A specific radiation source identification method based on a Bayesian attention mechanism

Abstract

This invention belongs to the field of wireless communication signal processing and deep learning technology, and discloses a method for identifying specific radiation sources based on a Bayesian attention mechanism. First, the invention performs multi-scale frequency-normalized short-time Fourier transform on the baseband I / Q signal to generate a multi-channel time-frequency feature map without carrier masking. Then, it constructs a lightweight neural network with ResNet18 as the backbone and a Bayesian frequency attention module embedded in the front end. Finally, it trains the model using a loss function including an evidence lower bound and a KL divergence annealing strategy. This invention explicitly eliminates the masking effect of carrier energy on weak fingerprint features through multi-scale frequency normalization, and introduces weight probability distribution and uncertainty quantization through Bayesian frequency attention, effectively preventing overfitting under small sample conditions. This invention maintains high recognition accuracy while possessing low computational complexity and low inference latency, making it suitable for real-time radiation source identification in edge devices.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication signal processing and deep learning technology, and particularly relates to a method for identifying specific radiation sources based on a Bayesian attention mechanism. Background Technology

[0002] With the increasing complexity of electromagnetic environments, the issues of authentication and secure access for wireless devices are becoming increasingly prominent. Specific Source Identification (SEI) technology, which utilizes unavoidable and non-clonable physical fingerprints (such as phase noise, I / Q imbalance, and power amplifier nonlinearity) generated during transmitter hardware manufacturing to distinguish different individual radiation sources, has become one of the key technologies for ensuring wireless network security. Currently, although deep learning-based SEI methods have made some progress in experimental environments, existing identification algorithms still face the following serious challenges in real-world complex electromagnetic environments and edge deployment scenarios:

[0003] First, due to the "spectral masking effect," weak fingerprint features are difficult to extract effectively. In actual communication signals, the energy of spurious and harmonic components carrying hardware fingerprint information is usually very weak, often several orders of magnitude lower than the energy of the main carrier signal. In the spectrum directly generated by traditional time-frequency analysis methods (such as Short-Time Fourier Transform (STFT), the high-energy carrier signal dominates, causing weak fingerprint features to be "submerged" under the carrier energy. Existing preprocessing methods lack an explicit energy balancing mechanism, making it difficult for neural networks to focus on key discriminative features, thus limiting recognition accuracy.

[0004] Secondly, models with limited sample sizes exhibit poor generalization ability and are prone to overfitting. Traditional convolutional neural networks (CNNs) or attention mechanisms are typically deterministic, meaning that for a given input, the weights within the network have fixed values. In non-cooperative scenarios, the number of labeled radiation source samples obtained is often very limited (e.g., only tens to hundreds of samples per class). In situations of data scarcity, deterministic networks are highly susceptible to overfitting to random noise in the training data, leading to a sharp decline in the model's generalization performance on the test set, making it difficult to meet the needs of practical applications.

[0005] Finally, existing high-precision models have high computational complexity, making them difficult to deploy on edge devices. To improve feature extraction capabilities, current research tends to use deep networks (such as ResNet50) or large Transformer-based models (such as ViT). While these models offer high accuracy, they have a huge number of parameters, high floating-point operations (FLOPs), and inference latency often reaches tens of milliseconds or even higher, failing to meet the millisecond-level real-time requirements of resource-constrained devices such as handheld terminals and drones.

[0006] In summary, existing radiation source identification methods struggle to achieve a balance between interference resistance in feature extraction, robustness under small sample sizes, and computational efficiency. Therefore, there is an urgent need to propose a lightweight, effective method for identifying specific radiation sources that can effectively remove carrier masking, possess uncertainty quantification capabilities, and is designed for efficient radiation source identification. Summary of the Invention

[0007] The purpose of this invention is to provide a specific radiation source identification method based on the Bayesian attention mechanism, so as to solve the technical problems of existing technologies such as difficulty in feature extraction under strong carrier masking, easy overfitting under small sample conditions, and limited edge computing resources.

