An open set communication radiation source individual identification method, system, device and storage medium

Abstract

This invention relates to the fields of network security and radio transmission signal processing technology, specifically to a method, system, device, and storage medium for identifying individual radio emission sources in open set communication scenarios. The method includes acquiring a discrete baseband complex signal transmitted by a radio signal source; obtaining backbone intermediate features from the discrete baseband complex signal; obtaining low-frequency structural components and high-frequency detail components from the backbone intermediate features; obtaining low-frequency enhancement features from the low-frequency structural components; obtaining high-frequency enhancement features from the high-frequency detail components; obtaining fusion features from the backbone intermediate features, low-frequency enhancement features, and high-frequency enhancement features; generating and outputting discriminant embedding and known class prediction scores based on the fusion features; calculating unknown scores based on the discriminant embedding; and combining the unknown scores and known class prediction scores to identify the category of the radio signal source. This solution can improve the known recognition capability and unknown rejection robustness of radio signal sources in open set scenarios.

Description

Technical Field

[0001] This invention relates to the fields of Internet security services and radio transmission signal processing technology, specifically to an open-source communication radiation source individual identification method, system, device, and storage medium. Background Technology

[0002] With the rapid proliferation of wireless communication and IoT devices, the types of signals and the number of transmitters in the electromagnetic spectrum continue to grow, leading to increasing competition for spectrum resources and growing wireless security risks. In complex and open wireless environments, problems such as unauthorized frequency occupation and transmission, identity spoofing and impersonation, malicious interference, and the infiltration of abnormal devices occur frequently, posing significant challenges to wireless access control, situational awareness, and regulatory evidence collection. Therefore, how to achieve transmitter identification and tracing at the physical layer without relying on high-level protocol interactions or key systems has become an important research direction in the fields of wireless security and spectrum monitoring.

[0003] Specific Emitter Identification (SEI) of communication radiation sources utilizes the radio frequency fingerprint (RFF) formed by differences and non-ideal effects in the transmitter's radio frequency front-end devices. This enables individual differentiation and traceability among devices of the same standard, providing traceable evidence for abnormal transmissions, spoofing, and unauthorized access. Given that radio frequency signals are unavoidable, difficult to forge, and independent of signal content, SEI is of significant value. However, the key challenges of SEI lie in the fact that fingerprint clues are typically weak and strongly coupled with complex channel distortions and noise disturbances. The observational distribution of the same device is prone to significant drift under different receiving conditions. Furthermore, the differences between different devices often manifest as fine-grained perturbations in the time or frequency domains across scales, making the identification of reliable and generalizable fingerprint representations a core challenge.

[0004] In recent years, deep learning has driven advancements in SEI (Search Engine Imaging) in terms of accuracy and condition adaptability across multiple categories. However, most existing methods still assume that the training and testing categories are the same, i.e., they are based on the closed-set assumption. In real-world wireless monitoring, this assumption is often difficult to satisfy. During specific testing phases, new emission sources not covered in the training phase may appear, and the model may easily force unknown samples to be classified into a known category, giving an overconfident and erroneous judgment. At the same time, noise disturbances, channel variations, and device feature drift can further amplify this risk. Therefore, a more practically applicable goal is open-set SEI: while ensuring the ability to identify known emission sources, it can reliably reject unknown / rogue emission sources and maintain a stable false alarm-false negative tradeoff under different openness conditions.

[0005] To address closed-set identification, researchers have proposed various deep learning SEI architectures and training strategies. These include hybrid CNN-Transformer architectures (combining convolutional neural networks and Transformers) to balance local modeling and global dependencies, hybrid augmentation to improve generalization, domain generalization and robust training to mitigate the impact of cross-channel and label noise, and meta-learning and feature augmentation to support rapid adaptation for small samples. While these methods effectively improve the ability to distinguish known categories, their optimization objectives often remain focused on closed-set classification accuracy, lacking direct constraints on overconfidence in categories not encountered during training. Furthermore, as the risk in the open space increases, the discrimination boundary becomes more easily squeezed, leading to a simultaneous exacerbation of false acceptances of unknown categories and false rejections of known categories.

[0006] To address the challenges of open sets, existing open set SEI methods can be summarized into two paradigms: one focuses on reconstruction or distribution modeling, using reconstruction errors, generated outliers, or intra-class statistical models as the basis for anomalies. This approach is generally more adaptable to noise and distribution drift, but often comes with additional training or inference overhead and is sensitive to threshold calibration. The other focuses on discriminative representation learning and decision boundary construction, achieving unknown rejection through more separable embedding spaces and distance rules. This approach has higher inference efficiency but places higher demands on feature quality and cross-condition stability. Therefore, the upper limit of open set rejection capability still strongly depends on whether the representation can stably expose unknown cues under different conditions.

[0007] In summary, existing methods for identifying individual sources of communication radiation sources have not fully utilized the multi-scale discriminative structure of radio frequency fingerprints in the frequency domain, where the influence of channel and noise interference on fingerprint clues varies significantly across different frequency bands. Single-path representations are insufficient to continuously provide stable known and unknown boundaries as openness increases. Summary of the Invention

[0008] The purpose of this invention is to provide an open set communication radiation source individual identification method, system, device and storage medium to solve the technical problem that the existing communication radiation source individual identification methods do not make full use of the multi-scale discrimination structure of radio frequency fingerprints in the frequency domain and are difficult to continuously provide stable boundaries when the openness is increased.

[0009] In a first aspect, the present invention provides a method for identifying an individual radiation source in an open set communication system, comprising the following steps: Acquire the discrete baseband complex signal transmitted by the wireless signal transmitter, and obtain the backbone intermediate features based on the discrete baseband complex signal; Input the backbone intermediate features into the Open Set Communication Radiation Source Individual Identification Network. The Open Set Communication Radiation Source Individual Identification Network performs the following operations: obtain low-frequency structural components and high-frequency detail components based on the backbone intermediate features; obtain low-frequency enhancement features based on the low-frequency structural components; obtain high-frequency enhancement features based on the high-frequency detail components; obtain fusion features based on the backbone intermediate features, low-frequency enhancement features, and high-frequency enhancement features; and generate and output discriminative embeddings and known class prediction scores based on the fusion features. The unknown score is calculated based on the discriminant embedding, and the category of the wireless signal source is identified by combining the unknown score with the known class prediction score.

