A radio frequency fingerprinting method and system based on adaptive soft subspace evolution

Abstract

The application relates to the technical field of radio frequency fingerprint identification, in particular to a radio frequency fingerprint identification method based on adaptive soft subspace evolution, which comprises the following steps: inputting a preprocessed radio frequency signal into a feature extraction network to perform high-dimensional feature extraction and obtain high-dimensional features; based on a physical classification weight sharing mechanism, combining a soft mask generated by using an improved soft mask generation algorithm with temperature annealing, constructing logical classification centers corresponding to different categories; continuously monitoring the in-class distance variance of the features of each category to the logical classification center thereof; when the in-class distance variance exceeds a first preset threshold, triggering a mask splitting mechanism to generate a new soft mask; performing in-group style mixing on the high-dimensional features by using the soft mask to obtain enhanced features guided by the soft mask; based on the enhanced features guided by the soft mask, combining an extreme value routing strategy to perform category prediction and obtain a radio frequency fingerprint identification result; and performing ternary joint optimization on the radio frequency fingerprint identification result in combination with a classification loss, an orthogonal constraint and a sparse constraint.

Description

Technical Field

[0001] The embodiments of this application relate to the field of radio frequency fingerprint recognition technology, and in particular to a radio frequency fingerprint recognition method and system based on adaptive soft subspace evolution. Background Technology

[0002] Radio Frequency Fingerprint Identification (RFFI) and Specific Emitter Identification (SEI) are key technologies in physical layer security and reconnaissance. Their core lies in using the minute tolerances (such as frequency offset, I / Q imbalance, phase noise, etc.) that are inevitably introduced during the manufacturing process of transmitter hardware to extract a unique "hardware fingerprint" from the intercepted signal, thereby realizing the identification of the device's operating mode or the accurate authentication of the individual's identity.

[0003] With the widespread application of modern wireless communication and radar technologies such as 5G and LoRa (a long-range low-power radio technology standard), the electromagnetic environment is becoming increasingly complex, and signals exhibit diverse modulation patterns and rapid parameter changes. In practical scenarios such as non-cooperative reconnaissance or large-scale IoT deployments, systems often face severe cross-domain generalization challenges. On the one hand, factors such as time, location, line-of-sight / non-line-of-sight (LOS / NLOS) conditions, and dynamically changing signal-to-noise ratio (SNR) introduce significant environmental noise (i.e., domain shift), which can easily mask the already weak hardware fingerprint characteristics. On the other hand, when facing new high-threat radiation sources or dynamically accessed edge devices, there are often only a very small number of samples available for analysis. Traditional deep learning methods rely on a large amount of labeled data in a single scenario, which is prone to overfitting under small sample conditions and exhibits a sharp decline in performance under complex and variable channel environments.

[0004] Therefore, how to construct an RF fingerprint recognition system with both high accuracy and strong generalization ability under conditions of scarce samples, dynamic environment, and uncertain channel has become a key technical bottleneck that urgently needs to be overcome.

[0005] Chinese invention patent application number 202210652213.7 discloses a radar radiation source identification method based on relation awareness and prototype optimization, aiming to solve the problems of large intra-class differences and weak inter-class separability under small sample conditions. This method introduces a graph convolutional network to construct a class prototype relationship graph, explicitly modeling the topological relationships between different signal classes, and enhances inter-class separability through prototype-level contrastive loss. Simultaneously, channel gating and sample soft attention mechanisms are integrated in the prototype aggregation stage to suppress noise interference.

[0006] Chinese invention patent application number 202011258729.0 discloses a method for individual identification of open-set radar radiation sources based on transfer learning. This method trains the radar by mixing intermediate frequency signal data from different radars under various operating modes, and utilizes transfer learning strategies to reduce the interference of operating modes on individual identification. This method emphasizes improving the mode invariance of features through data obfuscation to address identification fluctuations caused by the diversity of operating modes.

[0007] To address the channel domain offset problem, traditional domain generalization and multi-branch networks often employ brute-force data augmentation using global style mixing (such as global MixStyle) or hybrid expert models (MoE) that physically superimpose multiple network branches to process data under different channel states.

[0008] Through in-depth analysis of the above technical solutions, the inventors of this application have discovered that the above technical solutions have the following fatal drawbacks when dealing with the extraction of radio frequency micro-features in complex electromagnetic environments.

[0009] First, the indiscriminate data augmentation and feature mixing severely damage the underlying micro-fingerprint. Existing technologies typically employ global data obfuscation or style mean / variance exchange when dealing with cross-modal / cross-domain problems. However, radio frequency fingerprints are extremely vulnerable to physical layer damage features. If, during the feature extraction stage, features from "high SNR clean channels" are indiscriminately mixed globally with those from "severely multipath-distorted channels," a "distorted" signal distribution that is fundamentally nonexistent in a physical and electromagnetic sense will be generated. This brute-force mixing not only fails to learn domain-invariant features but also completely obliterates the device's original micro-fingerprint, leading to "negative transfer" in the model and difficulty in convergence.

[0010] Second, the rigid classification prototype structure cannot cope with multimodal distributions in cross-domain environments. Although existing technologies optimize inter-class relationships, they still essentially assume that each device has a single central prototype in the feature space. In cross-domain RFFI scenarios, the feature distribution of the same device under "clear line-of-sight channels" and "complex multipath channels" often exhibits severe "bimodal" or "multimodal" characteristics. In order to accommodate all channel states, the single rigid classification head is forced to expand the decision boundary, becoming a fuzzy "greatest common divisor," resulting in the loss of fine-grained fingerprint features, and radiation sources of highly similar categories are easily confused.

