Open-set emitter individual identification method and system based on spatial-frequency domain fusion and hybrid-evt

Abstract

This invention relates to a method and system for identifying individual open-set radiation sources based on space-frequency domain fusion and Hybrid-EVT, belonging to the field of radiation source identification technology. It addresses the problems of frequent occurrences of unknown emission sources and the failure of the closed-set hypothesis in radiation source identification under complex electromagnetic environments. The method includes: acquiring raw IQ signals, preprocessing them to obtain a time-frequency map; extracting features from the time-frequency map based on the SFFNet model to obtain a discriminant activation vector; training the SFFNet model and a classification head based on samples of known radiation source categories, and calculating the mean activation vector for each category; constructing a Hybrid-EVT hybrid extremum model based on the training set activation vectors; inputting test samples into the SFFNet model with frozen parameters to obtain the activation vector of the test samples; evaluating the tail probability of the distance between the test samples and candidate categories, introducing unknown channel probabilities; if the unknown confidence level meets preset conditions, it is determined to be an unknown radiation source; otherwise, the corresponding known category is output. This invention is applicable to scenarios involving enhanced security for wireless device authentication.

Description

Technical Field

[0001] This invention belongs to the field of radiation source individual identification technology, specifically relating to an open-set radiation source individual identification technology based on spatial-frequency domain fusion and Hybrid-EVT. Background Technology

[0002] With the rapid increase in the number of wireless transmitting devices, the radio spectrum environment is characterized by diverse signal types, complex systems, and significant dynamic changes in services. At the same time, security threats such as signal spoofing and unauthorized access are becoming more prominent, easily posing risks to communication network security and personal privacy. Existing wireless terminal authentication largely relies on upper-layer protocol identifiers or key mechanisms, such as using MAC addresses, serial numbers, or shared keys for access control and authentication. However, this identifier information can be copied or tampered with in open wireless environments, making it difficult to provide sufficiently strong security guarantees in high-adversarial and high-dynamic scenarios. To improve the security and robustness of wireless device authentication, Specific Emitter Identification (SEI) has attracted attention as a physical layer authentication technology. SEI utilizes radio frequency fingerprints (RFF) to characterize the non-ideal hardware characteristics of transmitting devices, thereby achieving the differentiation and authentication of transmitting device identities. RFFs typically originate from subtle physical layer features such as amplitude and phase distortion, frequency drift, and nonlinear distortion caused by differences in device manufacturing. Therefore, compared to protocol layer identifiers that can be directly modified, they have stronger unforgeability and are more difficult to replicate, and have shown application prospects in fields such as wireless access control and spectrum regulation.

[0003] Existing SEI methods can be broadly categorized into two types: manual feature methods and deep learning methods. Manual feature methods typically extract fingerprint features based on prior mechanisms, such as IQ imbalance, carrier frequency offset, and Hilbert-Huang transform, and then combine these with traditional classifiers for identification. These methods have relatively low requirements for training data, but feature design relies on experience and prior assumptions, resulting in limited adaptability under complex channels, non-stationary interference, and multi-mode signal conditions. Deep learning methods, on the other hand, achieve fingerprint representation and classification through end-to-end learning. Recent advancements have been made in few-shot learning, domain adaptation, multimodal fusion, and lightweight models; however, most research still relies on the closed-set assumption, meaning that only device categories seen in the training phase exist during the testing phase. In real-world open wireless environments, unknown transmitting devices may appear at any time, and the number and operational status of devices are dynamic. Therefore, Open-Set Specific Emitter Identification (OS-SEI) has become a relevant research direction. OS-SEI aims to simultaneously achieve accurate classification of known transmitters and effective detection of unknown transmitters, thereby providing more robust authentication capabilities for dynamic and complex wireless communication scenarios. Existing OS-SEI research mainly follows two approaches: one is prototype learning or metric learning, which constructs a discriminant space by learning class centers or similarity metrics; the other is tail modeling combined with Extreme Value Theory (EVT) to fit class boundaries and reduce open space risks. In addition, some works have attempted to introduce generative methods or one-class learning to improve unknown detection capabilities through pseudo-sample expansion or boundary constraints.

[0004] While the aforementioned methods improve open set recognition performance to some extent, they still have limitations in OS-SEI scenarios with high adversarial nature, strong similarity, and complex channel conditions. Firstly, existing methods are insufficient in extracting highly discriminative RF fingerprint features, especially when multiple devices have highly similar fingerprints, significant channel variations, or strong noise interference. This makes it difficult to simultaneously ensure inter-class separability and intra-class compactness of the feature space, thus affecting the consistency and stability of known classifications and unknown detections. Secondly, many methods rely on tail distribution fitting to characterize class boundaries, but tail samples are naturally scarce and susceptible to noise and domain shifts, resulting in insufficient accuracy and generalization ability in tail modeling, and potentially high open space risk. In high-similarity scenarios, calibration or boundary fitting strategies may also suffer from threshold sensitivity and calibration failure. Furthermore, existing open set decision mechanisms often lack sufficient flexibility, making it difficult to simultaneously achieve low false negatives and low false positives for unknown detections under different application constraints, and placing higher demands on the dynamic emergence of unknown devices, incremental access, and stable authentication during long-term operation. Therefore, given the shortcomings of existing OS-SEI technology in discriminative feature learning, tail modeling accuracy, and robustness of open set decision-making, there is an urgent need for a technical solution that can more reliably distinguish known devices and more accurately identify unknown devices in complex open environments, in order to solve the key problems faced by open set radiation source individual identification in practical deployment. Summary of the Invention