[0010] The significant advantages of this invention are as follows: This solution addresses the problem in existing technologies that do not fully utilize the multi-scale discriminative structure of RF fingerprints in the frequency domain. By performing frequency decomposition on the intermediate features of the backbone, the mixed features are split into low-frequency structural components and high-frequency detail components. These components are then enhanced separately and fused with the backbone features to obtain a fused feature. Since the low-frequency structural components and high-frequency detail components carry different discriminative information, and the influence of channel and noise on different frequency bands varies, this technical solution can separate and enhance the originally mixed-distributed discriminative information at different scales, thereby making fuller use of the multi-scale discriminative structure of RF fingerprints in the frequency domain and improving the relevance and effectiveness of feature representation. Simultaneously, addressing the problem in existing technologies that struggle to maintain a stable known / unknown class distinction boundary when openness increases, this invention, by fusing high-frequency enhanced features, low-frequency enhanced features, and backbone features, can dynamically adjust the contribution of each component based on the frequency domain feature differences of the input samples. This reduces representational shifts caused by factors such as channel variations and noise disturbances, thus helping to maintain a more stable known / unknown class distinction boundary when openness increases, improving the known recognition capability and unknown rejection robustness in open set scenarios.

[0011] Furthermore, the step of obtaining the low-frequency structural components and high-frequency detail components based on the intermediate features of the main trunk includes: The bandwidth coefficient is obtained based on the intermediate characteristics of the backbone, and the effective half-bandwidth is calculated based on the bandwidth coefficient. Construct differentiable center-limited masks and complementary masks based on the effective half-bandwidth; The frequency domain representation is obtained based on the intermediate features of the backbone, and the frequency domain representation is divided into low-frequency and high-frequency components based on the differentiable center band-limited mask and complementary mask. The low-frequency frequency domain components are transformed into low-frequency structural components, and the high-frequency frequency domain components are transformed into high-frequency detail components.

[0012] In open wireless environments, transmitter fingerprint clues and receiver conditional perturbations often superimpose on the feature spectrum in different frequency bands. Directly modeling the intermediate features of the time-domain backbone might misinterpret conditionally related spectral changes as a basis for discrimination, thus weakening the generalization ability across channels and noise conditions. Therefore, a frequency domain decomposition module is used to decompose the intermediate features of the backbone. A sample-adaptive center band-limited mask is used to decouple the intermediate features of the backbone into two complementary components, which respectively characterize the low-frequency structure of the center band-limited component and the high-frequency details of the complementary residual, providing structured input for subsequent enhancement and fusion operations.

[0013] Furthermore, the steps of obtaining low-frequency enhancement features based on low-frequency structural components and obtaining high-frequency enhancement features based on high-frequency detail components include: Build a low-frequency expert database and a high-frequency expert database, and generate low-frequency expert routes and high-frequency expert routes; Calculate the weight of low-frequency routes based on low-frequency expert routes, and calculate the weight of high-frequency routes based on high-frequency expert routes. Residual updates for low-frequency structural components are calculated based on low-frequency routing weights and a low-frequency expert database, while residual updates for high-frequency detail components are calculated based on high-frequency routing weights and a high-frequency expert database. Calculate low-frequency sparse weights based on low-frequency routing weights, and calculate high-frequency sparse weights based on high-frequency routing weights. Low-frequency enhancement features are obtained by weighted calculation based on the residual update of low-frequency sparse weights and low-frequency structural components, and high-frequency enhancement features are obtained by weighted calculation based on the residual update of high-frequency sparse weights and high-frequency detail components.

[0014] Since channel distortion and noise perturbations conditionally alter the available information density of the two components, the contributions of the low-frequency structural component and the high-frequency detail component to discrimination and rejection exhibit a clear sample dependence. Therefore, this scheme introduces gated sparse expert enhancements to both the high-frequency and low-frequency components, and uses adaptive fusion gating to inject the enhanced information into the backbone representation in a stable manner, resulting in low-frequency and high-frequency enhanced features for subsequent classification and rejection.

[0015] Furthermore, the step of obtaining the fused features based on the backbone intermediate features, low-frequency enhancement features, and high-frequency enhancement features includes: Fusion weights are generated based on the main intermediate features, low-frequency enhancement features, and high-frequency enhancement features. The fused features are obtained by weighting and fusing the backbone intermediate features, low-frequency enhanced features, and high-frequency enhanced features based on the fusion weights.

[0016] The beneficial effects of the above technical solution are that by generating fusion weights based on the main intermediate features, low-frequency enhanced features and high-frequency enhanced features and performing adaptive weighted fusion, the contribution ratio of the three types of information can be dynamically adjusted for different input samples. This allows the model to retain the stability of the main representation while introducing high-frequency and low-frequency enhanced information as needed, avoiding the excessive dominance of one type of feature due to a fixed fusion method. This results in more consistent fusion features and improves the consistency and stability of subsequent known class classification and unknown rejection.

[0017] Furthermore, the training steps of the open set communication radiation source individual identification network include: Obtain discrete baseband complex signal sample data and construct a training set of known classes; Discriminant embeddings and predicted scores for known classes are output from the training set of known classes; Based on discriminative embedding and known class prediction scores, a classification loss, angular margin metric constraint, and routing auxiliary regularization term are constructed, and an overall training objective is constructed based on the classification loss, angular margin metric constraint, and routing auxiliary regularization term. Optimize the parameters of the open set communication radiation source individual identification network according to the overall training objectives.

[0018] To suppress route collapse and promote expert load balancing, a route-aiding term is introduced during training. This term constrains route allocation to tend towards equilibrium on mini-batch samples and regularizes route uncertainty. Optimization is achieved by jointly applying classification loss, angular margin metric constraints, and the route-aiding regularization term. This improves both the intra-class compactness and inter-class separability of known class representations, enhancing the discriminative power of known classes. Furthermore, it suppresses route collapse and promotes load balancing among different experts, enabling multiple augmentation experts to complement each other, thereby improving network training stability and the generalization ability of representations in open-set scenarios.