[0011] Third, the network has poor adaptive evolution capabilities and extremely high parameter redundancy. When faced with complex distributions, existing methods that attempt to introduce multi-branch networks (expert models) for divide-and-conquer often exponentially increase the number of physical parameters in the model, which is unacceptable for edge IoT devices or spaceborne / airborne reconnaissance nodes with extremely limited storage and computing power. In addition, rigid network gating or physical splitting can cause severe gradient oscillations during training, resulting in an unsmooth model evolution process that is prone to getting trapped in local optima. Summary of the Invention

[0012] In view of this, embodiments of this application propose a radio frequency fingerprinting method and system based on adaptive soft subspace evolution, which can dynamically construct multiple logical subspaces through a learnable soft masking mechanism without increasing the number of model parameters, thereby achieving high-precision and robust identification of multimodal radiation source signals in complex cross-domain scenarios.

[0013] To achieve the above objectives, embodiments of this application propose a radio frequency fingerprinting method based on adaptive soft subspace evolution, comprising: preprocessing the radio frequency signals used as training samples, inputting the preprocessed radio frequency signals into a feature extraction network for high-dimensional feature extraction to obtain high-dimensional features; constructing logical classification centers corresponding to different categories based on a physical classification weight sharing mechanism and a soft mask generated by an improved soft mask generation algorithm with temperature annealing; continuously monitoring the intra-class distance variance from the features of each category to their logical classification centers during training, and triggering a mask splitting mechanism to generate a new soft mask when the intra-class distance variance exceeds a first preset threshold; performing intra-group style mixing on the high-dimensional features using the soft mask to obtain enhanced features guided by the soft mask; calculating the cosine similarity between the enhanced features guided by the soft mask and all logical classification centers under each category, performing category prediction based on an extreme value routing strategy to obtain the radio frequency fingerprinting result, and performing ternary joint optimization by combining classification loss, orthogonality constraint, and sparsity constraint.

[0014] To achieve the above objectives, embodiments of this application also propose a radio frequency fingerprinting system based on adaptive soft subspace evolution, comprising: a preprocessing module for preprocessing radio frequency signals used as training samples; a soft subspace construction module for constructing logical classification centers corresponding to different categories based on a physical classification weight sharing mechanism and soft masks generated using an improved soft mask generation algorithm with temperature annealing; and an adaptive evolution monitoring module for inputting the preprocessed radio frequency signals into a feature extraction network for high-dimensional feature extraction to obtain high-dimensional features, and continuously monitoring during training. The intra-class distance variance from each category's features to its logical classification center is calculated. When the intra-class distance variance exceeds a first preset threshold, a mask splitting mechanism is triggered to generate a new soft mask. The mask-guided data augmentation module is used to perform intra-group style mixing on high-dimensional features using the soft mask to obtain enhanced features guided by the soft mask. The joint optimization module is used to calculate the cosine similarity between the enhanced features guided by the soft mask and all logical classification centers under its category, perform category prediction based on the extreme value routing strategy, obtain the RF fingerprint recognition result, and perform ternary joint optimization by combining classification loss, orthogonality constraint, and sparsity constraint.

[0015] To achieve the above objectives, embodiments of this application also propose an electronic device, the electronic device comprising: a processor and a memory, the memory storing instructions executable by the processor, the processor being configured to execute the instructions such that the electronic device can implement the radio frequency fingerprinting method based on adaptive soft subspace evolution as described above.

[0016] To achieve the above objectives, embodiments of this application also propose a computer-readable storage medium storing a computer program that, when executed by a processor, enables a radio frequency fingerprinting method based on adaptive soft subspace evolution as described above.

[0017] Optionally, the high-dimensional features include the device's inherent micro-hardware fingerprint and the domain offset noise introduced by the current channel, and the high-dimensional features are denoted as... , , This represents the preprocessed radio frequency signal, i.e., the I / Q sequence. This represents a feature extraction network, which can be a one-dimensional ResNet or CNN. For dimension The complex field of .

[0018] Optionally, based on the physical classification weight sharing mechanism, and combined with the soft mask generated by the improved soft mask generation algorithm with temperature annealing, logical classification centers corresponding to different categories are constructed, including: Maintain a globally shared physical classification weight matrix , , For dimension The complex field of , the first The weight vector of a class is denoted as , Shared with all experts to ensure consistency of the underlying physical semantics. ; For the first The class of An expert initializes a learnable parameter vector. , ,based on Through temperature coefficient The sigmoid function generates a series of soft masks, the th... The class of A soft mask for experts Represented as: ; in, For the Sigmoid function, The value of is gradually annealed to 0 as training progresses; By combining physical weights with soft masks through Hadamard product, logical classification centers oriented towards a specific domain are generated. The class of Logical classification center Represented as: ; in, This represents the Hadamard product operation.

[0019] Optionally, a mask splitting mechanism is triggered to generate a new soft mask, including: Copy the current soft mask parameters Apply a small Gaussian perturbation to it Generate two new soft masks and , , .

[0020] Optionally, within-group style blending of high-dimensional features is performed using soft masking to obtain soft-mask-guided enhanced features, including: Based on the distance between each high-dimensional feature and each logical classification center, the high-dimensional features are grouped; each group corresponds to a category. Iterate through each group, and within the current group, exchange the mean and standard deviation of each high-dimensional feature to generate the preliminary mixed features of the current group; where, the th The initial mixed characteristics of the group are denoted as ; Style fusion is performed on the high-dimensional features and preliminary mixed features of the current group using the soft mask corresponding to the current group, resulting in soft-mask-guided enhanced features; where, the first Group 1 Enhanced features guided by a soft mask The calculation process is expressed as follows: , Indicates the first The soft mask corresponding to the group, For the first Group 1 A high-dimensional feature.

[0021] Optionally, during inference, the cosine similarity between the enhanced features guided by the soft mask and all logical classification centers under each category is calculated. Category prediction is then performed based on an extreme value routing strategy to obtain the RF fingerprint recognition result, including: Iterate through each category, calculate the cosine similarity between the enhanced features guided by the soft mask and all logical classification centers under the current category, and take the maximum value as the prediction score of the current category; After traversing all categories, the category with the highest predicted score is selected as the RFID fingerprint recognition result. No. The calculation process for the predicted score of a class is expressed as follows: ; in, Enhanced features guided by soft masks, For the first The predicted score of the class. This indicates the calculation of the L2 norm.