[0008] To address the aforementioned problems in existing technologies, this invention provides a method for identifying specific radiation sources based on multi-scale normalization and Bayesian frequency attention. This invention explicitly removes carrier energy interference through a multi-scale frequency normalization strategy, introduces uncertainty modeling using a Bayesian frequency attention module to prevent overfitting, and combines it with a ResNet18 backbone network to achieve lightweight identification.

[0009] The specific technical solution of the present invention is as follows:

[0010] A method for identifying specific radiation sources based on a Bayesian attention mechanism includes the following steps:

[0011] Step 1: Obtain the baseband I / Q signal of the radiation source to be identified, perform multi-scale frequency normalized short-time Fourier transform on the baseband I / Q signal, and generate a multi-channel time-frequency feature map with carrier masking removed;

[0012] Step 2: Construct a lightweight neural network model based on Bayesian frequency attention; the lightweight neural network model uses ResNet18 as the backbone network and embeds a Bayesian frequency attention module at its front end to perform frequency domain weighting on the input time-frequency feature map;

[0013] Step 3: Construct a loss function that includes a lower bound on evidence, and train the model using the KL divergence annealing strategy; optimize the network parameters by learning the posterior distribution of the frequency weights through variational inference.

[0014] Step 4: Input the signal to be tested into the trained neural network model, and determine the identity category of the radiation source through the probability distribution output by the model.

[0015] Furthermore, the multi-scale frequency normalized short-time Fourier transform in step 1 specifically includes: first, performing a short-time Fourier transform on the baseband I / Q signal to obtain the original time-frequency matrix; then, setting multiple different time scale factors, performing time-domain energy normalization operations on each scale, linearly mapping the energy value of each frequency point on the time axis to a fixed interval to eliminate the absolute energy difference between different frequency points; finally, stitching the normalized spectrum diagrams generated by the I and Q paths at different scales along the channel dimension to form a multi-channel input tensor.

[0016] For baseband signals A short-time Fourier transform (STFT) is performed to obtain the original time-frequency matrix. To capture fingerprint features at different temporal granularities, three time scale factors are set. .

[0017] For each scale, time-domain energy normalization is performed, calculated using the following formula:

[0018]

[0019] in, Indicates time ,frequency Power spectral density at; and These represent the frequency components respectively. Throughout the timeline The maximum and minimum energy values ​​on the surface; To prevent tiny constants with a denominator of zero.

[0020] This step eliminates the huge differences in absolute energy between different frequency points through normalization (i.e., eliminates carrier masking), forcing the network to focus on the relative fluctuation characteristics of each frequency point over time.

[0021] Finally, the normalized spectrograms generated by the I-path and Q-path at the three scales (Scale 1, 10, 20) are concatenated along the channel dimension to form a spectrum with dimension 1. Input tensor .

[0022] Furthermore, the construction method of the Bayesian frequency attention module in step 2 includes: performing global average pooling on the input feature map to obtain a global frequency context vector; predicting the distribution parameters of the frequency weights using the prior network and the posterior network respectively, assuming that the weights follow a log-normal distribution; generating random attention weights using reparameterized sampling technology; and applying the generated random weights to the frequency dimension of the input feature map to obtain the enhanced feature map. The reparameterized sampling introduces a random noise vector sampled from a standard normal distribution, combines it with the mean vector output by the posterior network and a preset variance hyperparameter, to generate random attention weights that follow a log-normal distribution.

[0023] The BFA module aims to learn the probability distribution of frequency attention rather than fixed values. Specifically, it includes:

[0024] 1. Global information aggregation: This involves processing the input feature map. Perform global average pooling (GAP) to obtain the global frequency context vector. :

[0025]

[0026] in, Indicates: the number of channels in the multi-channel time-frequency feature map; Indicates: Time bins; Represents: Local element values ​​of the input time-frequency feature map, that is, the specific values ​​of the feature tensor at the c-th channel, the f-th frequency point, and the t-th time step.

