A Radiation Source Individual Identification Method Based on Dual-Branch Feature Decoupling

Abstract

This invention discloses a radiation source individual identification method based on dual-branch feature decoupling, constructing a dual-branch parallel decoupling architecture: a shared feature extractor extracts hybrid features; the identity branch extracts domain-invariant hardware fingerprint features; and the environment branch captures domain-specific features that drift over time. A dual-discriminator collaborative game mechanism is introduced, utilizing adversarial learning to actively remove time-domain information from the identity branch, while explicit supervision guides the environment branch to absorb non-stationary noise. A multi-dimensional joint optimization system including orthogonal constraints, center constraints, and entropy maximization regularization is established, coupled with a dynamic weight adjustment strategy to achieve stable convergence. This invention effectively eliminates the interference of environmental noise on identity features, significantly improving the model's generalization ability and recognition robustness in complex electromagnetic environments. Only the identity branch is retained during the inference stage, resulting in low computational complexity.

Description

Technical Field

[0001] This invention relates to the fields of radio signal processing and deep learning technology, specifically to a radiation source individual identification technology. Background Technology

[0002] Specific Emitter Identification (SEI) is a key technology in modern electronic reconnaissance, spectrum monitoring, wireless network security, and military countermeasures. Its core principle is based on the physical non-cloning property of hardware: during the manufacturing process of any radio frequency electronic device, manufacturing tolerances and microscopic physical defects are unavoidable in hardware components (such as power amplifiers, mixers, and digital-to-analog converters). These subtle non-ideal characteristics couple into the radio signals emitted by the device, forming unique and distinguishable "fingerprint" features. These fingerprint features are contained in multi-dimensional information such as signal amplitude, phase, frequency drift, and modulation trajectory, possessing uniqueness and stability, similar to human biometric fingerprints. Through in-depth analysis and feature extraction of received radio frequency signals, SEI technology can achieve unique authentication and identification of the transmitting source device.

[0003] To effectively capture and utilize these fingerprint features, existing radiation source individual identification technologies mainly employ the following two mainstream methods:

[0004] The first category is machine learning methods based on expert feature engineering. These methods focus on manually extracting the physical properties of signals, using signal processing algorithms to extract explicit fingerprint features, such as higher-order statistics, wavelet transform coefficients, integral bispectral features, and spurious component features. Subsequently, traditional classifiers such as Support Vector Machines (SVM), Random Forest, or K-Nearest Neighbors (KNN) are used to train and classify the extracted feature vectors, thereby achieving individual identification.

[0005] The second category is data-driven deep learning methods. With the development of artificial intelligence technology, this type of method has gradually become mainstream. Its basic idea is to directly utilize the powerful representation learning capabilities of deep neural networks (such as one-dimensional convolutional neural networks (CNNs), recurrent neural networks (RNNs), or residual networks (ResNets) without relying on complex manual feature design. By directly using raw signal data (such as I / Q orthogonal sequences, time-frequency diagrams, or constellation diagrams) as network input, the model can automatically learn the nonlinear mapping relationship from low-level signal waveforms to high-level semantic identities in massive amounts of data, thereby completing feature extraction and classification tasks end-to-end.

[0006] Although the above-mentioned technologies can achieve good recognition results in relatively stable or simple experimental environments, their inherent structural defects become increasingly prominent when facing increasingly complex electromagnetic environments, cross-domain scenarios, and long-term monitoring tasks in practical applications. These defects are mainly reflected in the following three aspects:

[0007] First, existing models generally suffer from severe entanglement between identity features and environmental features, lacking the ability to generalize to domain shifts. Traditional deep learning methods typically employ a black-box feature extraction approach, where models tend to use any information that helps reduce training error for classification. In real-world radio frequency signals, the inherent fingerprint (identity essence) of a radiation source is often highly coupled with channel noise, multipath effects, and the characteristics of the acquisition equipment (environmental appearance). Existing methods cannot distinguish between these two types of features, often mistakenly treating time-drifting environmental features as identity features for learning. This feature entanglement leads to severe performance degradation when models face unknown domains (such as new environments or channels) or perform long-term cross-domain recognition (such as equipment aging or seasonal changes) due to significant changes in the distribution of environmental features, resulting in domain overfitting.