[0005] This invention provides a method and system for identifying open-set radiation sources based on spatial-frequency domain fusion and Hybrid-EVT. The purpose is to solve problems such as insufficient high-discriminative RF fingerprint feature extraction, difficulty in balancing inter-class separability and intra-class compactness of the feature space in complex scenarios, insufficient tail modeling accuracy and generalization ability, susceptibility to noise and domain offset and threshold sensitivity, lack of flexibility in open-set decision mechanism, and difficulty in adapting to dynamic and complex real-world application constraints.

[0006] Firstly, the purpose of this invention is to provide a method for identifying individual open-set radiation sources based on spatial-frequency domain fusion and Hybrid-EVT, comprising the following steps: Signal acquisition and preprocessing steps: Acquire the raw IQ signal, perform preprocessing to obtain the time-frequency diagram, which serves as the input for radio frequency fingerprint characterization; Feature extraction steps: Feature extraction is performed on the time-frequency map based on a spatial-frequency domain fusion CNN model (SFFNet) to obtain discriminative activation vectors. The SFFNet model includes a three-level segmented network, namely: First stage: Multi-scale spatial features are extracted using ConvNeXt, and after being convolved with 1×1 to unify the number of channels and upsampled to a consistent spatial resolution, they are concatenated to obtain the input for the second stage; The second stage involves constructing four complementary representations through a frequency-space mapping module, including global and local features in the spatial domain, as well as low-frequency and high-frequency features in the frequency domain. The third stage: Semantic alignment and feature selection are performed on local features and high-frequency features, and global features and low-frequency features through two multi-scale dual representation alignment filters, respectively. Then, after two levels of learnable weighted fusion and lightweight classification head processing, the discriminant activation vector is output. Closed-set training steps: Train the SFFNet model and classifier head based on samples of known class radiation sources, and calculate the mean activation vector for each class; Open set training steps: Construct a Hybrid-EVT hybrid extremum model based on the activation vectors of the training set; Open set testing steps: Input the test sample into the SFFNet model with frozen parameters to obtain the activation vector of the test sample. Use the Hybrid-EVT hybrid extreme value model to evaluate the tail probability of the distance between the test sample and the candidate class. Use OpenMax recalibration to introduce the probability of unknown channels. If the unknown confidence meets the preset conditions, it is determined to be an unknown radiation source; otherwise, the corresponding known class is output.

[0007] Furthermore, a preferred solution is provided: the frequency-space mapping module has three branches, including: Global branch: Global feature mapping is performed using an improved Swing Transformer; Local branch: used for multi-scale local feature mapping; Wavelet transform branch: The Haar wavelet transform is used to map the original features to high-frequency and low-frequency features in the frequency domain.

[0008] Furthermore, a preferred solution is provided: the improved method of the improved Sw-Transformer is to replace the Sw-Trans block in the Sw-Transformer with a vertical strip convolution.

[0009] Furthermore, a preferred embodiment is provided: the implementation process of the multi-scale dual-representation aligned filter includes: The input features are processed and concatenated using multi-scale vertical bar convolution, and then mapped to a matrix of uniform scale, which serves as the input for the next stage. A domain attention fusion method is adopted, which calculates attention by querying the key values ​​of the other party and itself and then performs feature weighting to achieve feature selection; The final output feature is obtained by concatenating the two outputs.

[0010] Furthermore, a preferred solution is provided: in the third stage of the SFFNet model, learnable weighted fusion introduces learnable weight vectors to normalize and weight features from different branches and combine them with convolutional rectification.

[0011] Furthermore, a preferred solution is provided: the open set training step includes: Calculate the feature center and the set of distances from the sample to the feature center for each known category of radiation source; Arrange the distance set in descending order, adaptively select the tail sample size, and extract the tail sample with the largest distance; Based on the tail samples, the Weibull model and the generalized Pareto distribution model were fitted respectively. Calculate the AIC values ​​of the two distribution models, determine the dynamic weights of the models based on the AIC values, and obtain the mixed extreme value distribution based on the dynamic weights.

[0012] Furthermore, a preferred solution is provided: in the open set testing step, the unknown confidence level meets the preset condition specifically as follows: the probability of outliers in the test sample exceeds a preset threshold.