[0022] Optionally, a ternary joint optimization can be performed by combining classification loss, orthogonality constraint, and sparsity constraint, as shown in the following formula: ; ; ; in, The total loss for ternary joint optimization, For classifying losses, For orthogonal constraints, For sparsity constraints, This indicates L1 regularization.

[0023] The radio frequency fingerprinting method based on adaptive soft subspace evolution proposed in this application brings the following advantages compared with traditional radio frequency fingerprinting methods.

[0024] First, it achieves a multi-expert architecture with "zero parameter proliferation," effectively solving the problems of limited computing power and storage capacity of edge devices. Current mainstream hybrid expert models or multi-branch networks typically address complex cross-domain data distributions by physically stacking multiple sub-networks, leading to an exponential increase in model parameters and computational overhead, making them difficult to deploy on resource-constrained IoT edge nodes. This application innovatively proposes a logical subspace construction mechanism based on soft masks and shared physical weights. By introducing only a minimal number of one-dimensional learnable mask parameters, multiple logical classification sub-centers can be dynamically combined based on shared weights. Without increasing the physical parameters of the feature extractor and classification head, this application achieves multimodal representation capabilities comparable to or even superior to massive physical multi-branch networks, significantly reducing model deployment costs and inference latency, and providing a feasible path for edge intelligence.

[0025] Second, it achieves precise data augmentation while effectively avoiding the damage to the underlying physical fingerprint caused by global style mixing. Traditional domain generalization methods employ a globally indiscriminate style mixing strategy during data augmentation, crudely fusing features from samples with vastly different channel conditions. This easily damages the extremely sensitive and fragile device-level micro-physical fingerprint in RF signals, leading to a "negative migration" phenomenon, where augmentation actually degrades performance. This application proposes a mask-guided style mixing mechanism, utilizing a soft mask as a natural feature selector. Style exchange between samples within the same domain is performed only within the currently activated feature dimension (i.e., the high-response region), while strictly preserving the original fingerprint information in the mask-suppressed region. This mechanism improves the model's robustness to environmental disturbances while ensuring the integrity of the device fingerprint at the electromagnetic-physical level, significantly improving the accuracy of RF fingerprint recognition in cross-domain scenarios.

[0026] Third, this application achieves a high degree of smoothness in the network evolution process, completely avoiding the training oscillation problem caused by traditional hard splitting. Traditional dynamic network expansion or routing switching mechanisms often rely on discrete "hard decisions," which can instantly disrupt the converged feature manifold, causing severe oscillations in the loss function and easily leading to training instability or even collapse. This application designs a "mitotic" evolution mechanism based on mask replication and perturbation, combined with a temperature annealing strategy. When the complexity of a certain category's feature distribution exceeds a threshold, only its corresponding soft mask is replicated and a small Gaussian noise is injected to break the symmetry, while keeping the shared weights unchanged. Subsequently, during the annealing process, new and old experts gradually complete the functional differentiation. This application ensures seamless knowledge inheritance and continuous gradient backpropagation, enabling the model to maintain training stability during adaptive expansion, significantly shortening the convergence time and improving the final performance.

[0027] Fourth, it achieves explicit feature decoupling and adaptive noise reduction, effectively addressing the challenge of feature confusion under harsh channel conditions. Traditional methods model all feature dimensions uniformly. Once some frequency bands suffer from strong multipath fading or narrowband interference, the entire category prototype is contaminated, leading to a sharp decline in discrimination performance. This application introduces two key constraints into the optimization objective: orthogonality constraint and sparsity constraint. The former forces different logic experts under the same category to focus on mutually exclusive feature subspaces, thereby achieving functional complementarity. The latter, through L1 regularization, prompts the mask to automatically close channels that are severely contaminated by noise or have redundant information. The two work together to enable the model to have strong adaptive noise reduction capabilities and feature fault tolerance in real electromagnetic environments with extremely low signal-to-noise ratios or strong interference, significantly improving the robustness and reliability of the system. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies of this application will be briefly introduced below. The following drawings are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. The drawings described herein are only used to explain this application and are not intended to limit this application.

[0029] Figure 1 This is a flowchart of a radio frequency fingerprinting method based on adaptive soft subspace evolution provided in one embodiment of this application; Figure 2 This is a detailed schematic diagram of a radio frequency fingerprinting method based on adaptive soft subspace evolution provided in one embodiment of this application; Figure 3 This is a schematic diagram of the structure of a radio frequency fingerprint recognition system based on adaptive soft subspace evolution provided in another embodiment of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided in another embodiment of this application. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the various embodiments of this application will be described in detail below with reference to the accompanying drawings. Those skilled in the art will understand that many technical details have been presented in the embodiments of this application to facilitate better understanding. However, the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments. The division of the following embodiments is for ease of description and should not constitute any limitation on the specific implementation of this application. The following embodiments can be combined with and referenced by each other without contradiction.

[0031] This application primarily addresses the technical problem of poor generalization ability and susceptibility to crashing of RF fingerprint recognition models in complex and variable physical channel environments (such as multipath propagation and varying signal-to-noise ratios). Specifically, it includes: 1) Solve the "semantic collapse" problem caused by the accidental deletion of fingerprint features in domain generalization.

[0032] Traditional domain generalization methods typically force the alignment of feature distributions across different domains. This crude alignment mechanism cannot distinguish between the subtle "RF fingerprints" that characterize device identity and the "domain noise" that characterizes environmental changes. During the alignment process, key fingerprint features are easily treated as noise and erased, resulting in the model achieving domain invariance but losing its core ability to identify individual devices.

[0033] This application proposes a decoupled feature purification mechanism that explicitly separates fingerprint representation and noise representation in the feature space and applies domain alignment constraints only to the noise component, thereby significantly improving the domain adaptability of the model while preserving weak fingerprint features.