[0008] Secondly, existing architectures lack explicit feature decoupling mechanisms, making it impossible to construct independent domain-invariant subspaces within the feature space. Most existing technologies map all extracted information into the same hybrid feature space, lacking corresponding structural designs to physically isolate features with different attributes. The model cannot enforce at the network structure level that identity fingerprints and environmental interference are independent of each other. Due to the lack of such explicit structural constraints, the model struggles to ensure that the final classification decision relies solely on stable hardware fingerprints, making it highly susceptible to interference from non-steady-state environmental noise. This results in difficulty in guaranteeing the reliability and stability of recognition in complex dynamic environments.

[0009] Finally, existing training strategies are simplistic and lack joint optimization mechanisms based on multi-dimensional collaborative constraints. Traditional methods often rely on a single classification cross-entropy loss for parameter updates. This optimization approach only focuses on the classification accuracy of samples in the source domain, neglecting the intrinsic geometric structure and domain adaptability of feature representations. Due to the lack of multi-objective collaborative optimization, the model struggles to proactively remove environmental noise to purify identity fingerprints, thus limiting its robustness under long-term cross-domain conditions. Summary of the Invention

[0010] The technical problem to be solved by this invention is to address the shortcomings of existing radiation source individual identification technologies, which suffer from poor generalization ability and severely degraded recognition performance in cross-domain and long-term scenarios due to the high coupling between essential identity features and environmental interference features. This invention provides a radiation source individual identification technology that can separate non-steady-state environmental information from identity features, thereby extracting pure and stable hardware fingerprint features.

[0011] The technical solution adopted by this invention to solve the above-mentioned technical problems is a method for individual radiation source identification based on dual-branch feature decoupling, comprising the following steps:

[0012] Signal preprocessing and shared feature extraction steps: Receive the raw radio frequency signal, perform preprocessing to obtain a one-dimensional time-domain signal sample, and extract shared feature representations through a shared feature extractor;

[0013] Dual-branch feature projection and decoupling steps: The shared feature representation is input into the identity branch and the environment branch respectively, and the identity feature representation and the environment feature representation are output; the identity branch is used to extract domain-invariant hardware fingerprint features, and the environment branch is used to capture domain-specific features that drift over time;

[0014] The dual-discriminator adversarial constraint steps are as follows: After the identity branch, connect the identity domain discriminator, perform adversarial learning in conjunction with the gradient reversal layer, and calculate the adversarial loss; after the environment branch, connect the environment domain discriminator, perform domain classification supervision, and calculate the domain classification loss.

[0015] Collaborative constraint steps: Apply orthogonal constraints, central constraints, and entropy maximization regularization to the identity feature representation and the environment feature representation, and calculate the orthogonal constraint loss, central constraint loss, and entropy maximization regularization loss respectively;

[0016] The main task classification and joint optimization steps are as follows: input the identity feature representation into the classifier, calculate the main task classification loss, construct a joint objective function that includes the main task classification loss, adversarial loss, domain classification loss, orthogonal constraint loss, center constraint loss and entropy maximization regularization loss, and use a dynamic weight adjustment strategy for joint optimization.

[0017] Model inference steps: After training, discard the environment branch and all discriminator modules, retain only the shared feature extractor, identity branch and classifier, and perform forward propagation on the test signal to obtain the recognition result.

[0018] Specifically,

[0019] A dual-branch parallel network architecture is constructed, explicitly building two independent feature projection subspaces for identity and environment within a unified feature extraction framework. Through physical isolation, the model is able to extract time-varying, non-stationary environmental information from identity features, ensuring that the final classification decision relies solely on hardware fingerprint features.

[0020] A collaborative game mechanism based on dual discriminators is introduced to impose push-pull constraints on the feature space. On the one hand, an adversarial learning mechanism (gradient inversion) is used to actively remove time-domain information from the identity branch to prevent the infiltration of environmental noise; on the other hand, an explicit supervision mechanism is used to guide the environment branch to actively absorb and lock in differential features such as time and channel. Through this two-way collaboration, the structural decoupling of features is forced.