[0027] Note: Average pooling along the C and T dimensions is performed to extract the average energy response of a specific frequency across all channels in global time.

[0028] 2. Distribution Parameter Prediction: The distribution parameters of the frequency weights are predicted using a prior net and a posterior net, respectively. It is assumed that the weights follow a log-normal distribution, and the posterior net outputs the mean. .

[0029] 3. Reparameterization Trick: Introducing standard Gaussian noise. Random attention weights are generated using the following formula. :

[0030]

[0031] in, To control the hyperparameters of the distribution variance; The term "posterior mean" refers to the mean vector of the weights for the f-th frequency point, inferred by the posterior network based on the current input.

[0032] 4. Feature weighting: The generated random weights are... The frequency dimension of the input feature map is applied to obtain the enhanced feature map. .

[0033] Furthermore, the loss function in step 3 is composed of a weighted sum of the cross-entropy classification loss and the KL divergence regularization loss, and the total loss function is... :

[0034]

[0035] in, Cross-entropy is used for classification loss to ensure recognition accuracy; Kullback-Leibler divergence is used to constrain the posterior distribution. Approximating the prior distribution (Usually set to a standard normal distribution), which serves a regularization function.

[0036] Furthermore, in the KL divergence annealing strategy in step 3, the weight coefficient of KL divergence increases linearly with the training rounds: in the early stage of training, the weight coefficient is zero or extremely small, so that the model focuses on learning classification features; after the warm-up period, the weight coefficient gradually increases to the target value, and Bayesian constraints are gradually introduced to solve the problem that Bayesian neural networks are difficult to converge.

[0037] Annealing strategy: These are the weighting coefficients for the KL divergence. During training, It is not fixed, but increases linearly with the number of training epochs:

[0038]

[0039] in, For the current training round, This is the preheating cycle.

[0040] Furthermore, the backbone network in step 2 contains four residual blocks, each consisting of two convolutional layers, a batch normalization layer, and a ReLU activation function, ultimately outputting a feature vector with a dimension of 512.

[0041] The method for identifying specific radiation sources based on the Bayesian attention mechanism of the present invention has the following advantages:

[0042] 1. Solved the problem of spectrum masking effect: Through innovative multi-scale frequency normalization (MSFN) preprocessing, strong carrier energy is explicitly removed, allowing weak hardware fingerprint features to be revealed, and the recognition accuracy is improved from 62.87% (traditional STFT) to over 97% under the same conditions.

[0043] 2. Robust recognition under small sample conditions: By using the Bayesian Frequency Attention (BFA) module to introduce the probability distribution of weights and KL divergence constraints, the model can still maintain high accuracy in scenarios with scarce data (such as 100 samples per class), effectively overcoming the overfitting problem of traditional deterministic networks.

[0044] 3. Balancing high accuracy and low latency: Through lightweight design, this invention achieves 98.49% high accuracy while maintaining extremely low computational complexity (2.22G FLOPs) and inference latency (2.53ms), providing a feasible technical solution for the real-time application of specific radiation source identification technology in edge devices such as drones and handheld terminals. Attached Figure Description

[0045] Figure 1 is a schematic diagram of the overall process of the method proposed in the embodiment of the present invention;

[0046] Figure 2 is a schematic diagram of the principle and process of multi-scale frequency normalized short-time Fourier transform (MSFN-STFT) in an embodiment of the present invention;

[0047] Figure 3 is a schematic diagram of the lightweight neural network structure based on Bayesian frequency attention (BFA) constructed according to an embodiment of the present invention;

[0048] Figure 4 is a diagram showing the internal structure and data flow of the Bayesian Frequency Attention (BFA) module in an embodiment of the present invention.