[0013] Secondly, the purpose of this invention is to propose an open-collection radiation source individual identification system based on spatial-frequency domain fusion and Hybrid-EVT. The system is implemented based on the open-collection radiation source individual identification method based on spatial-frequency domain fusion and Hybrid-EVT as described in any one or more of the above-mentioned schemes. The system includes: Signal acquisition and preprocessing module: used to acquire the raw IQ signal, perform preprocessing, and obtain a time-frequency diagram as the input for radio frequency fingerprint characterization; Feature extraction module: used to extract features from the time-frequency graph based on the SFFNet model to obtain discriminative activation vectors. The SFFNet model includes a three-level segmented network, namely: First stage: Multi-scale spatial features are extracted using ConvNeXt, and after being convolved with 1×1 to unify the number of channels and upsampled to a consistent spatial resolution, they are concatenated to obtain the input for the second stage; The second stage involves constructing four complementary representations through a frequency-space mapping module, including global and local features in the spatial domain, as well as low-frequency and high-frequency features in the frequency domain. The third stage: Semantic alignment and feature selection are performed on local features and high-frequency features, and global features and low-frequency features through two multi-scale dual representation alignment filters, respectively. Then, after two levels of learnable weighted fusion and lightweight classification head processing, the discriminant activation vector is output. Closed-set training module: used to train the SFFNet model and classifier head based on samples of known class radiation sources, and to calculate the mean activation vector for each class; Open set training module: used to build a Hybrid-EVT hybrid extremum model based on the activation vectors of the training set; The open set test module is used to input the test sample into the SFFNet model with frozen parameters to obtain the activation vector of the test sample. The distance between the test sample and the candidate class is evaluated by tail probability through the Hybrid-EVT hybrid extreme value model. The unknown channel probability is introduced by OpenMax recalibration. If the unknown confidence meets the preset conditions, it is determined to be an unknown radiation source; otherwise, the corresponding known class is output.

[0014] 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 the open set radiation source individual identification method based on spatial-frequency domain fusion and Hybrid-EVT according to any one or more of the above-described schemes.

[0015] Fourthly, the present invention aims to provide a computer-readable storage medium for storing a computer program that executes the open-set radiation source individual identification method based on spatial-frequency domain fusion and Hybrid-EVT described in any one or more of the above-described schemes.

[0016] Compared with the prior art, the advantages of the present invention are: To address the core challenge of identifying individual radiation sources in complex electromagnetic environments—namely, the frequent appearance of unknown emission sources and the failure of the closed-set hypothesis—this invention proposes an open-set SEI method based on space-frequency domain fusion and Hybrid-EVT. The method uses SFFNet as its backbone: multi-scale features are extracted using ConvNeXt, and a feature mapping separation module is used to construct global / local representations in the spatial domain and decompose low / high frequencies in the frequency domain using wavelet-based time-frequency decomposition. Subsequently, a multiscale dual-representation alignment filter completes semantic alignment in the time-frequency dimension, and a discriminative embedding is obtained through two-level learnable weighted fusion and a lightweight classification head. On the decision-making side, Hybrid-EVT is proposed: the tail of the class center distance is simultaneously modeled using both Weibull and Generalized Pareto distributions, and adaptively weighted using Akaike's information criterion. Unknown class probabilities are introduced through OpenMax-style recalibration. Experimental results show that this method outperforms existing state-of-the-art methods in both closed-set accuracy and open-set detection, demonstrating promising application prospects.

[0017] This invention is applicable to scenarios that enhance the security of wireless device identity authentication. Attached Figure Description

[0018] 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.

[0019] Figure 1 The radiation source individual open set identification model described in the specific embodiment of the present invention is shown in (a) closed set training, (b) open set training, and (c) open set testing. Figure 2 This is a schematic diagram of the SFFNet structure according to a specific embodiment of the present invention; Figure 3 This is a comparison chart showing the impact of the number of unknown radiation sources on the performance of the OS-SEI system according to a specific embodiment of the present invention; Figure 4This is the confusion matrix under different numbers of unknown radiation sources as described in a specific embodiment of the present invention; Figure 5 This is a visualization of TSNE characteristics under different numbers of unknown radiation sources, as described in a specific embodiment of the present invention. Detailed Implementation

[0020] 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.

[0021] 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.

[0022] 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.

[0023] Implementation Method 1 Reference Figure 1 This implementation method is described below.

[0024] This implementation provides an open-set radiation source individual identification method based on spatial-frequency domain fusion and Hybrid-EVT. First, the raw IQ signal is captured by a Universal Software Radio Peripheral (USRP) receiver. After a series of processing steps, a time-frequency map is generated through time-frequency analysis as the input for the RFF representation. Subsequently, in the feature extraction stage, the model extracts multi-scale spatial features using ConvNeXt as the backbone. The frequency-space mapping module (FMS) constructs two representations in the spatial domain: "global" and "local". The wavelet transform feature decomposition module (WTFD) decouples "low" and "high" frequencies in the frequency domain. Then, the multi-scale dual representation alignment filter (MDAF) completes cross-representation semantic alignment in the time-frequency plane. Finally, the discriminative activation vector is output through WF and a lightweight classification head.

[0025] like Figure 1As shown, the feature extraction and classification process is divided into three stages: (a) closed set training, (b) open set training, and (c) open set testing.

[0026] During the closed-set training phase, SFFNet and the classification head are trained with samples of known categories. Cross-entropy optimization is used to make intra-class compact and inter-class separated, resulting in fixed parameters and closed-set baseline performance for inference.

[0027] In the open set training phase, the mean activation vectors of each class are calculated based on the activation vectors of the training set, and the distance from the sample to the mean activation vectors is calculated. For the tail of each class distance, Weibull and generalized Pareto distributions (GPD) are fitted simultaneously, and a Hybrid-EVT hybrid extreme value model is formed by adaptive weighting with AIC.