[0021] Simultaneously, a multi-dimensional joint optimization system is established, comprising orthogonal constraints, central constraints, and entropy maximization regularization. Orthogonal constraints minimize the leakage of mutual information between two branches; central constraints ensure that the decoupled identity features remain compact within the class; entropy maximization further eliminates residual domain information in identity features; and a dynamic weight adjustment strategy is employed to achieve stable convergence and efficient optimization of the model under multi-objective constraints.

[0022] The beneficial effects of this invention are:

[0023] Significantly improves recognition robustness in cross-domain and long-term scenarios: By explicitly decoupling the two branches, non-steady-state environmental information is separated from identity features, so that classification decisions rely only on stable hardware fingerprint features, effectively mitigating performance degradation caused by environmental drift, channel changes and device aging.

[0024] Achieving more thorough identity-environment feature separation: Unlike the single adversarial domain adaptation method in the prior art, the dual discriminator push-pull collaborative mechanism of the present invention applies constraints to the feature space from both positive and negative directions, achieving more thorough and controllable feature decoupling.

[0025] Maintaining the strong discriminative power of decoupled identity features: Orthogonal constraints prevent cross-branch information leakage, central constraints maintain the tight clustering of features of similar samples in space, and entropy maximization regularization completely eliminates residual domain information in identity features. The joint optimization of these three factors avoids the feature divergence problem common in adversarial training.

[0026] Efficient and deployment-friendly in the inference phase: After training, the environment branch and discriminator module can be completely discarded, and only the identity branch is retained for classification inference, achieving high robustness while maintaining low computational complexity and memory usage. Attached Figure Description

[0027] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0028] This invention is a radiation source individual identification architecture that integrates shared feature extraction, dual-branch dynamic decoupling, and dual-discriminator collaborative constraints. This architecture maps the original signal to a decoupled feature space through explicit structural design and multi-dimensional loss constraints, thereby solving the identification challenge in long-term, cross-domain scenarios. Specifically, it includes the following components:

[0029] 1. Shared Feature Extraction Module: This module serves as the system's perceptual foundation, responsible for general high-dimensional feature encoding of the input raw radio frequency signal (such as a one-dimensional time-domain sequence). It employs ResNeXt1D as the backbone network, leveraging its grouped convolution characteristics to reduce parameter redundancy while enhancing the nonlinear representation capability of high-dimensional signals. This module maps the raw signal into a shared feature vector rich in mixed semantics (containing both device fingerprints and environmental noise), providing a unified information foundation for subsequent decoupling operations.

[0030] 2. Dual-branch dynamic decoupling module: This module, connected after the shared feature extraction module, is the core component of this invention. It introduces two parallel and structurally independent projection branches, explicitly decomposing the mixed shared features into two non-interfering subspaces:

[0031] The Identity Branch employs a non-linear mapping structure and is responsible for extracting domain-invariant device hardware fingerprint features from shared features. During training, this branch is jointly constrained by the main task classification loss and adversarial gradients, aiming to extract feature representations that contain only identity information while discarding temporal information. .

[0032] Environment Branch: This branch also employs a nonlinear mapping structure, but it is completely independent of the Identity Branch in terms of parameter space. It is specifically designed to capture domain-specific features such as time-varying channel interference and acquisition noise. During training, this branch is guided by explicit domain classification supervision signals, aiming to extract environmental feature representations that maximize the distinction between different time domains. .

[0033] Through this parallel dual-channel design, combined with different subsequent supervision signals, the model can automatically learn to direct the information flow of domain-invariant features and domain-correlated features to different parameter branches respectively.

[0034] 3. Dual Discriminator and Cooperative Constraint Module: This module uses two complementary discriminators to apply push-pull game constraints to the two branches mentioned above.

[0035] Push (Adversarial Removal): The first discriminator is connected after the identity branch and performs adversarial learning in conjunction with a gradient inversion layer (GRL). Its goal is to distinguish the temporal domain of the signal as much as possible, while the identity branch attempts to confuse the discriminator. This game forces the identity branch to actively remove any feature components containing temporal context information.

[0036] Supervised Adsorption: The second discriminator is attached to the environment branch and performs explicit domain classification supervised learning. It guides the environment branch to maximize the retention of temporal environmental differences, enabling it to actively adsorb non-steady-state noise, thereby indirectly purifying the identity branch.

