint recognition system for cross-receiver radiation sources based on receiver model and feature degradation isolation, characterized in that, The system is implemented based on the cross-receiver radiation source fingerprinting method based on receiver model and feature degradation isolation as described in any one of claims 1-7, and the system includes: Feature extraction module: used to establish receiver mathematical model and cross-receiver data migration model, derive feature extraction scheme against cross-receiver distortion based on the receiver mathematical model and cross-receiver data migration model, and extract cross-receiver insensitive features; The pseudo-label generation module is used to construct a domain-adaptive neural network based on the feature extraction scheme, and uses the pseudo-label method to isolate feature degradation and generate high-confidence pseudo-labels. Network training module: used to construct a master recognition network guided by a triple alternation of pseudo-label, CDAN, and LMMD. The master recognition network is initialized using high-confidence pseudo-labels, trained by alternating CDAN and LMMD, and then used to achieve fingerprint recognition across receiver radiation sources.

9. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes a cross-receiver radiation source fingerprinting method based on receiver model and feature degradation isolation according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program that executes a cross-receiver radiation source fingerprinting method based on receiver model and feature degradation isolation as described in any one of claims 1-7.

Images