[0034] 2) It resolves the contradiction between limited resources for edge devices and model complexity.

[0035] While traditional hybrid expert models can improve their ability to fit complex distributions by adding branches, their parameter count and computational cost increase exponentially. This high resource consumption fundamentally conflicts with the low storage and low computing power characteristics of IoT edge computing nodes (such as sensors and embedded devices), making the models difficult to deploy in practice.

[0036] This application designs a dynamic sparse activation architecture, which introduces lightweight gating units to dynamically select active experts and combines parameter sharing mechanisms and gradient blocking strategies to strictly control the growth of parameters while maintaining the model's expressive power, so as to adapt it to the resource constraints of edge devices.

[0037] 3) It solves the problems of training instability and convergence difficulties caused by network hard splitting.

[0038] When using hard gating or physically adding classification heads for dynamic network evolution, discrete splitting actions can disrupt the existing feature manifold structure, causing the loss function to oscillate violently. This oscillation not only prolongs training time but also easily causes the model to get stuck in local optima or even diverge, seriously affecting the model's stability.

[0039] This application proposes a progressive network evolution strategy that uses a soft routing mechanism and a continuity regularization term to smoothly transition changes in network structure, eliminates gradient shocks caused by structural abrupt changes, and ensures that the loss function can converge smoothly during the dynamic adjustment process.

[0040] 4) It solves the problems of physical distortion and negative migration caused by global data augmentation.

[0041] Traditional style mixing methods blindly mix features with different signal-to-noise ratios in the global domain, ignoring the causal logic of the signal in the physical generation process. This unconstrained mixing is prone to generating distorted signal features that do not exist physically, misleading the model's learning and leading to a negative performance degradation (i.e., negative transfer).

[0042] This application introduces a physical consistency enhancement strategy, which constrains the spatial and intensity range of feature mixing based on prior channel knowledge, ensuring that the generated enhanced samples conform to the propagation law of the physical channel, thereby improving data diversity while avoiding the introduction of distortion interference.

[0043] In summary, this application proposes a decoupled, lightweight RF fingerprinting framework for complex channel environments. Through a collaborative mechanism of feature decoupling, dynamic sparse activation, progressive evolution, and physical consistency enhancement, it systematically solves the key bottlenecks of the RFFI model in terms of generalization, deployment, training stability, and data augmentation rationality, and significantly improves the robustness and practicality of the model in variable physical environments.

[0044] One embodiment of this application proposes a radio frequency fingerprinting method based on adaptive soft subspace evolution. The implementation details of the radio frequency fingerprinting method based on adaptive soft subspace evolution proposed in this embodiment are described in detail below. The following implementation details are provided for ease of understanding and are not necessary for implementing this solution.

[0045] The specific process of the radio frequency fingerprint recognition method based on adaptive soft subspace evolution proposed in this embodiment can be described as follows: Figure 1 As shown, its specific details are as follows: Figure 2 As shown, the method includes: S1, preprocess the radio frequency signals used as training samples, and input the preprocessed radio frequency signals into the feature extraction network for high-dimensional feature extraction to obtain high-dimensional features.

[0046] During training, the radio frequency (RF) signals used as training samples are first preprocessed. The preprocessed RF signals are then input into a shared feature extraction network for high-dimensional feature extraction, resulting in high-dimensional features. These high-dimensional features include the device's inherent microscopic hardware fingerprint and domain offset noise introduced by the current channel.

[0047] In one example, the high-dimensional feature is denoted as , , This represents the preprocessed radio frequency signal, i.e., the I / Q sequence. This represents a feature extraction network, which can be a one-dimensional ResNet or CNN. For dimension The complex field of .

[0048] S2, based on the physical classification weight sharing mechanism, combines the soft mask generated by the improved soft mask generation algorithm with temperature annealing to construct logical classification centers corresponding to different categories.

[0049] In the specific implementation, in addition to extracting high-dimensional features, it is necessary to construct logical classification centers corresponding to different categories based on the physical classification weight sharing mechanism and the soft mask generated by the improved soft mask generation algorithm with temperature annealing.

[0050] To avoid the parameter inflation and information fragmentation caused by hard partitioning in traditional multi-branch structures, this application designs a dynamic subspace mechanism with zero parameter redundancy, which includes three parts: physical classification weight sharing, soft mask generation and temperature annealing, and logical classification center construction.

[0051] In the physical classification weight sharing part, a globally shared physical classification weight matrix is ​​maintained. , , The total number of categories, For dimension The complex field of , the first The weight vector of a class is denoted as , Shared with all experts to ensure consistency of the underlying physical semantics. .

[0052] In the soft mask generation stage, for the first The class of An expert initializes a learnable parameter vector. , ,based on Through temperature coefficient The sigmoid function generates a series of soft masks, the th... The class of A soft mask for experts Represented as: ,in, For the Sigmoid function, The value of is set to a larger value in the early stage of training to make the mask smooth and promote feature sharing. In the later stage, it is gradually annealed to 0 to make the mask tend to be binary and achieve adaptive decoupling of feature dimensions.

[0053] Finally, logical classification centers are constructed by combining physical weights with soft masks through Hadamard product to generate logical classification centers oriented towards a specific domain. The class of Logical classification center Represented as: ,in, This represents the Hadamard product operation. This operation is equivalent to "focusing" on fingerprint dimensions that are stable in the current domain and "masking" dimensions that are severely affected by noise.

[0054] In one example, during the generation of the soft mask, in addition to using the sigmoid function with temperature annealing, Gumbel-Softmax can be used to achieve a differentiable hard mask approximation, or activation functions such as Sparsemax or ReLU with sparse regularization can be used to directly output a sparse mask. Furthermore, the mask parameters can be replaced by a static class-level vector with a dynamic instance-level mask, for example, by using a lightweight hypernetwork or MLP, taking channel state information (such as SNR, CSI, RSSI) as input to generate a sample-adaptive mask.