[0019] Furthermore, the step of calculating the unknown score based on the discriminant embedding includes: Construct a known class prototype and calculate the distance to the embedding in the known class prototype; The unknown score is calculated based on the distance between the discriminant embedding and the known class prototype.

[0020] Furthermore, the method of identifying the category of a wireless signal emission source by combining the unknown score and the known class prediction score includes: When the unknown score is greater than the preset unknown score threshold, the wireless signal source is considered to be an unknown and unidentifiable category. When the unknown score is not greater than the preset unknown score threshold, the wireless signal transmitter is considered to be of the known class, and the known class label of the wireless signal transmitter is output based on the predicted score of the known class.

[0021] By decoupling known class classification and unknown rejection classification, the known class prediction score is used only for known class discrimination, while rejection is entirely determined by the unknown score of the prototype distance. This decoupling of known class classification decisions from unknown rejection decisions avoids the high-confidence output of the classification branch being directly passed as an acceptance decision, thereby improving the reliability of unknown emission source detection and making the threshold calibration process clearer and the deployment decision rules more stable.

[0022] Secondly, the present invention provides an open set communication radiation source individual identification system for implementing the above-mentioned open set communication radiation source individual identification method, comprising: A wireless signal acquisition unit is used to acquire the discrete baseband complex signal of the wireless signal under test. The backbone intermediate feature extraction unit is used to extract backbone intermediate features based on discrete baseband complex signals; The discriminant prediction unit is used to generate discriminant embeddings and known class prediction scores based on the intermediate features of the backbone. A wireless signal source identification unit is used to identify wireless signal sources based on discriminant embedding and known class prediction scores.

[0023] Thirdly, the present invention provides an open set communication radiation source individual identification device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above-described open set communication radiation source individual identification method.

[0024] Fourthly, the present invention provides a computer-readable storage medium containing a computer program, wherein the computer program is stored thereon, and when the computer program is executed by one or more processors, implements the steps of the above-described method for identifying an individual open set communication radiation source. Attached Figure Description

[0025] Figure 1 This is a flowchart of an open set communication radiation source individual identification method according to an embodiment of the present invention; Figure 2 This is an architecture diagram of the open set communication radiation source individual identification method in an embodiment of the present invention; Figure 3 This is a diagram of the frequency routing adaptive expert enhancement structure in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the complementary relationship between the differentiable center band-limited mask and the complementary mask in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the effect of the temperature coefficient on the transition zone between the differentiable center band-limited mask and the complementary mask in an embodiment of the present invention; Figure 6This is the ROC curve of the open set communication radiation source individual identification method under a 9:1 openness setting in this embodiment of the invention; Figure 7 This is the ROC curve of the open set communication radiation source individual identification method under the 9:4 openness setting in this embodiment of the invention; Figure 8 This is the ROC curve of the open set communication radiation source individual identification method under the 9:8 openness setting in this embodiment of the invention; Figure 9 This is a router preference heatmap in an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, a clear and complete description will be provided below in conjunction with the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0027] See appendix Figure 1 and attached Figure 2 The method for identifying an individual radiation source in an open set communication system includes the following steps: S1. Acquire discrete baseband complex signals transmitted by wireless signal transmitters in an open environment. The discrete baseband complex signals can be acquired by wireless signal monitoring equipment or by sampling wireless signal segments from a publicly available dataset. Each discrete baseband complex signal consists of an in-phase component I and a quadrature component Q, and is represented in the form of an I / Q sequence.

[0028] In an open wireless environment, the discrete baseband complex signal of the acquired wireless signal can be uniformly represented as:

[0029] In the formula, It is a discrete baseband complex signal for wireless signals. The effects on the propagation channel and the receiving link include multipath fading, frequency selectivity, amplitude and phase distortion, and receiver response. For the first The equivalent mapping of each transmitter's RF front-end link is used to characterize stable individual features introduced by manufacturing process differences and device non-idealities, such as power amplifier nonlinearity, IQ imbalance, phase noise, frequency offset, and sampling clock skew. These subtle but repeatable distortions together constitute the physical source of the Radio Frequency Fingerprint (RFF). This refers to the ideal transmission signal sequence under the same communication standard. It is additive noise.

[0030] S2. Extracting the backbone intermediate features from the discrete baseband complex signal; specifically, the discrete baseband complex signal is processed by a pre-trained backbone feature extraction network to extract its features, resulting in channel-priority sequence features. The specific details of extracting the features of the discrete baseband complex signal through the backbone feature extraction network are existing technologies and will not be elaborated here.

[0031] The extracted sequence features are processed through an adaptive pooling Transformer encoding operation to perform context modeling, resulting in context-enhanced backbone intermediate features. The expression for the backbone intermediate features is as follows:

[0032] In the formula, The main trunk features For the set of real numbers, For small batch sample size, The number of feature channels, is the length of the feature sequence.

[0033] S3. Construct an open set communication radiation source individual identification network based on frequency routing and adaptive expert enhancement, including a frequency decomposition module, a dual-component expert enhancement module, a fusion gating module, and an output module. Input the backbone intermediate features to the open set communication radiation source individual identification network. The frequency decomposition module is used to decompose the backbone intermediate features to obtain low-frequency structural components and high-frequency detail components. The dual-component expert enhancement module is used to obtain low-frequency enhanced features based on the low-frequency structural components and high-frequency enhanced features based on the high-frequency detail components. The fusion gating module is used to obtain fused features based on the backbone intermediate features, low-frequency enhanced features, and high-frequency enhanced features. The output module is used to generate and output discriminant embedding and known class prediction scores based on the fused features.

[0034] In open wireless environments, transmitter fingerprint clues and receiver conditional perturbations often superimpose on the feature spectrum in different frequency bands. Directly modeling the intermediate features of the time-domain backbone might misinterpret conditionally related spectral changes as a basis for discrimination, thus weakening the generalization ability across channels and noise conditions. Therefore, a frequency domain decomposition module is used to decompose the intermediate features of the backbone. A sample-adaptive center band-limited mask is used to decouple the intermediate features of the backbone into two complementary components, which respectively characterize the low-frequency structure of the center band-limited component and the high-frequency details of the complementary residual, providing structured input for subsequent enhancement and fusion operations.