[0049] Figure 5 is a confusion matrix diagram of the identification results of an embodiment of the present invention;

[0050] Figure 6 This is a t-SNE clustering visualization of 30 types of radiation source characteristics. Detailed Implementation

[0051] To better understand the purpose, structure, and function of this invention, the following detailed description of a specific radiation source identification method based on a Bayesian attention mechanism is provided in conjunction with the accompanying drawings.

[0052] like Figure 1 As shown, a specific radiation source identification method based on Bayesian attention mechanism according to the present invention includes the following steps:

[0053] Step 1: Obtain the baseband I / Q signals of the radiation source to be identified and construct an active interference dataset containing different individual radiation sources;

[0054] In this embodiment, a dataset is constructed using real-world commercial walkie-talkie signals, specifically:

[0055] Step 1.1: Data Acquisition: The data acquisition targets are 30 commercial communication radios (radiation sources), including 15 radios of model A and 15 radios of model B, to verify the algorithm's ability to distinguish fingerprints of devices of the same model.

[0056] Step 1.2: Signal parameters: The baseband signal sampling rate is 600 kHz. The acquisition environment includes real channel noise, and different signal-to-noise ratio environments are simulated by adding additive white Gaussian noise (AWGN) in subsequent experiments.

[0057] Step 1.3: Data Segmentation: The acquired continuous I / Q signal stream is segmented into non-overlapping sample segments. Each sample segment contains 15,000 sampling points.

[0058] Step 1.4: Dataset Construction: Collect 2,500 sample fragments from each radiation source, for a total of 75,000 samples. Divide the dataset into training, validation, and test sets in a 4:1:5 ratio.

[0059] Step 2: Perform multi-scale frequency normalized short-time Fourier transform (MSFN-STFT) on the dataset to generate a multi-channel time-frequency feature map with carrier masking removed;

[0060] like Figure 2 As shown, to address the problem of strong carrier energy masking weak fingerprints, this invention designs an MSFN-STFT preprocessing module. This step specifically includes:

[0061] Step S2.1: Original time-frequency transformation:

[0062] For each I / Q signal sample Perform short-time Fourier transform (STFT) on each.

[0063] Set the STFT window length to 256 and the number of overlap points to 128.

[0064] Calculate the power spectral density matrices of the in-phase component (I-path) and the quadrature component (Q-path) respectively. and .

[0065] Step S2.2: Multi-scale segmentation:

[0066] To capture the performance of fingerprint features at different temporal granularities (such as instantaneous phase jitter and long-term frequency drift), a time scale factor is introduced. This embodiment selects three scales. :

[0067] Scale 1: Preserves the original temporal resolution for capturing high-frequency transient fingerprints.

[0068] Scale 10: Downsamples the time axis by 10 times (or average pooling) to capture medium-scale symbolic changes.

[0069] Scale 20: Downsamples the time axis by 20 times to capture long-term frequency drift trends.

[0070] Step S2.3: Time-domain energy normalization (a core step of this invention):

[0071] For the time-frequency plot at each scale, perform time-domain energy normalization to eliminate absolute energy differences between frequency points.

[0072] For any frequency point Its timeline The energy distribution on the surface is rescaled. The normalization formula is as follows:

[0073]

[0074] in:

[0075] The original power spectrum values ​​represent the values ​​at time 100000. ,frequency Power spectral density at;

[0076] For this frequency point The minimum energy value within the current time window;

[0077] For this frequency point The maximum energy value within the current time window;

[0078] The numerical stability constant (taken as 1 × 10⁻⁶) -8 ).

[0079] Using the above formula, the carrier frequency points with extremely high energy (whose time-domain fluctuations may be small) are normalized to the [0,1] interval; the spurious frequency points with weak energy (including fingerprint fluctuations) are also stretched to the [0,1] interval. This processing makes the weak fingerprint features numerically on the same order of magnitude as the strong carrier, thus explicitly "demasking".