[0028] Finally, in the open-set testing phase, the test samples are processed by the frozen SFFNet to obtain activation vectors; Hybrid-EVT is used to evaluate the tail probability of the distance to the candidate class, and "unknown" channel probabilities are introduced through OpenMax recalibration. When the confidence level of unknown is high (or the distance exceeds the threshold), it is judged as unknown; otherwise, the corresponding known class is output.

[0029] The method described in this embodiment specifically includes the following steps: Signal acquisition and preprocessing steps: Acquire the raw IQ signal, perform preprocessing to obtain the time-frequency diagram, which serves as the input for radio frequency fingerprint characterization; Feature extraction steps: Feature extraction is performed on the time-frequency graph based on the SFFNet model to obtain discriminative activation vectors. The SFFNet model includes a three-level segmented network, namely: First stage: Multi-scale spatial features are extracted using ConvNeXt, and after being convolved with 1×1 to unify the number of channels and upsampled to a consistent spatial resolution, they are concatenated to obtain the input for the second stage; The second stage involves constructing four complementary representations through a frequency-space mapping module, including global and local features in the spatial domain, as well as low-frequency and high-frequency features in the frequency domain. The third stage: Semantic alignment and feature selection are performed on local features and high-frequency features, and global features and low-frequency features through two multi-scale dual representation alignment filters, respectively. Then, after two levels of learnable weighted fusion and lightweight classification head processing, the discriminant activation vector is output. Closed-set training steps: Train the SFFNet model and classifier head based on samples of known class radiation sources, and calculate the mean activation vector for each class; Open set training steps: Construct a Hybrid-EVT hybrid extremum model based on the activation vectors of the training set; Open set testing steps: Input the test sample into the SFFNet model with frozen parameters to obtain the activation vector of the test sample. Use the Hybrid-EVT hybrid extreme value model to evaluate the tail probability of the distance between the test sample and the candidate class. Use OpenMax recalibration to introduce the probability of unknown channels. If the unknown confidence meets the preset conditions, it is determined to be an unknown radiation source; otherwise, the corresponding known class is output.

[0030] Implementation Method 2 Reference Figure 2 This implementation method is described below.

[0031] This embodiment is a further illustrative example of the SFFNet model described in the open-set radiation source individual identification method based on space-frequency domain fusion and Hybrid-EVT described in Embodiment 1.

[0032] The structure of the SFFNet model is as follows: Figure 2 As shown: The first stage utilizes ConvNext to extract four-level multi-scale spatial features, denoted as... in, , , and To facilitate cross-scale fusion, interpolation and 1×1 convolution are first used to unify the number of channels. (This implementation method) Then upsample to the same spatial resolution as the lower layer res1. ( and (Input the height and width of the image separately), and then concatenate them along the channel dimension to obtain intermediate features: in express Convolution mapping, This indicates an interpolation operation, while Cat indicates a channel splicing operation. As input for the second stage, while retaining This is used for subsequent residual and detail compensation. This fusion operation effectively combines deep semantic information and surface information, providing more powerful spatial features for subsequent operations.

[0033] The second stage is the feature mapping stage, where the FMS module transforms the raw data containing redundant information into a more representative and discriminative feature space. The FMS module has three branches: a global branch performs global feature mapping using an improved SwingTransformer, a local branch handles multi-scale local feature mapping, and a wavelet transform branch uses Haar wavelet transform to map the raw features to high-frequency and low-frequency features in the frequency domain. Subsequently, two MDAF branches (MDAF_L and MDAF_H) align, select, and combine local and high-frequency features, as well as global and low-frequency features, to obtain a more robust blended feature space.

[0034] The third stage is the feature fusion stage, which uses WF weighted fusion of global and local features, then adds the residuals with middleres, and then fuses them with res1 through a second WF. Finally, the lightweight classification head outputs activation vectors through "pooling-Dropout-fully connected" for subsequent open set recognition.

[0035] Implementation Method 3 This embodiment is a further example illustrating the FMS module in the SFFNet model described in Embodiment 2.

[0036] The FMS module described in this embodiment has three branches, including: Global branch: Global feature mapping is performed using an improved Swing Transformer. The improvement method of the improved Swing Transformer is to replace the Sw-Trans block in the Swing Transformer with a vertical strip convolution. Local branch: used for multi-scale local feature mapping; Wavelet transform branch: The Haar wavelet transform is used to map the original features to high-frequency and low-frequency features in the frequency domain.

[0037] For the multi-scale features extracted in the first stage of the SFFNet model Simultaneously, it constructs complementary four-way representations in the frequency and spatial domains: low frequency ,high frequency Let G be the global spatial coordinates and S be the local spatial coordinates. Let the three mappings be: , FMS performs two tasks: extracting low / high frequency subband features from Haar wavelet decomposition; and constructing global / local spatial mapping branches in parallel, which are then paired with frequency domain components as input to MDAF.

[0038] 1) Global Branch: To enable the Swing Transformer to perform global mapping more effectively and better adapt to the SEI task, this implementation enhances it by designing a global branch. In this branch, the features are first downsampled, then W-Trans blocks are used to divide the window and build global features within each window. However, for information exchange between windows, this implementation abandons the complex Sw-Trans blocks in the Swing Transformer and chooses the more efficient vertical stripe convolution. The vertical stripe convolution is designed as a set of kernels with a size of... and The convolution.