[0055] S3 continuously monitors the intra-class distance variance from the features of each category to their logical classification center during training. When the intra-class distance variance exceeds the first preset threshold, a mask splitting mechanism is triggered to generate a new soft mask.

[0056] During training, it is necessary to continuously monitor the intra-class distance variance from the features of each category to their logical classification center. When the intra-class distance variance exceeds the first preset threshold, it indicates that the current subspace cannot fully fit the multimodal distribution. At this time, the mask splitting mechanism is immediately triggered to generate a new soft mask.

[0057] In one example, the mask splitting mechanism is similar to mitosis. When a split is triggered, the physical weight matrix remains unchanged, only the mask parameter vector for that class is copied, and a small amount of Gaussian noise is applied to the newly copied parameters to break the symmetry, i.e., copying the current soft mask parameters. Apply a small Gaussian perturbation to it Generate two new soft masks and , , .

[0058] This process does not add new physical weights, but only expands the number of logical experts, ensuring knowledge inheritance and gradient smoothing, and effectively avoiding the violent oscillations in the loss function caused by traditional hard splitting operations.

[0059] In one example, in addition to using intra-class distance variance as a basis, validation loss stagnation detection, prediction probability entropy threshold, prediction cross-entropy threshold, feature clustering contour coefficient, and even the introduction of reinforcement learning agents can be used to dynamically decide whether to trigger mitosis of the mask.

[0060] S4. Soft masking is used to perform intra-group style blending on high-dimensional features to obtain soft mask-guided enhanced features.

[0061] In practical implementation, for high-dimensional features, in order to prevent global style mixing from destroying the fragile microscopic physical fingerprint in the radio frequency signal, the data augmentation process is directly controlled by soft masking. That is, the soft mask is used to perform intra-group style mixing on the high-dimensional features to obtain the enhanced features guided by the soft mask.

[0062] In the process of feature enhancement, the first step is to perform extreme feature grouping, which is to group each high-dimensional feature based on the distance between each high-dimensional feature and each logical classification center. Each group corresponds to a category, and the same group can be considered to have similar channel domain backgrounds.

[0063] Next, iterate through each group, swapping the mean and standard deviation of each high-dimensional feature within the current group to generate the preliminary mixed features for the current group. Among these, the first... The initial mixed characteristics of the group are denoted as .

[0064] Finally, the high-dimensional features and preliminary mixed features of the current group are style-fused using the soft mask corresponding to the current group, resulting in soft-mask-guided enhanced features. Among these, the... Group 1 Enhanced features guided by a soft mask The calculation process is expressed as follows: , Indicates the first The soft mask corresponding to the group, For the first Group 1 A high-dimensional feature.

[0065] This mechanism ensures that enhancements are applied only to dimensions of interest to experts, while the original fingerprint information is fully preserved in inactive areas, achieving precise enhancement.

[0066] S5 calculates the cosine similarity between the enhanced features guided by the soft mask and all logical classification centers under each category, performs category prediction based on the extreme value routing strategy, obtains the RF fingerprint recognition result, and performs ternary joint optimization by combining classification loss, orthogonality constraint and sparsity constraint.

[0067] In the specific implementation, during inference, it is necessary to calculate the cosine similarity between the enhanced features guided by the soft mask and all logical classification centers under each category, perform category prediction based on the extreme value routing strategy, and obtain the RF fingerprint recognition result. During training, it is also necessary to perform ternary joint optimization by combining classification loss, orthogonality constraint and sparsity constraint.

[0068] In one example, during reasoning, each category is traversed, the cosine similarity between the enhanced features guided by the soft mask and all logical classification centers under the current category is calculated, and the maximum value is taken as the predicted score of the current category. After traversing all categories, the category with the highest predicted score is selected as the RFID fingerprint recognition result.

[0069] In one example, the first The calculation process for the predicted score of a class is expressed as follows: ; in, Enhanced features guided by soft masks, For the first The predicted score of the class. This indicates the calculation of the L2 norm.

[0070] In one example, a ternary joint optimization is performed by combining classification loss, orthogonality constraint, and sparsity constraint, achieved through the following formula: ; ; ; in, The total loss is for ternary joint optimization. For classification loss, cross-entropy loss or Circle loss is applied to bring the samples closer to the optimal subcenter. To enforce orthogonality constraints, cosine similarity between different masks of the same type is penalized, forcing experts to focus on complementary feature dimensions. To constrain sparsity, L1 regularization is used to encourage mask sparsity, activating only the most discriminative dimension. This indicates L1 regularization.

[0071] In one example, orthogonal constraints can replace cosine similarity penalties by minimizing the absolute value of the inner product between masks, KL divergence, or JS divergence, while sparse constraints can be replaced by L0 norm approximation or structured Dropout mechanisms instead of L1 regularization.

[0072] Any technical approach that employs "dynamically constructing a logical feature subspace using learnable soft / hard masks under a fixed physical classification weight dimension" and combines it with "a mask-based smooth evolution expansion mechanism" or "using masks as filters to implement restricted data augmentation" to achieve highly robust and low-redundancy identification of cross-domain, multimodal signals, regardless of its specific implementation form, should be considered to fall within the protection scope of this embodiment.

[0073] The radio frequency fingerprinting method based on adaptive soft subspace evolution proposed in this embodiment brings the following beneficial effects compared with traditional radio frequency fingerprinting methods.

[0074] First, it achieves a "zero-parameter surge" multi-expert architecture, effectively solving the problems of limited computing power and storage capacity of edge devices. Current mainstream hybrid expert models or multi-branch networks typically address complex cross-domain data distributions by physically stacking multiple sub-networks, leading to an exponential increase in model parameters and computational overhead, making them difficult to deploy on resource-constrained IoT edge nodes. This embodiment innovatively proposes a logical subspace construction mechanism based on soft masks and shared physical weights. By introducing only a minimal number of one-dimensional learnable mask parameters, multiple logical classification sub-centers can be dynamically combined based on shared weights. Without increasing the physical parameters of the feature extractor and classification head, this embodiment achieves multimodal representation capabilities comparable to or even superior to massive physical multi-branch networks, significantly reducing model deployment costs and inference latency, and providing a feasible path for edge intelligence.