[0035] As attached Figure 3 As shown, the decomposition steps of the frequency domain decomposition module include: A1. Obtain the bandwidth coefficient based on the intermediate features of the backbone; specifically, process the intermediate features of the backbone using a lightweight... Conv convolutional projection is used for channel alignment, and the bandwidth coefficient is obtained by global statistical prediction based on the intermediate features of the projection backbone. The specific details of obtaining the bandwidth coefficient through convolutional projection are existing technologies and will not be elaborated here.

[0036] A2. Calculate the effective half-bandwidth based on the bandwidth coefficient; to improve the stability of bandwidth control, the bandwidth control quantity is aggregated in the channel dimension to form a sample-level scalar, and the frequency band mask is shared among the channels of the same sample. First, define the search step size:

[0037] In the formula, For the search step size, For piecewise constants, The length of the feature sequence; The effective half-bandwidth of the b-th sample is obtained based on the search step size and bandwidth coefficient:

[0038] In the formula, The effective half bandwidth of the b-th sample, Indicates the first The bandwidth coefficient of each sample, b∈{1,2,…,B} is the sample index in the mini-batch. Unlike discretization search, this embodiment keeps the half-bandwidth of the sample as a continuous quantity to achieve end-to-end trainability.

[0039] A3. Construct a differentiable center-limited mask and a complementary mask based on the effective half-bandwidth; the calculation formula for the differentiable center-limited mask is:

[0040]

[0041] In the formula, This indicates that the b-th sample is at frequency point Differentiable center band-limited mask value at the location, For frequency point index, For center frequency index, For the floor operation, The length of the feature sequence. For the Sigmoid function, This represents the mask smoothing temperature coefficient.

[0042] The formula for calculating complementary masks is:

[0043] In the formula, This represents the complementary mask value of the b-th sample at frequency point f.

[0044] When the temperature coefficient approaches 0, the mask will approximate a binary central band-limited window, thus ensuring differentiability during the training phase while maintaining approximately hardened bandwidth boundaries. This allows gradients to be backpropagated from downstream task loss to the bandwidth coefficient, thereby achieving end-to-end joint optimization. Figure 4-5 The complementary relationship between the micronable center band-limited mask and the complementary mask, as well as the effect of the temperature coefficient on the transition zone, are visually demonstrated.

[0045] A4. Based on the intermediate features of the backbone, a frequency domain representation is obtained. Then, based on the differentiable central band-limited mask, complementary mask, and frequency domain representation, the frequency band is divided to obtain low-frequency and high-frequency components. Specifically, the intermediate features of the backbone are subjected to a discrete Fourier transform along the length dimension of the feature sequence to obtain the frequency domain representation of the intermediate features of the backbone.

[0046] In the formula, The frequency domain representation of the main intermediate features. This is the Discrete Fourier Transform operation. The main intermediate features; The low-frequency components are obtained by dividing the frequency domain representation using a differentiable center band-limited mask:

[0047] In the formula, For low-frequency components, For element-wise multiplication, For a differentially oriented band-limited mask; The frequency domain component is obtained by dividing the frequency domain representation using complementary masks:

[0048] In the formula, For high-frequency components, For complementary masks.

[0049] A5. Transform the low-frequency frequency domain components into low-frequency structure components, and the high-frequency frequency domain components into high-frequency detail components. Specifically, the low-frequency and high-frequency frequency domain components are respectively subjected to inverse Fourier transforms to return to the time domain to obtain the low-frequency structure components and high-frequency detail components. The calculation formula is as follows:

[0050]

[0051] In the formula, These are low-frequency structural components, primarily representing relatively smooth and stable structural information. This is the inverse Fourier transform operation. For modulo operation, The high-frequency detail components primarily represent local perturbations and fine-grained differences. This step decomposes the original mixed features into two subspace representations with different discriminative properties, providing an input basis for subsequent differential enhancement. During training, to ensure consistency and numerical stability, the inverse transformation result is mapped to real-valued features for use by subsequent network modules.

[0052] The frequency decomposition described above has decoupled the high-frequency and low-frequency structures at the feature level, but relying solely on static decomposition is still insufficient to cover the diverse conditions in open wireless environments. Since channel distortion and noise disturbances conditionally alter the available information density of the two components, the contributions of the low-frequency structure component and the high-frequency detail component to discrimination and rejection exhibit a clear sample dependence. Therefore, this scheme introduces gated sparse expert enhancements to both the high-frequency and low-frequency components, and uses adaptive fusion gating to inject the enhancement information into the backbone representation in a stable manner, obtaining low-frequency and high-frequency enhancement features for subsequent classification and rejection. Specifically, low-frequency and high-frequency expert databases are constructed for the low-frequency and high-frequency enhancement features, each containing multiple lightweight enhancement experts. In this embodiment, an expert database consisting of four lightweight enhancement experts is constructed for each component.

[0053] Here, "expert" refers to a lightweight network branch capable of enhancing input features; "expert pool" refers to a collection of multiple experts connected in parallel with different structural parameters and enhancement preferences. Different experts employ different convolutional kernel sizes and dilation rates, thus corresponding to different receptive field scales and feature enhancement emphases. Each expert uses an ELPlus architecture, introducing multi-scale inductive biases in both channel and temporal dimensions: firstly, modeling local channel correlations and achieving channel recalibration through ECA channel attention; secondly, expanding the sequence-dimensional receptive field through depthwise-temporal dilated convolutions, and combining pointwise mapping with nonlinear transformations to form residual updates. Simultaneously, a learnable scaling factor is introduced to suppress excessive perturbations in the early stages of training, improving optimization stability. To construct an interpretable function space, the four experts employ different combinations of channel correlation scales and temporal receptive field scales. The expressions for the expert's ECA convolutional kernel size and dilation rate are:

[0054] In the formula, For the first The size of the one-dimensional convolution kernel used in the ECA channel attention module among the experts is used to control the modeling range of channel local correlations. For the first The dilation rate of the temporally dilated convolution in each expert is used to control the effective receptive field in the sequence dimension. All the above parameters are preset structural hyperparameters, their values ​​given by human design, used to construct a multi-scale expert library with different scale inductive biases. In the subsequent enhancement process, different experts extract enhancement features of different scales from the input components according to their respective ECA convolution kernel size and dilation rate, and then the router adaptively selects and weights them. These correspond to complementary enhancement paths from fine-grained local details to coarser-grained global structures, thus providing a clear scale division of labor for routing selection. The configuration and functional explanation of each expert are shown in Table 1. Table 1. Multiscale expert configuration and functional interpretation