[0039] Specifically, for input features This step first uses The convolution downsamples it, and then it is processed through a W-Trans block: , in This indicates the operation of using the W-Trans block on the feature. This represents the features obtained through window partitioning and window self-attention operations. express Indicates the core size is The convolution. The entire global branch can be described as follows: , in, Indicates window size. This indicates batch normalization. This represents the calculation result of the global branch. This represents the global branch extraction function. This represents the tensor concatenation function.

[0040] 2) Local Branches: To obtain sufficient multi-scale features during the feature mapping stage, this implementation uses classic multi-scale maximum merging (5×5, 9×9, 13×13) for the downsampled features. Multi-scale maximum merging of the downsampled features effectively captures feature information at different scales, thereby enhancing the model's ability to recognize multi-scale features.

[0041] , in, Indicates the bottleneck layer. This represents a local branch extraction function. This indicates a spatial pyramid merging operation. Indicates the kernel size as Maximum pooling, This indicates the output characteristics of the local branch.

[0042] 3) Wavelet transform eigendecomposition To incorporate frequency domain information into the segmented network, this implementation introduces a WTFD for frequency domain feature mapping. This module utilizes convolution and Haar wavelet transform to convert spatial features into four frequency domain components: one low-frequency component and three horizontal, vertical, and diagonal high-frequency components. Then, feature representation learning is performed on the low-frequency information through convolution to obtain low-frequency features. The three high-frequency components are concatenated to form a single high-frequency signal, which is then subjected to nonlinear operations for feature mapping. In this process, the channel dimensionality is reduced, filtering out redundant information as much as possible to obtain high-frequency features. This method fully leverages the Haar wavelet transform for simple and efficient decomposition of image signals, introducing frequency information into the network and achieving a more comprehensive feature representation.

[0043] Specifically, for input First, point convolution operations are performed to increase the non-linearity of the features, thereby generating new features with invariant size. Then, the Haar wavelet transform is used to transform the features into a low-frequency component, a horizontal high-frequency component, a vertical high-frequency component, and a diagonal high-frequency component. For each channel... The first-order Haar wavelet transform is calculated as follows: , in This represents the low-frequency approximation coefficients for channel c. Represents the high-frequency detail coefficients of channel c, where i is the row index and j is the column index. , , , These represent the approximation coefficients, horizontal detail coefficients, vertical detail coefficients, and diagonal detail coefficients for a single channel. Finally, the three high-frequency components are concatenated and then subjected to point convolution for low-dimensional mapping to obtain the final high-frequency and low-frequency features. , in This indicates the operation of WTFD. and These represent high-frequency and low-frequency features, respectively. WTFD decomposes spatial features into high-frequency features with local details and low-frequency features with global characteristics, providing a more comprehensive feature set for segmented networks to consider frequency domain information.

[0044] Implementation Method 4 This embodiment is a further example of the MDAF described in the open set radiation source individual identification method based on space-frequency domain fusion and Hybrid-EVT described in Embodiment 1.

[0045] The implementation process of the multi-scale dual-representation aligned filter includes: 1) The input features are processed and concatenated using multi-scale vertical bar convolution, and then mapped to a matrix of uniform scale as the input for the next stage; 2) Employing a domain attention fusion method, feature selection is achieved by calculating attention scores and weighting features by querying the key values ​​of both the other party and itself; 3) Connect the two outputs to obtain the final output feature.

[0046] The semantic basis of the frequency domain and the spatial domain are inherently different, and direct addition / concatenation can easily introduce semantic mismatch. To achieve semantic alignment and feature selection between frequency and spatial domain features and promote their combination, this implementation introduces MDAF. First, vertical bar convolutions of different scales are used to process each feature. Then, they are concatenated and mapped to a matrix of uniform scale using 1×1 convolutions. , and As input for the next stage, spatial domain features and frequency domain features respectively yield two sets of matrices ( , , )and( , , This paper proposes a Domain Attention Fusion (DAF) method. This method calculates attention by querying key-value pairs of the other party and itself, then performs feature weighting to achieve feature selection. Finally, the two are concatenated as the final output feature. Through DAF, the model can dynamically adjust the attention based on task requirements and feature attributes, giving greater attention and weight to semantically similar features, while suppressing or ignoring semantically dissimilar features. This ensures that the model focuses more on semantically similar features during feature selection, thus achieving better semantic alignment.

[0047] For input The multi-scale orthogonal convolution operation is defined as: , in, Indicates the kernel size as and bar convolution, Representation layer normalization, This represents multi-scale mapping operations. , , It is the matrix obtained after multi-scale orthogonal convolution. The steps for DAF calculation are as follows: , in It is the attention formula. , These are the two outputs of the DAF. Therefore, the MDAF operation can be defined as: .

[0048] Implementation Method 5 This embodiment is a further illustrative example of the learnable weighted fusion (WF) and lightweight classification head processing described in the open set radiation source individual identification method based on spatial-frequency domain fusion and Hybrid-EVT described in Embodiment 1.

[0049] The learnable weighted fusion described in this embodiment introduces a learnable weight vector. and By normalizing and weighting features from different branches and combining this with convolutional rectification, the interpretability and robustness of the fusion process are improved. This can be expressed as: , Finally, the final embedding vector / classification score is output using global pooling + Dropout + fully connected layers.