[0080] Step S2.4: Feature stitching:

[0081] The I-path normalized features and Q-path normalized features at three scales (Scale 1, 10, 20) are concatenated along the channel dimension to form the final input tensor. .

[0082] The dimension is ,in For batch size, For frequency points, This represents the normalized time dimension.

[0083] Step 3: Construct a lightweight neural network model based on Bayesian Frequency Attention (BFA); the model uses ResNet18 as its backbone and embeds a BFA module at its front end;

[0084] As shown in Figures 3 and 4, the recognition network constructed in this embodiment mainly consists of an input layer, a Bayesian Frequency Attention (BFA) module, a ResNet18 backbone network, and a classification layer.

[0085] Step 3.1: Network Input:

[0086] The network input is the multi-scale normalized feature tensor generated in step S2. Its dimensions are ,in This refers to the batch size. Number of frequency points (in this embodiment) ), This represents the number of time frames.

[0087] Step 3.2: Construction of the Bayesian Frequency Attention (BFA) module:

[0088] To address the problem that existing deterministic attention mechanisms cannot quantify feature uncertainty, this embodiment designs a probabilistic BFA module. This module does not directly learn fixed weight values, but rather learns the probability distribution of the weights. The specific calculation process is as follows:

[0089] Step 3.2.1: Global Information Aggregation (Context Aggregation):

[0090] First, the input feature map Global Average Pooling (GAP) is performed along the time and channel dimensions to obtain a context vector containing global frequency distribution information. :

[0091]

[0092] Represents: The precise numerical value of the input time-frequency feature tensor at a specific location. Here, b represents the batch size, i.e., which sample in the batch is being processed; c represents the channel index, corresponding to the feature channels after multi-scale concatenation; f represents the frequency bin; and t represents the time frame index. It is a standard mathematical notation used to define the space and dimension of a vector; Represents the set of real numbers. This represents the total number of frequency points.

[0093] The vector Reflects the first The average energy intensity of each frequency point in the global range.

[0094] Step 3.2.2: Variational Inference:

[0095] This embodiment assumes frequency attention weight. The parameters follow a log-normal distribution to ensure that the generated attention weights are always positive. Two parallel multilayer perceptron (MLP) branches are constructed to fit the parameters of this distribution:

[0096] Prior Net: Used to learn the global prior distribution of frequency weights. This network consists of two fully connected layers (FC). The first layer divides the network from... Dimensional reduction (scaling ratio) The activation function is ReLU; the second layer restores the dimensions to [the specified value]. The output is a prior mean vector. .

[0097] Posterior Net: Used to infer the posterior distribution based on the current input samples. Its network structure is exactly the same as the prior net, and the input is a context vector. The output is the posterior mean vector. .

[0098] Note: To reduce model complexity and training difficulty, the variance of the distribution in this embodiment is... Set the parameters to be learnable or fixed hyperparameters (e.g., 0.1) and do not predict them through the network.

[0099] Step 3.2.3: Reparameterization Trick:

[0100] During the training phase, to ensure that the gradient can propagate backwards, this invention does not sample directly from the distribution, but instead introduces auxiliary noise variables. To achieve differentiability of the sampling process, the specific sampling formula is as follows:

[0101]

[0102] in, It is a random noise vector sampled from the standard normal distribution. The operation ensured the generated weights It is a positive value.

[0103] Step 3.2.4: Feature Recalibration:

[0104] Generate random weights Perform normalization (such as Softmax or Simplex constraints), then broaden (Broadcast) to match the input feature map. Same dimension, and perform element-wise multiplication:

[0105]

[0106] in, This refers to the enhanced feature tensor obtained after feature recalibration; The input is the original feature tensor; These are the normalized random attention weight coefficients generated by the Bayesian network for the f-th frequency point.

[0107] Through this step, the model can adaptively enhance key frequency bands (such as harmonic components) containing fingerprint information based on posterior probabilities, while suppressing noisy frequency bands.