[0055] For each low-frequency and high-frequency enhancement feature component, a lightweight router is introduced to generate low-frequency expert routes and high-frequency expert routes. The low-frequency expert routes and high-frequency expert routes are then subjected to temperature scaling and softmax operations to obtain the distribution of low-frequency route weights and high-frequency route weights, respectively. ; then execute Sparse activation, in this embodiment This allows a single sample to simultaneously activate two enhanced paths and perform a weighted combination, thereby reducing the variance caused by a single routing decision under conditions of strong distortion and low signal-to-noise ratio.

[0056] The residual update of the low-frequency structural component is calculated based on the low-frequency route weight and the low-frequency expert database, and the residual update of the high-frequency detail component is calculated based on the high-frequency route weight and the high-frequency expert database. The low-frequency and high-frequency route weights are normalized on the candidate experts to obtain low-frequency sparse weights and high-frequency sparse weights. The low-frequency enhancement feature and high-frequency enhancement feature components are calculated based on the low-frequency and high-frequency sparse weights, and their calculation formulas are based on a single path component and its... Write out the following samples:

[0057] In the formula, For the first Augmentation components of each sample The current component is the first in the batch Feature representation of a sample For expert index collection, For the first An expert updates the residuals generated by the component inputs. For the first Sparse weights of individual experts.

[0058] Specifically, low-frequency enhancement features are calculated by inputting low-frequency structural components into a low-frequency router and the corresponding expert database, using the formula described above. High-frequency enhancement features are calculated by inputting high-frequency detail components into a high-frequency router and the corresponding expert database, using the same formula. The calculation framework is consistent, but the input components, routing weights, and the set of activated experts differ. The router assigns weights to multiple enhancement experts at different scales based on the feature state of the current input sample and selects... Multiple experts participate in the augmentation calculation of the current sample to achieve adaptive multi-scale augmentation of the sample.

[0059] The fusion gating module obtains the fused features based on the main intermediate features, low-frequency enhanced features, and high-frequency enhanced features. Specifically, in this scheme, FusionGate (fusion gating module) is used to adaptively fuse the three information sources of the main intermediate features, low-frequency enhanced features, and high-frequency enhanced features, rather than a fixed linear superposition. FusionGate generates fusion weights using the three features as input, and obtains the fused features after softmax normalization.

[0060]

[0061] In the formula, As a feature of fusion, The weights are fused from the intermediate features of the backbone. The main trunk features For low-frequency enhanced feature fusion weights, It is a low-frequency enhancement feature. For high-frequency enhanced feature fusion weights, It is a high-frequency enhancement feature. For normalization function, This is a fusion weight generation function used to generate corresponding unnormalized weight scores based on the main intermediate features, low-frequency enhancement features, and high-frequency enhancement features.

[0062] To improve stability in the early stages of training, FusionGate employs an initialization strategy that favors preserving intermediate features of the backbone. This causes the model to prioritize backbone representations in the early optimization phases and gradually learn to incorporate low-frequency and high-frequency augmentation information as training progresses. This avoids unstable convergence caused by augmentation branches dominating the representation in the early stages. Specifically, it achieves conditional adaptive adjustment of the contribution of low-frequency and high-frequency information through sparse expert augmentation and three-way adaptive fusion. While preserving the stable representation of the backbone, it injects component-level discriminative gains, resulting in more consistent fused features for subsequent known class discrimination and open set rejection scoring.

[0063] The output module includes an embedding output branch and a classification output branch. The embedding output branch generates a discriminative embedding based on the fused features, while the classification output branch obtains and outputs the predicted scores for the known classes based on the fused features. Specifically, the embedding output branch maps the fused features to discriminative embeddings:

[0064] The classification output branch maps the fused features to predicted scores for the known classes:

[0065] in, ( ) represents an embedded output branch. ( ) represents the output branch for classification. For d-dimensional discriminative embedding, As a feature of fusion, for A known class prediction score vector. The number of known classes.

[0066] In training the open-set communication radiation source individual identification network, only the supervision signal from the known class training set is used. Given an input sample and its known labels, the network outputs a discriminant embedding and a known class prediction score. The known class training set refers to a dataset formed by selecting samples belonging to a predefined set of known transmitters from all transmitter samples; each sample in the set carries a corresponding known class label. The training set is obtained by partitioning the original dataset according to transmitter categories; no unknown transmitter samples are introduced during the training phase.

[0067] Discriminant embeddings are continuous vectors representing the location of sample features, primarily used for class center construction, distance calculation, and open set unknown modeling. Known class prediction scores are the output scores of the classification output branch on each known class, mainly used for cross-entropy supervised training and class prediction of known class samples. The overall training objective consists of three parts: classification loss, angular margin metric constraint, and routing-aided regularization term.

[0068] In the formula, Indicates the overall training objective. For classifying losses, Weights are used to measure angular intervals. For angular interval measurement constraints, For routing auxiliary item weights, This is a routing auxiliary regular expression.

[0069] The classification loss uses cross-entropy supervised recognition of known classes, and its expression is:

[0070] In the formula, Represents the cross-entropy loss function. For the first The known class prediction score vector corresponding to each sample For the first The true known class labels of each sample This is the sample index.

[0071] To improve the intra-class compactness and inter-class separability of the embeddings, an angular margin constraint (ArcFace) is applied to the normalized features and normalized classification weights, with the following expression:

[0072]

[0073]

[0074] In the formula, Let i be the classification score vector of the i-th sample after angular interval adjustment. As a scale factor, It is a cosine function. To perform discriminative embedding The eigenvectors obtained after normalization For the first Class classification weight The weight vector obtained after normalization It is an additive angular interval.