[0050] Implementation Method Six This embodiment is a further illustrative example of the Hybrid-EVT hybrid extremum model described in the open-set radiation source individual identification method based on space-frequency domain fusion and Hybrid-EVT described in Embodiment 1.

[0051] This implementation introduces a Hybrid-EVT model to support open-set recognition, building upon closed-set training. Ordinary OpenMax uses a single Weibull distribution to model the activation tails of each class to estimate the probability that a sample belongs to a known class. However, the Weibull distribution may be insufficient in capturing complex tail behaviors (such as heavy or super-heavy tails), especially when extreme data distributions are uneven. Based on extreme value theory, the GPD excels at modeling extreme data exceeding high thresholds and adapts to heavy-tail characteristics. The Hybrid-EVT model fits the Weibull and GPD distributions of the tail data by analyzing the activation vectors of the training samples and their class center activation values. Furthermore, it constructs a Hybrid Extreme Value Distribution based on dynamic weighting, leveraging the complementarity of both to improve the robustness and accuracy of tail modeling and provide more flexible tail distribution estimation. The algorithm flow of the Hybrid-EVT model is shown in Table 1.

[0052] Table 1 Hybrid-EVT Model

[0053] Implementation Method Seven This embodiment is a further illustrative example of the open set testing steps described in the open set radiation source individual identification method based on space-frequency domain fusion and Hybrid-EVT described in Embodiment 1.

[0054] During the open-set testing phase, RFF indicates that activation vectors are generated using a pre-trained SFFNet. It is also a feature of the input samples of the Hybrid-EVT model. First, the activation vector is processed using the softmax function. Obtain the closed set confidence vector Next, find the closed set confidence vector. Category index with maximum confidence Obtain the closed-set predicted class of the test sample. Then use the first... Hybrid-EVT model for individual radiation sources and feature center Calculate the probability that the test sample is an outlier. Then, based on the probability of outliers... right To correct this, we introduce the confidence level of the unknown category, resulting in the corrected open-set confidence vector. .

[0055] Implementation Method Eight This embodiment is a further illustrative example of the open-set radiation source individual identification method based on spatial-frequency domain fusion and Hybrid-EVT, as described in any one of Embodiments 1 to 7.

[0056] This implementation is based on a signal dataset from 10 WiFi transmitters collected from two receivers, Rx(1-1) and Rx(14-7). The data is divided into training and test sets in a 7:3 ratio. The WiFi 802.11 a / g modules include the Atheros 5212, 9220, 9280, and 9580 series, while the USRP receivers include models B210, X310, and N210. The WiFi 802.11 a / g transmitters and access points are configured to operate on WiFi channel 13, with a center frequency of 2462 MHz and a bandwidth of 20 MHz. The USRP receivers capture IQ samples at a sampling rate of 25 Msps, with a center frequency of 2462 MHz and a duration of 0.512 seconds.

[0057] After preprocessing and feature generation, the acquired signals were categorized by different emission sources and stored in separate folders. All time-frequency images input to the model were adjusted to 128×128. The model was trained on a GPU using a stochastic gradient descent (SGD) optimizer with a momentum of 0.9 and a weight decay of 5e-4. The training epochs were 100, with a weight decay coefficient of 0.00001 and a batch size of 32. The initial learning rate was 0.001, with an exponential decay of 0.9 every two epochs. Training was terminated early if the accuracy on the validation set did not improve or decreased after five consecutive evaluations. Notably, the model used mixed accuracy training.

[0058] To evaluate the robustness of the OS-SEI system as the number of unknown radiation sources increases, this implementation designed a series of experiments. Five known radiation sources were fixed, and the number of unknown radiation sources was gradually increased (from one to five) to observe its impact on system performance. The experimental results are as follows: Figure 3 As shown.

[0059] from Figure 3 As can be seen, the open set accuracy improved from 96.45% with one unknown radiation source to 97.29% with three unknown radiation sources, then decreased slightly, but the decrease was relatively gradual, and the overall performance remained at a high level. The initial improvement may be because the introduction of a small number of unknown class samples prompted the model to better learn the boundary between known and unknown classes. As the number of unknown classes further increased, the complexity of the feature space increased, leading to some known class samples being misclassified as unknown, resulting in a slight decrease in accuracy, but it remained above 97% overall, indicating that the model has good robustness. The AUEOC value remained above 99%, indicating that the model's ability to distinguish between known and unknown classes was very stable, and the Hybrid-EVT model's ability to model the tail distribution of activation vectors remained excellent even with an increase in the number of unknown classes. As the number of unknown classes increased to five, UCDR decreased slightly, possibly due to the increased diversity of the unknown class distribution, causing the activation vectors of some unknown class samples to be closer to those of known classes, thus being misclassified as known classes.

[0060] To further analyze the impact of increasing the number of unknown radiation sources on classification performance, this implementation uses three scenarios with 1, 3, and 5 unknown radiation sources as examples to demonstrate the confusion matrix under different scenarios, such as... Figure 4 As shown, the model has a very high detection rate for unknown classes, with misclassifications mainly concentrated in known classes. As the number of unknown radiation sources gradually increases, the misclassification rate for unknown radiation sources rises, but the system still maintains high classification accuracy.

