Intelligent Detection Method for Malicious Transmitters Based on Out-of-Distribution Data

Abstract

This invention presents an intelligent method for detecting malicious transmitters based on out-of-distribution data, comprising: acquiring radio frequency signals; constructing a deep residual shrinking network to perform feature extraction and signal denoising; determining whether the network training has converged, and if not, increasing the number of training iterations for further optimization; sampling from the out-of-distribution radio frequency signals to form subsets categorized by class; constructing and training a bi-headed bias network for each class, generating a bias score through a bias loss function; determining whether the network training has converged, and if not, increasing the number of training iterations for further optimization; integrating the deep residual shrinking network and the bi-headed bias network; inputting test set data into the trained network; and outputting the score result of malicious transmitter rejection and comparing it with a threshold to determine the detection result. This invention significantly enhances the malicious transmitter detection capability by introducing out-of-distribution samples, improves the accuracy of detecting unknown transmitters in open sets, and reduces the risk of malicious transmitters to IoT systems.

Description

Technical Field

[0001] This invention belongs to the field of Internet of Things security and wireless communication technology, specifically relating to an intelligent detection method for malicious transmitters based on out-of-distribution data. Background Technology

[0002] Transmitter identification is a core technology in the field of IoT security. It identifies devices by analyzing the unique characteristics of radio frequency signals (i.e., RF fingerprints) and is widely used in smart cities, industrial automation, and other scenarios. With the rise of 6G and satellite-drone networks, the scale of IoT devices is rapidly expanding, increasing the demand for spectrum resources. However, open wireless channels make systems vulnerable to malicious transmitter attacks, such as those that impersonate legitimate devices, threatening system reliability. Existing transmitter identification methods are mainly divided into two categories: methods based on traditional feature extraction and methods based on deep learning. Traditional methods rely on manually extracting signal features (such as amplitude and phase), but they have poor generalization ability in complex dynamic environments, are sensitive to noise, and require extensive manual tuning. For example, existing technologies (Qian Y, Qu J, Han X, et al., “Specific emitter identification based on multi-level sparse representation in automatic identification system,”) IEEE Transactions on Information Forensics and Security (vol. 16, pp. 3272-3282, 2021) proposed capturing signal patterns through multi-level sparse representation to improve robustness, but it is still affected by noise and is difficult to adapt to heterogeneous environments.

[0003] In recent years, deep learning-based methods have attracted attention due to their powerful feature extraction capabilities. Existing techniques (Ding L, Wang S, Wang F, et al., “Specific emitter identification via convolutional neural networks,”) IEEE Transactions on Information Forensics and Security(vol. 15, pp. 2589-2599, 2020) proposes using convolutional neural networks to extract features from raw signals to improve performance in dynamic environments. However, open-set transmitter identification requires detecting unknown malicious transmitters not included in the training data. Existing methods perform poorly under low signal-to-noise ratio conditions and do not fully utilize out-of-distribution radio frequency signals. For example, existing techniques (Wang Y, Gui G, Ohtsuka T, et al., “Scalable semi-supervised learning with semi-supervised kernel dictionary learning for specific emitter identification,”) IEEE Transactions on Aerospace and Electronic Systems The paper (vol. 56, no. 5, pp. 4012-4023, 2020) proposes a semi-supervised learning method to improve scalability, but its performance is limited in heterogeneous networks. Existing techniques (Xie X, Wang X, Gui G, “Transformer-based specific emitter identification in IoT environment,”) IEEE Internet of Things Journal (Vol. 8, no. 12, pp. 9876-9887, 2021) utilizes a converter to capture time-frequency dynamic features, but its performance degrades in complex noise environments. Patent application (application number CN202110456789.2, publication number CN113435374A) discloses a deep learning-based radio frequency fingerprinting method, which improves accuracy through multi-scale feature extraction, but does not fully utilize out-of-distribution radio frequency signals, resulting in limited detection capability for unknown transmitters.

[0004] These purely data-driven methods rely on a large number of training samples, which is difficult to implement in real-world, complex IoT systems. Furthermore, existing methods exhibit low detection accuracy under low signal-to-noise ratio (SNR) conditions, struggle to cope with heterogeneous network environments, and have limited effectiveness in rejecting unknown transmitters. Therefore, there is an urgent need to develop a novel open-set transmitter identification method to improve detection performance in low SNR environments, enhance generalization capabilities, and fully utilize distributed radio frequency signals to improve system robustness. Summary of the Invention

[0005] This invention addresses the shortcomings of existing transmitter identification technologies, such as poor detection performance in low signal-to-noise ratio environments, insufficient generalization ability in heterogeneous IoT environments, and insufficient utilization of distributed radio frequency signals. It proposes a malicious transmitter rejection system and method based on distributed data.