[0037] In addition, this module introduces orthogonal constraints to prevent cross-branch information leakage, central constraints to maintain intra-class compactness, and entropy maximization regularization to further solidify the decoupling effect in geometric space.

[0038] 4. Independent Classification and Joint Optimization Module: This module is connected to the output of the identity branch and is responsible for the final recognition decision and model training.

[0039] Independent Classification: The classifier only accepts decoupled and purified identity features as input, and outputs the class probability of the individual radiation source through a fully connected layer and a Softmax layer. Since the input features have been stripped of environmental interference, the classification decision will strictly rely on a stable hardware fingerprint.

[0040] Joint Optimization: This module defines a joint objective function that includes the main task loss, adversarial loss, domain classification loss, and multidimensional geometric constraint loss, and employs a dynamic weight adjustment strategy. During training, it adaptively balances the gradient contributions of each module to ensure that the model simultaneously achieves optimal performance on both feature decoupling and identity recognition objectives.

[0041] The embodiments of the present invention will be described in detail below with reference to the technical solution. The present invention proposes a radiation source individual identification network based on dual-branch feature decoupling, which aims to effectively separate the essential identity features and environmental interference features in a unified feature space through explicit structural design and multi-dimensional collaborative constraints, thereby improving the robustness of the model in long-term cross-domain scenarios.

[0042] Next, the following operations will be performed sequentially to complete the processing and classification of the radiation source signals. The specific steps are as follows:

[0043] Step 1: Signal preprocessing and shared feature extraction.

[0044] The raw radio frequency signal from the radiation source device is received, and preprocessed by down-conversion, filtering, and slicing to obtain a fixed-length one-dimensional time-domain signal sample. ,in For signal length. Through a shared feature extractor Feature extraction is performed to obtain shared feature representations. :

[0045] ;

[0046] in, Batch Size. The feature dimension is defined by the ResNeXt1D network architecture, which utilizes grouped convolution to extract a general high-dimensional semantic representation. The feature vector at this point contains both device fingerprint information and is coupled with environmental noise and channel features.

[0047] Specifically, shared feature extractor The ResNeXt1D network architecture is adopted, which utilizes the characteristics of grouped convolution to enhance the nonlinear representation capability of high-dimensional signals while reducing parameter redundancy.

[0048] Step 2: Bi-branch feature projection and decoupling.

[0049] In shared features Based on this, a two-branch dynamic decoupling module is used to project it into two orthogonal subspaces. This module contains two nonlinear projection branches with identical structures but independent parameters:

[0050] 2.1 Identity Branch: This branch aims to extract domain-invariant hardware fingerprint features. It consists of a fully connected layer, a batch normalization (BN) layer, and an activation function, serving as the identity branch mapping function. Input shared features Output identity feature representation :

[0051] ;

[0052] in, This is the dimension of identity features.

[0053] This branch preserves key identity features in G(x) through identity mapping, avoiding excessive transformation that could lead to fingerprint information loss. This feature It will be used for primary task classification, therefore it is required to be sensitive to device identity but not to environmental changes.

[0054] 2.2 Environment Branch: This branch aims to capture domain-specific features that drift over time. It also consists of fully connected layers and nonlinear activations, with parameters independent of the identity branch, and serves as a mapping function for the environment branch. Input shared features Output environment feature representation :

[0055] ;

[0056] in, This refers to the environmental characteristics dimension.

[0057] This branch learns environment-related feature representations independently through independent linear transformations and nonlinear activations.

[0058] Step 3: Dual discriminators to counteract constraints.

[0059] To drive the above decoupling process, two complementary discriminators are introduced:

[0060] 3.1 Identity Domain Discriminator: This discriminator Following the identity branch, adversarial training is implemented through a Gradient Reversal Layer (GRL). The goal is to determine identity as accurately as possible. The time domain in which it belongs, and the goal of the feature extractor is to maximize the classification error of the discriminator. Adversarial loss. Defined as:

[0061] ;

[0062] in, For sample index, For domain category indexing, The total number of domain categories, For the first The sample at the th The actual domain label on each domain. This represents the probability distribution output by the discriminator. GRL stands for Gradient Reversal Layer. Let be the identity features of the i-th sample.