[0061] To visually demonstrate the impact of increasing the number of unknown radiation sources on the feature space, this implementation uses t-SNE to reduce the dimensionality of the activation vectors for visualization. t-SNE diagrams are generated for 1, 3, and 5 unknown radiation sources, as shown below. Figure 5 As shown, when UUC=1, the known classes (KKC 1 to KKC 5) form compact clusters, while the unknown classes (UUC) are distributed outside the clusters, with clear inter-class boundaries. When UUC increases to 3, the feature distribution becomes wider, but the inter-class boundaries remain clear. As the number of UUC samples further increases, some feature points of some UUCs move closer to the known class clusters, resulting in slight overlap. However, overall, the system can effectively distinguish between known and unknown classes, demonstrating good robustness.

[0062] The above analysis shows that as the number of unknown radiation sources gradually increases, the overall performance remains stable, indicating that the model effectively adapts to the increase in the number of unknown types and is suitable for dynamic radiation source identification scenarios.

[0063] Implementation Method Nine This embodiment is a further illustrative example of the open-set radiation source individual identification method based on spatial-frequency domain fusion and Hybrid-EVT, as described in any one of Embodiments 1 to 7.

[0064] Ablation experiments: Table 2 shows the SEI results after deleting each module. Here, Indicates a global branch. The symbols represent local branches, WTFD represents wavelet transform eigenvalue decomposition (low / high frequency components), and MDAF represents multi-scale dual representation alignment filter. For experiments removing low-frequency or high-frequency components from SFFNet, the corresponding feature alignment of MDAF is also removed simultaneously. To verify the effectiveness of MDAF, this implementation also removes MDAF and directly concatenates and adds frequency and spatial domain features. Components removed in the ablation experiments are indicated by "(w / o)".

[0065] Overall, the complete model exhibits the best balance in performance. Removing WTFD_H results in a greater performance degradation than removing WTFD_L, highlighting the importance of high-frequency details. or It will slightly reduce accuracy; it is worth noting that removing The unknown category recognition rate dropped to 97.92% (from 100%), indicating the contribution of local branches to unknown detection. Without MDAF, the closed set accuracy decreased the most (to 92.08%), and the F1 score dropped by 2.89 points, although the open set accuracy was higher (99.33%). This suggests that the lack of cross-domain alignment biases decisions towards the unknown category, impairing overall balance. Removing WF caused a general decline in all metrics, while removing hybrid EVT significantly reduced open set processing capability (open set accuracy 94.83%, unknown recognition rate 99.17%), confirming the advantages of hybrid tail modeling.

[0066] Table 2 Component Contribution Analysis

[0067] Comparison methods: CNN / RNN / Transformer baseline, spatial or frequency domain-only methods, original OpenMax / Softmax thresholding, etc.

[0068] The experimental results are shown in Table 3 when the number of KKCs is 8 and the number of UUCs is 2. Table 3 shows that SoftMax achieves an accuracy of 99.67% in closed-set scenarios, but it cannot handle unknown samples. SoftMax-T and OpenMax introduce OSR capabilities through thresholding and calibration mechanisms, but their open-set accuracies are only 92.42% and 92.17%, respectively, indicating limited ability to detect unknown samples. G-SoftMax-T and G-OpenMax improve the robustness of the algorithms by introducing data augmentation based on generative adversarial networks, achieving 100% recognition rate for unknown classes and open-set accuracies of 93.17% and 94.33%, respectively, but are still limited by feature discriminativeness and boundary modeling accuracy. The classification boundary of the OvA classifier for a certain training radiation source individual is determined by the distribution of other training radiation source individuals. When the number of training radiation source individuals is small, the classification boundary of the OvA classifier is not tight enough, and it is easy to misidentify unknown radiation source individuals as known radiation source individuals. ARPL optimizes open-space risk through adversarial constraints, achieving an open set accuracy of 93.33% and an F1 score of 94.45%, with an unknown class recognition rate of 100%. However, its overall performance is still inferior to the proposed method.

[0069] Table 3 Performance Comparison of Methods

[0070] The proposed method achieves the most balanced and optimal overall performance, demonstrating stronger feature discrimination ability and robustness in open set detection. This is attributed to the highly discriminative feature extraction of SFFNet and the accurate boundary modeling of the Hybrid-EVT distribution, which adapts to complex distributions, proving its effectiveness and robustness in the OS-SEI task.

[0071] It is understood that the present invention has been described through some embodiments, and 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 method for identifying individual open-set radiation sources based on spatial-frequency domain fusion and Hybrid-EVT, characterized in that, Includes the following steps: Signal acquisition and preprocessing steps: Acquire the raw IQ signal, perform preprocessing to obtain the time-frequency diagram, which serves as the input for radio frequency fingerprint characterization; Feature extraction steps: Feature extraction is performed on the time-frequency graph based on the SFFNet model to obtain discriminative activation vectors. The SFFNet model includes a three-level segmented network, namely: First stage: Multi-scale spatial features are extracted using ConvNeXt, and after being convolved with 1×1 to unify the number of channels and upsampled to a consistent spatial resolution, they are concatenated to obtain the input for the second stage; The second stage involves constructing four complementary representations through a frequency-space mapping module, including global and local features in the spatial domain, as well as low-frequency and high-frequency features in the frequency domain. The third stage: Semantic alignment and feature selection are performed on local features and high-frequency features, and global features and low-frequency features through two multi-scale dual representation alignment filters, respectively. Then, after two levels of learnable weighted fusion and lightweight classification head processing, the discriminant activation vector is output. Closed-set training steps: Train the SFFNet model and classifier head based on samples of known class radiation sources, and calculate the mean activation vector for each class; Open set training steps: Construct a Hybrid-EVT hybrid extremum model based on the activation vectors of the training set; Open set testing steps: Input the test sample into the SFFNet model with frozen parameters to obtain the activation vector of the test sample. Use the Hybrid-EVT hybrid extreme value model to evaluate the tail probability of the distance between the test sample and the candidate class. Use OpenMax recalibration to introduce the probability of unknown channels. If the unknown confidence meets the preset conditions, it is determined to be an unknown radiation source; otherwise, the corresponding known class is output.