[0075] Second, it achieves precise data augmentation while effectively avoiding the damage to the underlying physical fingerprint caused by global style mixing. Traditional domain generalization methods employ a globally indiscriminate style mixing strategy during data augmentation, crudely fusing features from samples with vastly different channel conditions. This easily damages the extremely sensitive and fragile device-level micro-physical fingerprint in RF signals, leading to a "negative migration" phenomenon, where augmentation actually degrades performance. This embodiment proposes a mask-guided style mixing mechanism, utilizing a soft mask as a natural feature selector. Style exchange between samples within the same domain is performed only within the currently activated feature dimension (i.e., the high-response region), while strictly preserving the original fingerprint information in the mask-suppressed region. This mechanism improves the model's robustness to environmental disturbances while ensuring the integrity of the device fingerprint at the electromagnetic-physical level, significantly improving the accuracy of RF fingerprint recognition in cross-domain scenarios.

[0076] Third, it achieves a high degree of smoothness in the network evolution process, completely avoiding the training oscillation problem caused by traditional hard splitting. Traditional dynamic network expansion or routing switching mechanisms mostly rely on discrete "hard decisions," which can instantly destroy the converged feature manifold, causing the loss function to oscillate violently and easily leading to training instability or even collapse. This embodiment designs a "mitotic" evolution mechanism based on mask copying and perturbation, combined with a temperature annealing strategy. When the distribution complexity of a certain category of features exceeds a threshold, only its corresponding soft mask is copied and a small Gaussian noise is injected to break the symmetry, while keeping the shared weights unchanged. Subsequently, during the annealing process, new and old experts gradually complete the functional differentiation. This embodiment ensures seamless knowledge inheritance and continuous gradient backpropagation, enabling the model to maintain training stability during adaptive expansion, significantly shortening the convergence time and improving the final performance.

[0077] Fourth, it achieves explicit feature decoupling and adaptive noise reduction, effectively addressing the challenge of feature confusion under harsh channel conditions. Traditional methods model all feature dimensions uniformly. Once some frequency bands suffer from strong multipath fading or narrowband interference, the entire category prototype is contaminated, leading to a sharp decline in discrimination performance. This embodiment introduces two key constraints into the optimization objective: orthogonality constraint and sparsity constraint. The former forces different logic experts under the same category to focus on mutually exclusive feature subspaces, thereby achieving functional complementarity. The latter, through L1 regularization, prompts the mask to automatically close channels that are severely contaminated by noise or have redundant information. The two work together to enable the model to have strong adaptive noise reduction capabilities and feature fault tolerance in real electromagnetic environments with extremely low signal-to-noise ratios or strong interference, significantly improving the robustness and reliability of the system.

[0078] The radio frequency fingerprinting method based on adaptive soft subspace evolution proposed in this embodiment has significant application value in high-precision communication and large-scale Internet of Things scenarios. First, it is suitable for identifying specific radiation sources in complex dynamic channels, especially in highly time-varying environments such as satellite communications and UAV swarms. Traditional models lack robustness in the face of severe signal distortion caused by multipath propagation, Doppler shift, and weather disturbances. This embodiment achieves continuous high-precision authentication of device identity without increasing the number of physical parameters through an adaptive evolutionary logic expert subspace, meeting the low-power and high-reliability requirements of edge nodes.

[0079] Second, secure access control and intrusion detection for the Industrial Internet of Things (IIoT). In smart factories or power IoT systems with heterogeneous equipment and scarce samples, the zero-parameter redundancy architecture of this embodiment enables lightweight edge gateways to automatically construct fingerprint subspaces for new devices based on a very small number of observation signals, effectively identifying cloned terminals or fake base stations, and significantly reducing security deployment costs.

[0080] Third, intelligent authentication and anti-interference at the 6G physical layer. For high-frequency 6G scenarios such as terahertz / millimeter wave, hardware fingerprints are highly susceptible to interference. The mask-guided data enhancement in this embodiment performs style mixing only on the mask activation dimension when simulating extreme channel disturbances, strictly preserving the microscopic features of the physical layer. The system can quickly extract the fingerprint of the interference source after several interference observations, achieving physical layer identity removal and adaptive anti-interference coding.

[0081] The steps described above are merely for clarity in describing the technical solution. In actual implementation, they can be combined into one step, or certain steps can be broken down into multiple steps, as long as they involve the same logical relationship, they are all within the scope of protection of this application. Any insignificant modifications or designs added to the algorithm or process, as long as they do not change the core of the algorithm or process, are also within the scope of protection of this application.

[0082] In one embodiment, to verify the effectiveness of the radio frequency fingerprinting method based on adaptive soft subspace evolution proposed in this application (hereinafter referred to as the method), we conducted relevant simulation experiments.

[0083] The hardware platform used in this simulation experiment is: Intel(R) Core i5-13600KF processor with a main frequency of 3.50GHz and 32GB of memory.

[0084] The software platform used in this simulation experiment is: WINDOWS 11 operating system, MATLAB R2023b, Python 3.8, PyTorch 1.31.

[0085] The dataset used in this simulation experiment is a standard radio frequency fingerprint dataset (such as the LoRa-RFFI open source dataset), which contains 15 devices of the same model.

[0086] To simulate a cross-domain environment, the data is divided into: Source Domain: a high signal-to-noise ratio (SNR=20dB) line-of-sight transmission environment; Target Domain: a low signal-to-noise ratio (SNR varies from -10dB to 10dB) complex channel environment containing multipath fading and Doppler shift.

[0087] The comparison objects in this simulation experiment include: the basic CNN model (Base), the traditional hybrid expert model (Standard MoE, with added physics branches), and this method.

[0088] By comparing the traditional prototype network, relational network and the method of this invention, the recognition accuracy can be obtained as shown in Table 1.