[0075] To suppress route collapse and promote expert load balancing, a route helper term is introduced during training. This term constrains route allocation to tend towards equilibrium on mini-batch samples and regularizes route uncertainty. The expression is as follows:

[0076] In the formula, The average soft route probability constraint for experts is used to constrain the average soft route probability of each expert to tend to equilibrium on a small batch of samples. This is a discrete activation frequency constraint for experts, used to ensure that the discrete activation frequencies of each expert tend to be balanced on a small batch of samples. This is a constraint on the uncertainty of route distribution, used to regularize the uncertainty of route distribution; The average soft route probability constraint weighting coefficient. These are the weighting coefficients for the discrete activation frequency constraint. The weighting coefficients represent the uncertainty of the routing distribution.

[0077] The Adam optimizer is used when training the open set communication radiation source individual identification network, and the initial learning rate is set to... The learning rate is adaptively adjusted during training using a learning rate scheduling strategy. To stabilize route training, a small learning rate is used for router parameters; the weights of route auxiliary terms undergo linear warm-up.

[0078] In the formula, For the first The routing auxiliary item weights corresponding to each training round The maximum weight of the routing auxiliary item. This is an operation to extract the smaller value from the items within the parentheses. For the number of preheating cycles, This is the current training round.

[0079] S4. Calculate the unknown score based on the discriminant embedding, and combine the unknown score with the known class prediction score to identify the category of the wireless signal transmission source. Specifically, when calculating the unknown score based on the discriminant embedding, known class prototypes are constructed based on known samples in the training set, and the distance from any discriminant embedding to each prototype is calculated:

[0080] In the formula, To distinguish between embedding and the first The distance between known class prototypes To determine embedding, Given a known class prototype, for Norm, The number of known classes.

[0081] The unknown score is calculated based on the distance between the discriminant embedding and the known class prototype:

[0082]

[0083] In the formula, For unknowns, Given the number of classes, For the first The softmin weights corresponding to the known class prototypes. It is an exponential function. To determine embedding With the Known class prototypes The distance between them This refers to the temperature coefficient in the unknowns calculation; the larger the unknowns score, the more the sample deviates from the known class prototype distribution and the more likely it belongs to an unknown emission source. As the temperature coefficient approaches 0, the unknowns score gradually approaches zero. This definition makes the unknown continuously differentiable, which is beneficial for subsequent threshold determination and open set decision-making.

[0084] When identifying wireless signal sources based on unknown scores and known class prediction scores, if the unknown score is greater than a preset unknown score threshold, the wireless signal source is considered to be an unknown and unidentifiable class. When the unknown score is not greater than the preset unknown score threshold, the wireless signal transmitter is considered to be of the known class, and the known class label of the wireless signal transmitter is output based on the predicted score of the known class. The expression is as follows:

[0085] In the formula, The result of identifying the wireless signal transmission source. For unknown categories, For the threshold, To ensure that the source of the received wireless signal is of a known class, The goal is to select the label corresponding to the category with the highest predicted score from the known categories numbered 1, 2, ..., K. The known category label is the category identifier in a preset set of known wireless signal transmitters; in specific implementations, the category identifier can be provided by the known transmitter annotations in the dataset and can be remapped to integer labels as needed for training and inference.

[0086] The decision decouples known class classification and unknown rejection classification. The known class prediction score is only used for known class discrimination, while rejection is determined entirely by the unknown score of the prototype distance.

[0087] This scheme only uses the known validation set to scan candidate thresholds during the validation phase, and uses the highest accuracy rate for known samples as the threshold selection criterion: for each candidate threshold, a decision is first made based on the unknown score and the threshold value; for accepted samples, further decisions are made... Given a known class prediction, the threshold is calculated using the following formula:

[0088] In the formula, Let Ω be the threshold value, and Ω be the set of candidate threshold values. Let be the decision accuracy rate on the known validation set, where known samples that are rejected are counted as errors. That is, the decision accuracy rate of the known samples is the optimal criterion for selecting the threshold.

[0089] This scheme uses a publicly available ADS-B dataset (Automatic Dependent Surveillance Dataset for Broadcast) for evaluation. This dataset consists of ADS-B radio signals collected by a real airborne surveillance system. Each sample consists of an in-phase component I and a quadrature component Q, and is represented in I / Q sequence form.

[0090] The evaluation follows the K-way Known / R-way Rogue open set protocol: nine emitters are fixed as known classes (training uses only known class samples, and labels are remapped to 0-8), and the remaining emitters are considered anomalous emission candidates. Different openness settings for (number of known emission sources: number of abnormal emission sources) In each In the partitioning, the test set consists of a mixture of known class test sets and abnormal class test sets. The abnormal class test set is obtained by sampling 100 samples without replacement from each abnormal emission source class in the test partitioning. All thresholds, temperature parameters and model selections are determined only in the validation phase. The test set does not participate in any form of threshold search or hyperparameter selection. The test phase only uses fixed settings for a one-time evaluation.

[0091] Evaluation metrics include ROC-AUC (Receiver Operating Characteristic Curve - Area Under the Curve), EER (Equal Error Rate), and Overall Accuracy (AAC). AUC and EER are calculated from the ROC curve (FPR on the horizontal axis, TPR on the vertical axis). Overall Acc is calculated at a fixed threshold and measures the overall accuracy of the final deployment decision. EER is derived from the ROC curve obtained through threshold scanning: the threshold point that makes FPR and FNR closest is selected, and the error rate corresponding to that point is used as the EER. This metric reflects the overall error level when known false rejections and unknown false receptions are balanced.

[0092] The open space risk reflects the increasing pressure of misjudgments in the deployment environment when no emission source is observed during training. As the number of anomalous emission devices increases, known / unknown boundaries are more easily disturbed, making it more difficult to stabilize rejection. Considering that medium to high openness is closer to real-world scenarios, we evaluated our proposed identification method on a 9:R scale (R=4–8) and compared the results with those reported in the MeDAE paper. The comparison results are shown in Table 2. Table 2. MeDAE Comparison Results

[0093] As shown in Table 2, within the range of 9:4–9:8 between the number of known and anomalous emission sources, the AUC of this method is 0.8312–0.8787 (compared to 0.7054–0.7576 for the control method), and the Overall Acc is 0.7738–0.8173 (compared to 0.6494–0.6919 for the control method). Overall, even in settings with high openness and a denser concentration of anomalous emission devices, this method still maintains a relatively stable ability to distinguish between known and unknown sources and a high final decision accuracy.