[0108] Step S3.3: Backbone network feature extraction:

[0109] Enhanced feature map Input the ResNet18 backbone network. ResNet18 contains 4 residual blocks, each consisting of two... It consists of convolutional layers, batch normalization (BatchNorm) layers, and ReLU activation functions.

[0110] First convolutional layer: kernel size Step size 2, output channels 64.

[0111] Residual layer: The number of channels increases sequentially to 64, 128, 256, and 512.

[0112] Final output: After global average pooling, the output is a feature vector with a dimension of 512.

[0113] Step 4: Construct a loss function that includes the evidence lower bound (ELBO), and use the KL divergence annealing strategy to train and test the model.

[0114] To introduce Bayesian constraints while ensuring classification accuracy, this invention designs a composite loss function that includes an evidence lower bound (ELBO) and employs a KL divergence annealing strategy.

[0115] Step 4.1: Definition of loss function:

[0116] Total loss function It consists of two parts: classification loss and regularization loss.

[0117]

[0118] 1. Classification loss Cross-entropy loss is used to measure the predicted class probability. With real labels Differences between them:

[0119]

[0120] in Number of radiation source categories (in this embodiment) ).

[0121] 2. Regularization loss Kullback-Leibler (KL) divergence is used to constrain the posterior distribution. Approximating the prior distribution For a log-normal distribution, its KL divergence is calculated as follows:

[0122]

[0123] in, Indicates: the standard deviation of the prior distribution; Represents: the standard deviation of the posterior distribution; Represents: the posterior mean predicted by the posterior network for the f-th frequency point; It represents the prior mean predicted by the prior network for the f-th frequency point.

[0124] This loss condition forces the network to learn a posterior distribution that does not deviate too far from the prior distribution, thus playing a regularizing role and preventing overfitting.

[0125] Step 4.2: KL divergence annealing strategy:

[0126] To address the convergence problem of Bayesian neural networks, this invention proposes dynamically adjusting the weights. Strategy:

[0127] Warm-up phase (Epoch 0-20): Setup At this point, the network is dominated by classification loss and quickly learns discriminative features.

[0128] Annealing stage (Epoch 20-80): ... The value increases linearly from 0 to the target value (0.05 in this example). The formula is:

[0129]

[0130] This stage gradually introduces uncertainty constraints to regulate the distribution of the feature space.

[0131] Stable Phase (Epoch 80-100): Maintaining The parameters remain unchanged, but are fine-tuned.

[0132] Step 5: Specific experimental setup and parameters:

[0133] To verify the effectiveness of this invention, the specific implementation parameters are set as follows:

[0134] 1. Dataset generation:

[0135] Radiation source: 30 commercial communication radios (including the same model and different models).

[0136] Signal enhancement: Adds Gaussian white noise (AWGN) to the original signal, with a signal-to-noise ratio (SNR) range of 0dB to 20dB.

[0137] Sample size: 2,500 samples per class, for a total of 75,000 samples.

[0138] 2. Training hyperparameters:

[0139] Optimizer: Adam optimizer is used.

[0140] Learning Rate: A phased decay strategy is used, with an initial learning rate of 0.001, which decreases to 0.1 times the original value every 30 rounds, until... .

[0141] Batch Size: Set to 128.

[0142] Training epochs: 100 epochs.

[0143] Hardware configuration: Intel Core i9 CPU, NVIDIA RTX 4090 GPU, 64GB RAM.

[0144] Software framework: Python 3.9, PyTorch / PaddlePaddle.

[0145] Step 6: Setting Model Evaluation Metrics:

[0146] To comprehensively evaluate the performance of the specific radiation source identification method proposed in this invention, this embodiment selects accuracy, confusion matrix, t-SNE feature visualization, and computational complexity as evaluation criteria.

[0147] 1. Accuracy:

[0148] The calculation formula is:

[0149]

[0150] Among them, TP represents a true positive, TN represents a true negative, FP represents a false positive, and FN represents a false negative. This indicator intuitively reflects the model's overall ability to identify 30 types of radiation sources.