[0089] Table 1: Recognition accuracy of various methods under different signal-to-noise ratios (cross-domain testing)

[0090] As shown in Table 1, under harsh channel conditions with a signal-to-noise ratio (SNR) of 0 dB, the recognition accuracy of this method is improved by approximately 16.2% compared to CNN and by approximately 7.1% compared to MoE using physical branch expansion. This result fully verifies the superiority of the soft subspace masking mechanism in modeling cross-domain, multimodal radio frequency signal distributions.

[0091] Of particular note is that this method achieves significant performance gains while introducing only lightweight learnable mask vectors, resulting in an overall parameter increase of only 0.2%. In contrast, traditional MoE schemes require nearly four times the model storage overhead to achieve similar performance. This extremely low parameter redundancy makes it particularly suitable for resource-constrained edge device deployment scenarios.

[0092] The effectiveness of the mask evolution mechanism. After enabling the mitotic evolution mechanism triggered by intra-class distance variance, the training loss of the model decreases more smoothly and the convergence speed is improved by about 30% when dealing with mixed datasets of clean signals and multipath interference signals, indicating that the mechanism effectively enhances the model's adaptability to complex distributions.

[0093] The effect of mask guidance enhancement. Comparative experiments show that the model exhibits significant performance fluctuations at low signal-to-noise ratios when using the traditional MixStyle, while the recognition stability improves by 4.5% after introducing mask guidance. This proves that soft masks can effectively suppress noise-sensitive dimensions and accurately protect the key features of device fingerprints.

[0094] Visual verification of the intersection constraint. After introducing an orthogonal loss term, the feature responses activated by different masks exhibit significant complementarity in the frequency domain. For example, mask A mainly responds to the amplitude envelope information, while mask B focuses on the phase bias features. This phenomenon intuitively verifies the successful decoupling and functional differentiation of logic experts in the feature space.

[0095] Another embodiment of this application proposes a radio frequency fingerprint recognition system based on adaptive soft subspace evolution. The details of the radio frequency fingerprint recognition system based on adaptive soft subspace evolution proposed in this embodiment are described in detail below. The following content is only for the convenience of understanding and is not necessary for implementing this solution.

[0096] Figure 3 This is a schematic diagram of the structure of an RF fingerprint recognition system based on adaptive soft subspace evolution proposed in this embodiment, including: a preprocessing module M1, a soft subspace construction module M2, an adaptive evolution monitoring module M3, a mask-guided data enhancement module M4, and a joint optimization module M5.

[0097] The preprocessing module M1 is used to preprocess the radio frequency signals used as training samples.

[0098] The soft subspace construction module M2 is used to construct logical classification centers corresponding to different categories based on the physical classification weight sharing mechanism and the soft mask generated by the improved soft mask generation algorithm with temperature annealing.

[0099] The adaptive evolution monitoring module M3 is used to input the preprocessed radio frequency signal into the feature extraction network for high-dimensional feature extraction to obtain high-dimensional features. During the training process, it continuously monitors the intra-class distance variance from each category's features to its logical classification center. When the intra-class distance variance exceeds the first preset threshold, it triggers the mask splitting mechanism to generate a new soft mask.

[0100] The mask-guided data augmentation module M4 is used to perform intra-group style blending on high-dimensional features using soft masks to obtain soft-mask-guided augmented features.

[0101] The joint optimization module M5 is used to calculate the cosine similarity between the enhanced features guided by the soft mask and all logical classification centers under the same category. It performs category prediction based on the extreme value routing strategy to obtain the RF fingerprint recognition result, and performs ternary joint optimization by combining classification loss, orthogonality constraint and sparsity constraint.

[0102] It is worth noting that all modules involved in this embodiment are logical modules. In practical applications, a logical module can be a physical module, a part of a physical module, or an organic combination of multiple physical modules. Furthermore, to highlight the innovative aspects of this application, this embodiment does not introduce modules that are not closely related to solving the technical problems proposed in this application. However, this does not mean that other modules are absent from this embodiment.

[0103] It is not difficult to see that this embodiment is a system embodiment corresponding to the above method embodiments, and this embodiment can be implemented in conjunction with the above method embodiments. The relevant technical details and technical effects mentioned in the above method embodiments are still valid in this embodiment, and will not be repeated here to reduce repetition. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the above method embodiments.

[0104] Another embodiment of this application provides an electronic device, such as Figure 4 As shown, it includes a processor C1 and a memory C2. The memory C2 stores instructions that the processor C1 can execute. When the processor C1 is configured to execute the instructions, the electronic device can implement a radio frequency fingerprint recognition method based on adaptive soft subspace evolution as described in the above method embodiment.

[0105] The memory and processor are connected via a bus, which includes any number of interconnecting buses and bridges. The bus connects various circuits of one or more processors and memories. The bus can also connect various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and will not be described further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by the processor is transmitted over the wireless medium via an antenna, which further receives data and transmits it to the processor.

[0106] The processor manages the bus and general processing, and also provides various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory is used to store data used by the processor during operation.

[0107] Another embodiment of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, can implement a radio frequency fingerprinting method based on adaptive soft subspace evolution as described in the above method embodiments.

[0108] That is, those skilled in the art will understand that all or part of the steps in the above method embodiments can be implemented by a program instructing related hardware. The program is stored in a storage medium and includes several instructions to cause a device (such as a microcontroller, chip, etc.) or processor to execute all or part of the steps of the method described in the method embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0109] It will be understood by those skilled in the art that the above embodiments are specific implementations of this application, and in practical applications, various changes in form and detail may be made without departing from the spirit and scope of this application. For those skilled in the art, several improvements and modifications can be made without departing from the principles of this application, and these improvements and modifications are also considered to be within the scope of protection of this application.