[0094] ROC comparison with public baselines. To further verify the effectiveness of the rejection capability, a comparison was made with several representative open set SEI (OS-SEI) methods, including DMNet, SR2CNN, FRAME (i.e., the open set communication radiation source individual identification method in this scheme), TRiplet+KNN, and OpenMax.

[0095] All replications of the comparative methods employed the same data partitioning and open-set evaluation protocol as this scheme. Specifically, each method completed parameter learning on a training set containing only known classes; after training, class centers were constructed using training set embeddings, and the minimum distance to the nearest class center was used as the unknown score for ROC / AUC / EER evaluation. The evaluation results are attached. Figure 6-8 As shown, the ROC curves (horizontal axis is FPR, vertical axis is TPR) are presented under three openness settings (9:1, 9:4, 9:8). It can be observed that the proposed scheme exhibits a higher ROC envelope under different openness settings, and the advantage is more significant in the low FPR region, indicating that while strictly controlling known false rejection constraints, it can still achieve stronger unknown detection capabilities.

[0096] To further verify the effectiveness of this scheme, ablation experiments were conducted to analyze the proposed method for identifying individual communication radiation sources, which includes a Frequency Routing Adaptive Expert Enhancement (FRAE) module. To verify the effect of FRAE on open set rejection, the performance of three settings—Backbone only, MEE only, and FRAE—was compared while maintaining other training settings. Table 3 summarizes the AUC, Err, and OverallAcc at three representative openness levels (9:1, 9:4, and 9:8). Higher values ​​for AUC and OverallAcc are better, while lower values ​​for Err are better. Δ represents the difference calculated relative to the backbone-only baseline. Table 3: Comparison of Ablation Procedures for Three Representative Openness Levels

[0097] Table 3 shows that FRAE provides a stable improvement in rejection performance at different openness levels. Compared to the baseline without a frequency module (w / Backbone), FRAE improves AUC by +8.42pp, +9.53pp, and +6.82pp at openness settings of 9:1, 9:4, and 9:8, respectively, indicating that frequency enhancement helps improve the embedding space's ability to distinguish between known and unknown samples. Simultaneously, EER decreases by -12.06pp, -7.58pp, and -4.82pp, respectively, indicating that near a more balanced false acceptance / false rejection operating point, the decision error between known and unknown samples is significantly reduced. Regarding the overall accuracy related to final deployment, FRAE improves the baseline by +1.06pp, +2.00pp, and +2.61pp at openness settings of 9:1, 9:4, and 9:8, respectively, showing that it improves the overall accuracy of the final joint decision at different openness levels.

[0098] Further decomposition of the total gain reveals that MEE-only improves the Acc of 9:1 / 9:4 / 9:8 by +1.37 / +0.82 / +0.57pp respectively, compared to Backbone-only, indicating that gated expert enhancement has a direct gain on the discrimination of known representations and closed sets. Furthermore, the introduction of frequency decomposition further significantly improves the rejection-related indicators: FRAE further improves the AUC of MEE-only at 9:1, 9:4, and 9:8 by +3.90 / +8.09 / +4.25pp, while EER further decreases to -7.06 / -4.47 / -2.23pp respectively, reflecting the contribution of the decomposition-enhancement-fusion structure to open set boundary modeling. It is worth noting that at a low openness ratio of 9:1, while FD provides a stronger rejection gain above MEE, Acc exhibits slight fluctuations (90.63%→90.32%, -0.31pp), reflecting that when the proportion of unknowns is low, the overall accuracy is more easily affected by the combined impact of threshold trade-offs and known classification errors. However, at medium to high openness ratios (9:4, 9:8), the Acc increment of FRAE above MEE only is more stable (+1.18pp, +2.04pp respectively), indicating that as the open space risk increases, FRAE can more consistently improve robust rejection and final decision performance. In summary, FRAE brings more consistent performance improvements at medium to high openness ratios, demonstrating its ability to effectively enhance robust rejection capability when the open space risk increases.

[0099] To further verify that the gating expert enhancement mechanism did not degenerate into a single path, the expert occupancy ratio (routing ratio) and mean probabilities of soft routing were statistically analyzed for the optimal checkpoint at both gating layers during the testing phase. The statistical results show that neither gating layer exhibited significant collapse on the test set: taking a known class test sample set as an example, the expert occupancy ratio of Layer 0 was [0.265, 0.235, 0.232, 0.268], and that of Layer 1 was [0.233, 0.243, 0.303, 0.221], with corresponding effective expert numbers of 3.99 and 3.97 respectively, indicating that expert utilization was close to equilibrium. Meanwhile, the routing decisions have high confidence and almost no tie: the maximum probability values ​​(maxP) of the known class test sample set in Layer 0 and Layer 1 (Layer 0 / 1) are 0.969 and 0.940, respectively, and the maximum probability values ​​(maxP) of the anomalous emission source test set in Layer 0 and Layer 1 (Layer 0 / 1) are 0.975 and 0.929, respectively. The tie is 0%, indicating that the routing selection is stable rather than random perturbation.

[0100] There are differences in expert preferences between known and unknown samples. Figure 9 As shown, taking the average value of the routing weights at this layer as an example, the distribution of the known source test sample set is [0.224, 0.247, 0.313, 0.216], while the distribution of the abnormal source test set is [0.122, 0.188, 0.213, 0.477]. The unknown samples are significantly more assigned to expert 3 (E3) (0.477 vs 0.216), while the known samples are more inclined towards expert 2 (E2). This phenomenon indicates that different experts learn complementary augmentation responses corresponding to different augmentation scales and inductive biases. The gating network can adaptively select a more suitable augmentation path based on the input features, thereby forming a sample-level augmentation strategy for open set scenarios.

[0101] This invention aims to provide an open set communication radiation source individual identification system for implementing the aforementioned open set communication radiation source individual identification method, comprising: A wireless signal acquisition unit is used to acquire the discrete baseband complex signal of the wireless signal under test. The backbone intermediate feature extraction unit is used to extract backbone intermediate features based on discrete baseband complex signals; The discriminant prediction unit is used to generate discriminant embeddings and known class prediction scores based on the intermediate features of the backbone. A wireless signal source identification unit is used to identify wireless signal sources based on discriminant embedding and known class prediction scores.