2. The method for identifying individual open-set radiation sources based on spatial-frequency domain fusion and Hybrid-EVT according to claim 1, characterized in that, The frequency-space mapping module has three branches, including: Global branch: Global feature mapping is performed using an improved Swing Transformer; Local branch: used for multi-scale local feature mapping; Wavelet transform branch: The Haar wavelet transform is used to map the original features to high-frequency and low-frequency features in the frequency domain.

3. The method for identifying individual open-set radiation sources based on spatial-frequency domain fusion and Hybrid-EVT according to claim 2, characterized in that, The improvement method of the improved Swin Transformer is to replace the Sw-Trans block in the Swin Transformer with vertical strip convolution.

4. The method for individual identification of open-set radiation sources based on spatial-frequency domain fusion and Hybrid-EVT according to claim 1, characterized in that, The implementation process of the multi-scale dual-representation aligned filter includes: The input features are processed and concatenated using multi-scale vertical bar convolution, and then mapped to a matrix of uniform scale, which serves as the input for the next stage. A domain attention fusion method is adopted, which calculates attention by querying the key values ​​of the other party and itself and then performs feature weighting to achieve feature selection; The final output feature is obtained by concatenating the two outputs.

5. The method for identifying individual open-set radiation sources based on spatial-frequency domain fusion and Hybrid-EVT according to claim 1, characterized in that, In the third stage of the SFFNet model, learnable weighted fusion introduces learnable weight vectors to normalize and weight features from different branches and combine them with convolutional rectification.

6. The method for identifying individual open-set radiation sources based on spatial-frequency domain fusion and Hybrid-EVT according to claim 1, characterized in that, The open set training steps include: Calculate the feature center and the set of distances from the sample to the feature center for each known category of radiation source; Arrange the distance set in descending order, adaptively select the tail sample size, and extract the tail sample with the largest distance; Based on the tail samples, the Weibull model and the generalized Pareto distribution model were fitted respectively. Calculate the AIC values ​​of the two distribution models, determine the dynamic weights of the models based on the AIC values, and obtain the mixed extreme value distribution based on the dynamic weights.

7. The method for individual identification of open-set radiation sources based on spatial-frequency domain fusion and Hybrid-EVT according to claim 1, characterized in that, In the open set testing step, the unknown confidence level meets the preset condition specifically as follows: the probability of outliers in the test sample exceeds a preset threshold.

8. An open-set radiation source individual identification system based on spatial-frequency domain fusion and Hybrid-EVT, characterized in that, The system is implemented based on the open-set radiation source individual identification method based on space-frequency domain fusion and Hybrid-EVT as described in any one of claims 1-7. The system includes: Signal acquisition and preprocessing module: used to acquire the raw IQ signal, perform preprocessing, and obtain a time-frequency diagram as the input for radio frequency fingerprint characterization; Feature extraction module: used to extract features from the time-frequency graph based on the SFFNet model to obtain discriminative activation vectors. The SFFNet model includes a three-level segmented network, namely: First stage: Multi-scale spatial features are extracted using ConvNeXt, and after being convolved with 1×1 to unify the number of channels and upsampled to a consistent spatial resolution, they are concatenated to obtain the input for the second stage; The second stage involves constructing four complementary representations through a frequency-space mapping module, including global and local features in the spatial domain, as well as low-frequency and high-frequency features in the frequency domain. The third stage: Semantic alignment and feature selection are performed on local features and high-frequency features, and global features and low-frequency features through two multi-scale dual representation alignment filters, respectively. Then, after two levels of learnable weighted fusion and lightweight classification head processing, the discriminant activation vector is output. Closed-set training module: used to train the SFFNet model and classifier head based on samples of known class radiation sources, and to calculate the mean activation vector for each class; Open set training module: used to build a Hybrid-EVT hybrid extremum model based on the activation vectors of the training set; The open set test module is used to input the test sample into the SFFNet model with frozen parameters to obtain the activation vector of the test sample. The distance between the test sample and the candidate class is evaluated by tail probability through the Hybrid-EVT hybrid extreme value model. The unknown channel probability is introduced by OpenMax recalibration. If the unknown confidence meets the preset conditions, it is determined to be an unknown radiation source; otherwise, the corresponding known class is output.

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 the open set radiation source individual identification method based on spatial-frequency domain fusion and Hybrid-EVT 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 the open-set radiation source individual identification method based on spatial-frequency domain fusion and Hybrid-EVT as described in any one of claims 1-7.

Images