[0151] 2. Computational complexity metrics:

[0152] Floating-point operations (FLOPs): Measure the computational cost of a model, measured in gigabytes (Giga).

[0153] Parameters: Measures the space occupied by the model, measured in M ​​(Million).

[0154] Inference Time: Measures the time it takes for the model to process a single sample, measured in milliseconds (ms).

[0155] Step 7: Analysis of Recognition Results:

[0156] This embodiment conducted a system test on a dataset of 30 types of commercial radio station radiation sources and compared it with various existing feature extraction methods (such as STFT, EMD, VMD, DCD) and mainstream deep learning models (such as VGG16, ResNet50, ViT).

[0157] 1. Comparative analysis of existing feature extraction methods:

[0158] To verify the effectiveness of the MSFN-STFT (Multi-Scale Frequency Normalization) strategy proposed in this invention, it was compared with traditional feature extraction methods under the same ResNet18 backbone network. The results are shown in the table below:

[0159] Feature extraction methods Recognition accuracy (%) Preprocessing time (ms) STFT (Prior Art) 62.87 0.75 EMD (Empirical Mode Decomposition) 78.46 1748.13 VMD (Variational Mode Decomposition) 80.13 2040.01 DCD (Differential Constellation Diagram) 82.46 25.61 MSFN-STFT (This invention) 97.34 2.53

[0160] Results Analysis: Traditional STFT methods are limited by spectral masking effects, achieving an accuracy of only 62.87%. Although decomposition-based methods (EMD, VMD) can improve accuracy, their preprocessing time is too high (over 1.7 seconds), failing to meet real-time requirements. The MSFN-STFT method of this invention, while maintaining a low latency of milliseconds (2.53ms), significantly improves accuracy to 97.34%, demonstrating the significant advantage of normalization strategies in removing non-stationary carrier interference.

[0161] 2. Ablation experiment of Bayesian Frequency Attention (BFA) module:

[0162] To verify the role of the BFA module in uncertainty quantification and feature enhancement, ablation experiments were conducted:

[0163] Baseline model (ResNet18 + MSFN-STFT): Recognition accuracy is 97.34%.

[0164] The model of this invention (introducing the BFA module) improves the recognition accuracy to 98.49%.

[0165] Analysis of small sample scenarios: In situations with extremely limited data (only 100 training samples per class), the accuracy of the benchmark model drops significantly, while the model of this invention, benefiting from the regularization constraints introduced by Bayesian inference, achieves an accuracy improvement of approximately 20% compared to the benchmark model. This indicates that the BFA module effectively prevents the model from overfitting to a single noisy feature, significantly enhancing its generalization ability.

[0166] 3. Computational complexity and real-time performance evaluation:

[0167] The model of this invention is compared with mainstream deep learning models, and the results are as follows:

[0168] Model Architecture Number of parameters (M) FLOPs (G) Inference delay (ms) Accuracy (%) ResNet50 23.58 4.13 4.05 93.10 VGG16 134.39 15.47 10.86 88.54 ViT-B / 16 (Transformer) 85.07 17.29 13.10 (Non-convergent / Low) This invention (BFA-ResNet18) 11.59 2.32 2.87 98.49

[0169] Results Analysis: Compared to heavyweight networks such as VGG16 and ViT, the computational cost of the model in this invention is reduced by an order of magnitude (2.32 G FLOPs), with only 11.59 M parameters and a single inference time of only 2.87 ms. This demonstrates that this invention maintains state-of-the-art (SOTA) accuracy while possessing great potential for deployment on edge computing devices.

[0170] 4. Visual Analysis:

[0171] Confusion matrix analysis (see attached) Figure 5 The confusion matrix shows a clear diagonal, indicating that the vast majority of samples were correctly classified. False positives were mainly concentrated between different individuals of the same model (e.g., Radio 20 and Radio 23), which aligns with the physical laws governing hardware fingerprint similarity. False positives between different models were extremely rare, demonstrating the model's ability to distinguish between fine fingerprints.