[0102] The present invention also aims to provide an open set communication radiation source individual identification device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the above-described open set communication radiation source individual identification steps.

[0103] The present invention also aims to provide a computer-readable storage medium containing a computer program, wherein the computer program is stored thereon, and when the computer program is executed by one or more processors, implements the above-described steps for identifying individuals of open set communication radiation sources.

[0104] The present invention also aims to provide a computer program product, including a computer program that, when executed by a processor, implements the above-described steps for identifying individuals of open set communication radiation sources.

[0105] The present invention also aims to provide an electronic device, including at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the above-described steps of identifying individuals of open set communication radiation sources.

[0106] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the scope of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying individual radiation sources in open set communication, characterized in that, Includes the following steps: Acquire the discrete baseband complex signal transmitted by the wireless signal transmitter, and obtain the backbone intermediate features based on the discrete baseband complex signal; Input the backbone intermediate features into the Open Set Communication Radiation Source Individual Identification Network. The Open Set Communication Radiation Source Individual Identification Network performs the following operations: obtain low-frequency structural components and high-frequency detail components based on the backbone intermediate features; obtain low-frequency enhancement features based on the low-frequency structural components; obtain high-frequency enhancement features based on the high-frequency detail components; obtain fusion features based on the backbone intermediate features, low-frequency enhancement features, and high-frequency enhancement features; and generate and output discriminative embeddings and known class prediction scores based on the fusion features. The unknown score is calculated based on the discriminant embedding, and the category of the wireless signal source is identified by combining the unknown score with the known class prediction score.

2. The method for individual identification of communication radiation sources according to claim 1, characterized in that, The steps for obtaining low-frequency structural components and high-frequency detail components based on the intermediate features of the main trunk include: The bandwidth coefficient is obtained based on the intermediate characteristics of the backbone, and the effective half-bandwidth is calculated based on the bandwidth coefficient. Construct differentiable center-limited masks and complementary masks based on the effective half-bandwidth; The frequency domain representation is obtained based on the intermediate features of the backbone, and the frequency domain representation is divided into low-frequency and high-frequency components based on the differentiable center band-limited mask and complementary mask. The low-frequency frequency domain components are transformed into low-frequency structural components, and the high-frequency frequency domain components are transformed into high-frequency detail components.

3. The method for individual identification of communication radiation sources according to claim 1, characterized in that, The steps of obtaining low-frequency enhancement features based on low-frequency structural components and high-frequency enhancement features based on high-frequency detail components include: Build a low-frequency expert database and a high-frequency expert database, and generate low-frequency expert routes and high-frequency expert routes; Calculate the weight of low-frequency routes based on low-frequency expert routes, and calculate the weight of high-frequency routes based on high-frequency expert routes. Residual updates for low-frequency structural components are calculated based on low-frequency routing weights and a low-frequency expert database, while residual updates for high-frequency detail components are calculated based on high-frequency routing weights and a high-frequency expert database. Calculate low-frequency sparse weights based on low-frequency routing weights, and calculate high-frequency sparse weights based on high-frequency routing weights. Low-frequency enhancement features are obtained by weighted calculation based on the residual update of low-frequency sparse weights and low-frequency structural components, and high-frequency enhancement features are obtained by weighted calculation based on the residual update of high-frequency sparse weights and high-frequency detail components.

4. The method for individual identification of communication radiation sources according to claim 1, characterized in that, The step of obtaining the fused features based on the backbone intermediate features, low-frequency enhancement features, and high-frequency enhancement features includes: Fusion weights are generated based on the main intermediate features, low-frequency enhancement features, and high-frequency enhancement features. The fused features are obtained by weighting and fusing the backbone intermediate features, low-frequency enhanced features, and high-frequency enhanced features based on the fusion weights.

5. The method for individual identification of communication radiation sources according to claim 1, characterized in that, The training steps of the open set communication radiation source individual identification network include: Obtain discrete baseband complex signal sample data and construct a training set of known classes; Discriminant embeddings and predicted scores for known classes are output from the training set of known classes; Based on discriminative embedding and known class prediction scores, a classification loss, angular margin metric constraint, and routing auxiliary regularization term are constructed, and an overall training objective is constructed based on the classification loss, angular margin metric constraint, and routing auxiliary regularization term. Optimize the parameters of the open set communication radiation source individual identification network according to the overall training objectives.

6. The method for individual identification of communication radiation sources according to claim 1, characterized in that, The step of calculating the unknown score based on the discriminant embedding includes: Construct a known class prototype and calculate the distance to the embedding in the known class prototype; The unknown score is calculated based on the distance between the discriminant embedding and the known class prototype.

7. The method for individual identification of communication radiation sources according to claim 1, characterized in that, The method of identifying the category of a wireless signal emission source by combining the unknown score and the known class prediction score includes: When the unknown score is greater than the preset unknown score threshold, the wireless signal source is considered to be an unknown and unidentifiable category. When the unknown score is not greater than the preset unknown score threshold, the wireless signal transmitter is considered to be of the known class, and the known class label of the wireless signal transmitter is output based on the predicted score of the known class.

8. An open-set communication radiation source individual identification system, used to implement the open-set communication radiation source individual identification method according to any one of claims 1-7, characterized in that, include: A wireless signal acquisition unit is used to acquire the discrete baseband complex signal of the wireless signal under test. The backbone intermediate feature extraction unit is used to extract backbone intermediate features based on discrete baseband complex signals; The discriminant prediction unit is used to generate discriminant embeddings and known class prediction scores based on the intermediate features of the backbone. A wireless signal source identification unit is used to identify wireless signal sources based on discriminant embedding and known class prediction scores.

9. An open-set communication radiation source individual identification device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the open set communication radiation source individual identification method according to any one of claims 1-7.

10. A computer-readable storage medium containing a computer program, wherein the computer program is stored thereon, characterized in that, When the computer program is executed by one or more processors, it implements the steps of the open set communication radiation source individual identification method according to any one of claims 1-7.

Images