[0063] By flipping the inverse gradient of GRL, the identity branch is forced to actively remove feature components containing time-domain information.

[0064] 3.2 Environment Domain Discriminator: This discriminator Following the environment branch, a standard supervised classification task is performed. Its goal is to maximize the classification accuracy of environment domain labels, guiding... Actively absorbs environmental differences. Domain classification loss. Defined as:

[0065] ;

[0066] in, Let be the environmental features of the i-th sample.

[0067] This allows the environment branch to effectively absorb non-steady-state interference components in the environment, thereby indirectly filtering out environmental noise in the identity branch at the feature level.

[0068] Step 4: Collaborative Constraint Mechanism. To further optimize the geometric structure of the feature space, the following three constraint terms are introduced:

[0069] 4.1 Orthogonality Constraint: This is achieved by calculating the correlation matrix between the identity feature matrix and the environmental feature matrix. And minimize its Frobenius norm to obtain the orthogonal constraint loss. :

[0070] ;

[0071] in, The normalized identity feature matrix, This is the normalized environmental feature matrix. For transpose, It is an L2 norm.

[0072] Force the two subspaces to be perpendicular to each other to prevent information leakage across branches.

[0073] 4.2 Center Constraint: This is achieved by calculating the feature centers for each category. And minimize the distance from the sample to the corresponding class center:

[0074] ;

[0075] ;

[0076] Where k is the identity category index, Let K be the feature center of the k-th category, and K be the total number of identity categories. Loss is centered on constraint.

[0077] To prevent intra-class divergence of features caused by adversarial training, maintain the tight clustering of similar samples in the feature space, and preserve the discriminative power of identity features.

[0078] 4.3 Entropy Maximization: To further improve... Domain uncertainty is addressed by explicitly maximizing the environment discriminator. identity characteristics Predicted entropy value Obtain the entropy-maximizing regularized loss :

[0079] ;

[0080] in, It is the entropy function. This represents the probability that the i-th sample output by the environment domain discriminator belongs to the d-th domain.

[0081] force It exhibits the greatest uncertainty (i.e. uniform distribution) in the face of the environment discriminator, completely eliminating residual domain information.

[0082] Step 5: Classification and joint optimization of main tasks.

[0083] Decoupled identity features The input is fed into the classifier (Softmax layer) to calculate the cross-entropy loss, which is the main task for identifying individual radiation sources. :

[0084] ;

[0085] in and These are learnable classification weights. The loss is driven by minimizing the classification cross-entropy loss. Learning discriminative identity representation:

[0086] ;

[0087] in, The true label indicates whether the i-th sample belongs to the k-th class; To predict the probability, we have the probability that the i-th sample belongs to the k-th class.

[0088] Final total loss function A weighted combination of multiple objectives:

[0089] ;

[0090] in, , , , , , These are the weight coefficients for the main task classification loss, adversarial loss, domain classification loss, orthogonal constraint loss, center constraint loss, and entropy maximization regularization loss.

[0091] Step 6: Dynamic weight adjustment strategy. Specifically,

[0092] A dynamic weight adjustment strategy is employed to balance the training process for multi-objective optimization. Based on experience with DANN (Domain Adversarial Neural Network), this strategy is suitable for adversarial loss weights. Sum domain loss weights Set a weight coefficient that changes dynamically with the training progress, that is, = = As training progresses It exhibits Sigmoid growth. The value range is 0 to 1. This indicates the start of training. This indicates the end of training:

[0093] ;

[0094] This allows the model to focus more on the main task in the early stages and gradually increase the influence of adversarial constraints in the later stages.

[0095] Step 7: Model Inference

[0096] After the model training is complete, discard the environment branch. And all discriminator modules. Only the shared feature extractor is retained. Identity Branch and the main classifier .

[0097] For test signals The recognition result is obtained directly through forward propagation:

[0098] ;

[0099] Because auxiliary branches are eliminated during the inference phase, the model maintains low computational complexity.