[0172] t-SNE clustering analysis: such as Figure 6 As shown in the figure, the feature visualization results show that the features enhanced by BFA exhibit a clustered distribution with high cohesion and low coupling in two-dimensional space, and the inter-class boundaries are clear, further verifying the effectiveness of the method.

[0173] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A method for identifying specific radiation sources based on a Bayesian attention mechanism, characterized in that, Includes the following steps: Step 1: Obtain the baseband I / Q signal of the radiation source to be identified, perform multi-scale frequency normalized short-time Fourier transform on the baseband I / Q signal, and generate a multi-channel time-frequency feature map with carrier masking removed; Step 2: Construct a lightweight neural network model based on Bayesian frequency attention; the lightweight neural network model uses ResNet18 as the backbone network and embeds a Bayesian frequency attention module at its front end to perform frequency domain weighting on the input time-frequency feature map; Step 3: Construct a loss function that includes a lower bound on evidence, and train the model using the KL divergence annealing strategy; optimize the network parameters by learning the posterior distribution of the frequency weights through variational inference. Step 4: Input the signal to be tested into the trained neural network model, and determine the identity category of the radiation source through the probability distribution output by the model.

2. The method for identifying specific radiation sources based on Bayesian attention mechanism according to claim 1, characterized in that, The multi-scale frequency normalized short-time Fourier transform in step 1 specifically includes: first, performing a short-time Fourier transform on the baseband I / Q signal to obtain the original time-frequency matrix; then, setting multiple different time scale factors, performing time-domain energy normalization operations on each scale, linearly mapping the energy value of each frequency point on the time axis to a fixed interval to eliminate the absolute energy difference between different frequency points; finally, stitching the normalized spectrum diagrams generated by the I and Q paths at different scales along the channel dimension to form a multi-channel input tensor.

3. The method for identifying specific radiation sources based on Bayesian attention mechanism according to claim 1, characterized in that, The construction method of the Bayesian frequency attention module in step 2 includes: performing global average pooling on the input feature map to obtain a global frequency context vector; using prior and posterior networks to predict the distribution parameters of the frequency weights, assuming that the weights follow a log-normal distribution; using reparameterized sampling technology to generate random attention weights; and applying the generated random weights to the frequency dimension of the input feature map to obtain the enhanced feature map.

4. The method for identifying specific radiation sources based on Bayesian attention mechanism according to claim 3, characterized in that, The reparameterized sampling introduces a random noise vector sampled from a standard normal distribution, combines it with the mean vector output by the posterior network and a preset variance hyperparameter, and generates random attention weights that follow a log-normal distribution.

5. The method for identifying specific radiation sources based on Bayesian attention mechanism according to claim 1, characterized in that, The loss function in step 3 is composed of a weighted sum of cross-entropy classification loss and KL divergence regularization loss; where cross-entropy loss is used to ensure classification accuracy, and KL divergence is used to constrain the posterior distribution to approximate the prior distribution, playing a regularization role to prevent overfitting.

6. The method for identifying specific radiation sources based on Bayesian attention mechanism according to claim 5, characterized in that, In the KL divergence annealing strategy in step 3, the weight coefficient of KL divergence increases linearly with the training rounds: in the early stage of training, the weight coefficient is zero or very small, so that the model focuses on learning classification features. After the warm-up period, the weight coefficients are gradually increased to the target value, and Bayesian constraints are gradually introduced.

7. The method for identifying specific radiation sources based on Bayesian attention mechanism according to claim 1, characterized in that, The backbone network in step 2 contains four residual blocks. Each residual block consists of two convolutional layers, a batch normalization layer, and a ReLU activation function, ultimately outputting a feature vector with a dimension of 512.

Images