Claims

1. A radio frequency fingerprint recognition method based on adaptive soft subspace evolution, characterized in that, include: The radio frequency signals used as training samples are preprocessed, and the preprocessed radio frequency signals are input into the feature extraction network for high-dimensional feature extraction to obtain high-dimensional features; Based on the physical classification weight sharing mechanism, and combined with the soft mask generated by the improved soft mask generation algorithm with temperature annealing, logical classification centers corresponding to different categories are constructed. During training, the intra-class distance variance from the features of each category to their logical classification center is continuously monitored. When the intra-class distance variance exceeds the first preset threshold, the mask splitting mechanism is triggered to generate a new soft mask. By using soft masking to perform intra-group style blending on high-dimensional features, soft mask-guided enhanced features are obtained. The cosine similarity between the enhanced features guided by the soft mask and all logical classification centers under each category is calculated. Category prediction is performed based on the extreme value routing strategy to obtain the radio frequency fingerprint recognition result. The ternary joint optimization is then performed by combining classification loss, orthogonality constraint and sparsity constraint.

2. The radio frequency fingerprinting method based on adaptive soft subspace evolution according to claim 1, characterized in that, High-dimensional features include the device's inherent microscopic hardware fingerprint and domain offset noise introduced by the current channel. These high-dimensional features are denoted as... , , This represents the preprocessed radio frequency signal, i.e., the I / Q sequence. This represents a feature extraction network, which can be a one-dimensional ResNet or CNN. For dimension The complex field of .

3. The radio frequency fingerprinting method based on adaptive soft subspace evolution according to claim 2, characterized in that, Based on the physical classification weight sharing mechanism, and combined with the soft mask generated by the improved soft mask generation algorithm with temperature annealing, logical classification centers corresponding to different categories are constructed, including: Maintain a globally shared physical classification weight matrix , , For dimension The complex field of , the first The weight vector of a class is denoted as , Shared with all experts to ensure consistency of the underlying physical semantics. ; For the first The class of An expert initializes a learnable parameter vector. , ,based on Through temperature coefficient The sigmoid function generates a series of soft masks, the th... The class of A soft mask for experts Represented as: ； in, For the Sigmoid function, The value of is gradually annealed to 0 as training progresses; By combining physical weights with soft masks through Hadamard product, logical classification centers oriented towards a specific domain are generated. The class of Logical classification center Represented as: ； in, This represents the Hadamard product operation.

4. The radio frequency fingerprinting method based on adaptive soft subspace evolution according to claim 3, characterized in that, Trigger the mask splitting mechanism to generate a new soft mask, including: Copy the current soft mask parameters Apply a small Gaussian perturbation to it Generate two new soft masks and , , .

5. The radio frequency fingerprinting method based on adaptive soft subspace evolution according to claim 4, characterized in that, By using soft masking to perform within-group style blending on high-dimensional features, soft-mask-guided enhanced features are obtained, including: Based on the distance between each high-dimensional feature and each logical classification center, the high-dimensional features are grouped; each group corresponds to a category. Iterate through each group, and within the current group, exchange the mean and standard deviation of each high-dimensional feature to generate the preliminary mixed features of the current group; where, the th The initial mixed characteristics of the group are denoted as ; Style fusion is performed on the high-dimensional features and preliminary mixed features of the current group using the soft mask corresponding to the current group, resulting in soft-mask-guided enhanced features; where, the first Group 1 Enhanced features guided by a soft mask The calculation process is expressed as follows: , Indicates the first The soft mask corresponding to the group, For the first Group 1 A high-dimensional feature.

6. The radio frequency fingerprinting method based on adaptive soft subspace evolution according to claim 5, characterized in that, During inference, the cosine similarity between the enhanced features guided by the soft mask and all logical classification centers under each category is calculated. Category prediction is then performed based on an extreme value routing strategy to obtain the RF fingerprint recognition results, including: Iterate through each category, calculate the cosine similarity between the enhanced features guided by the soft mask and all logical classification centers under the current category, and take the maximum value as the prediction score of the current category; After traversing all categories, the category with the highest predicted score is selected as the RFID fingerprint recognition result. No. The calculation process for the predicted score of a class is expressed as follows: ； in, Enhanced features guided by soft masks, For the first The predicted score of the class. This indicates the calculation of the L2 norm.

7. The radio frequency fingerprinting method based on adaptive soft subspace evolution according to claim 6, characterized in that, A ternary joint optimization combining classification loss, orthogonality constraint, and sparsity constraint is achieved through the following formula: ； ； ； in, The total loss for ternary joint optimization, For classifying losses, For orthogonal constraints, For sparsity constraints, This indicates L1 regularization.

8. A radio frequency fingerprint recognition system based on adaptive soft subspace evolution, characterized in that, include: The preprocessing module is used to preprocess the radio frequency signals used as training samples; The soft subspace construction module is used to construct logical classification centers corresponding to different categories based on the physical classification weight sharing mechanism and the soft mask generated by the improved soft mask generation algorithm with temperature annealing. The adaptive evolution monitoring module is used to input the preprocessed radio frequency signal into the feature extraction network for high-dimensional feature extraction to obtain high-dimensional features. During the training process, it continuously monitors the intra-class distance variance from each category's features to its logical classification center. When the intra-class distance variance exceeds the first preset threshold, it triggers the mask splitting mechanism to generate a new soft mask. The mask-guided data augmentation module is used to perform intra-group style blending on high-dimensional features using soft masks to obtain enhanced features guided by soft masks. The joint optimization module is used to calculate the cosine similarity between the enhanced features guided by the soft mask and all logical classification centers under the same category. It performs category prediction based on the extreme value routing strategy to obtain the RF fingerprint recognition result, and performs ternary joint optimization by combining classification loss, orthogonality constraint and sparsity constraint.

9. An electronic device, characterized in that, include: The electronic device includes a processor and a memory, wherein the memory stores instructions that the processor can execute, and the processor is configured to execute the instructions such that the electronic device can implement a radio frequency fingerprinting method based on adaptive soft subspace evolution as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it can implement a radio frequency fingerprinting method based on adaptive soft subspace evolution as described in any one of claims 1 to 7.

Images