[0100] The above steps constitute the complete implementation process of the radiation source individual identification method based on dual-branch feature decoupling proposed in this invention. This method explicitly separates identity and environmental features through a dual-branch parallel architecture, achieves structural decoupling through push-pull constraints of dual discriminators, and maintains feature discriminability through a multi-dimensional joint optimization system. It effectively solves the problems of poor generalization ability and severe feature coupling in existing technologies in cross-domain and long-term scenarios.

Claims

1. A method for identifying a radiation source individual based on dual-branch feature decoupling, characterized in that, Includes the following steps: Signal preprocessing and shared feature extraction steps: Receive the raw radio frequency signal, perform preprocessing to obtain a one-dimensional time-domain signal sample, and extract shared feature representations through a shared feature extractor; Dual-branch feature projection and decoupling steps: The shared feature representation is input into the identity branch and the environment branch respectively, and the identity feature representation and the environment feature representation are output; the identity branch is used to extract domain-invariant hardware fingerprint features, and the environment branch is used to capture domain-specific features that drift over time; The dual-discriminator adversarial constraint steps are as follows: After the identity branch, connect the identity domain discriminator, perform adversarial learning in conjunction with the gradient reversal layer, and calculate the adversarial loss; after the environment branch, connect the environment domain discriminator, perform domain classification supervision, and calculate the domain classification loss. Collaborative constraint steps: Apply orthogonal constraints, central constraints, and entropy maximization regularization to the identity feature representation and the environment feature representation, and calculate the orthogonal constraint loss, central constraint loss, and entropy maximization regularization loss respectively; The main task classification and joint optimization steps are as follows: input the identity feature representation into the classifier, calculate the main task classification loss, construct a joint objective function that includes the main task classification loss, adversarial loss, domain classification loss, orthogonal constraint loss, center constraint loss and entropy maximization regularization loss, and use a dynamic weight adjustment strategy for joint optimization. Model inference steps: After training, discard the environment branch and all discriminator modules, retain only the shared feature extractor, identity branch and classifier, and perform forward propagation on the test signal to obtain the recognition result.

2. The method as described in claim 1, characterized in that, The shared feature extractor adopts the ResNeXt1D network architecture and uses the characteristics of grouped convolution to extract features.

3. The method as described in claim 1, characterized in that, The identity branch and the environment branch are nonlinear projection branches with the same structure but independent parameters.

4. The method as described in claim 1, characterized in that, The resistance loss Represented as: ； in, For sample index, For batch size, For domain category indexing, The total number of domain categories, For the first The sample at the th The actual domain label on the domain. The probability distribution is the output of the discriminator, and GRL is the gradient inversion layer. Let be the identity features of the i-th sample.

5. The method as described in claim 1, characterized in that, Domain classification loss Represented as: ； in, For sample index, For batch size, For domain category indexing, The total number of domain categories, For the first The sample at the th The actual domain label on the domain. The probability distribution is the output of the discriminator, and GRL is the gradient inversion layer. Let be the environmental features of the i-th sample.

6. The method as described in claim 1, characterized in that, The orthogonal constraint loss Represented as: ； in, For sample index, For batch size, Let i be the identity features of the i-th sample. For the environmental features of the i-th sample, It is an L2 norm.

7. The method as described in claim 1, characterized in that, The central constraint loss Represented as: ； ； in, For sample index, For batch size, Let k be the identity feature of the i-th sample, and k be the identity category index. Let K be the feature center of the k-th category, and K be the total number of identity categories.

8. The method as described in claim 1, characterized in that, The entropy maximization regularization loss Represented as: ； in, For sample index, For batch size, For domain category indexing, The total number of domain categories, This represents the probability that the i-th sample output by the environment domain discriminator belongs to the d-th domain.

9. The method as described in claim 1, characterized in that, The joint objective function is expressed as: ； in, Losses are categorized by primary task. To combat the losses, For domain classification loss, For orthogonal constraint loss, Loss is centered on constraint. To maximize the regularized loss for entropy, , , , , , These are the weighting coefficients for each loss term.

10. The method according to claim 9, characterized in that, In the dynamic weight adjustment strategy, the weight against loss is... Domain classification loss rights Set as weight coefficients that change dynamically with the training progress , p increases in a Sigmoid manner as training progresses: ； The training progress p ranges from [0,1], where p=0 indicates the start of training and p=1 indicates the end of training.

Images