[0006] This invention discloses an intelligent detection method for malicious transmitters based on out-of-distribution data. The complete technical method includes the following steps:

[0007] (1) Collect radio frequency signals within and outside the distribution to construct test set data;

[0008] (2) Construct a network based on an open set transmitter identification framework. The network based on the open set transmitter identification framework integrates a deep residual shrinking network and a dual-head bias network. The deep residual shrinking network is used to extract features of in-distribution and out-of-distribution radio frequency signals to determine the type of transmitter. The dual-head bias network detects whether the radio frequency signal belongs to a malicious transmitter based on the features.

[0009] (3) Input the test set data into the network based on the open set transmitter recognition framework to obtain classification and detection results.

[0010] Furthermore, the deep residual shrinkage network is used to extract features of the in-distribution and out-of-distribution radio frequency signals specifically as follows:

[0011] Based on the distributed radio frequency signals in step (1), a deep residual shrinkage network is constructed and trained. The deep residual shrinkage network classifies the distributed radio frequency signals to obtain the transmitter category to which each type of distributed radio frequency signal belongs. It is then determined whether the training of the deep residual shrinkage network has ended. If yes, the trained deep residual shrinkage network is obtained. If not, the number of training iterations is incremented by one and the deep residual shrinkage network is trained again.

[0012] The trained deep residual shrinkage network extractor is used to extract features of both in-distribution and out-of-distribution radio frequency signals.

[0013] Furthermore, the deep residual shrinkage network utilizes cross-entropy loss to extract radio frequency fingerprint features and trains a classifier to classify the latent spatial features of transmitters corresponding to radio frequency signals within the distribution. The deep residual shrinkage network includes an initial convolutional layer, a convolutional block attention module, multiple channel adaptive residual shrinkage units, a global average pooling layer, and a fully connected layer.

[0014] The initial convolutional layer is used to extract preliminary temporal features, and the convolutional block attention module optimizes the extraction of temporal features through channel and spatial attention mechanisms to obtain key radio frequency fingerprint features;

[0015] Multiple channel adaptive residual shrinkage units employ a learnable soft thresholding mechanism to suppress noise and enhance the robustness of the key RF fingerprint features;

[0016] The global average pooling layer is used to average the output of different channels.

[0017] Furthermore, the dual-head bias network detects whether the radio frequency signal belongs to a malicious transmitter based on the features, specifically as follows:

[0018] Based on the aforementioned characteristics, the correspondence between in-distribution and out-of-distribution radio frequency signals is determined. For each type of in-distribution radio frequency signal, several out-of-distribution radio frequency signals are randomly selected to form a subset. A dual-head bias network is constructed and trained. The dual-head bias network outputs a bias score to identify malicious transmitters.

[0019] Determine whether the training of the deep residual shrinking network has ended. If yes, obtain the trained bi-head bias network. If not, increment the training iteration count by one and continue training the deep residual shrinking network.

[0020] Furthermore, the dual-headed bias network includes an explicit bias branch and an implicit bias branch. The explicit bias branch is used to capture the explicit bias features of the category, and the implicit bias branch captures the implicit bias characteristics of the category based on the residual. Through bias loss function optimization, a bias score is generated to distinguish between known and unknown transmitters, where the known transmitter is a malicious transmitter.

[0021] As a preferred embodiment of this application, the deviation loss function includes:

[0022]

[0023] in, It is a normality deviation. Indicates the bias network score;

[0024] Deviation loss is defined as follows:

[0025]

[0026] Among them, if If it is an out-of-distribution sample, then ,if If it is an in-distribution sample, then ;parameter This represents the confidence interval threshold, controlling the degree of deviation of out-of-distribution samples;

[0027] Calculate the average deviation score for the explicit deviation branch and the implicit deviation branch respectively.

[0028] Furthermore, step (3) specifically involves:

[0029] The network outputs a classification score and a bias score based on the open set transmitter identification framework. The bias score is the average of the explicit and implicit branch scores. For each category, a bias score is calculated, and the minimum value among all categories is taken as the final score. This score is then compared with a preset bias score threshold. If the score is below the threshold, the transmitter is determined to be an unknown transmitter, and its access is denied. Otherwise, it is identified as a known transmitter, and its category is output.

[0030]

[0031] in It is a judgment label. It is the deviation score. It was judged as the first The probability of each category, It is the total number of categories. It is the deviation score threshold.

[0032] Compared with the prior art, the present invention has the following advantages:

[0033] First, by introducing off-distribution data, the system significantly improves its sensitivity to detecting malicious transmitters in open scenarios and exhibits greater robustness in heterogeneous IoT environments.

[0034] Second, the deep residual shrinkage network works in conjunction with the convolutional block attention module to effectively denoise and extract robust RF fingerprints in low signal-to-noise ratio environments.

[0035] Third, the dual-headed bias network learns explicit and implicit anomalies separately based on bias loss and directly outputs anomaly scores, which significantly reduces the false alarm rate and outperforms mainstream malicious denial strategies in unknown transmitter detection tasks.

[0036] Fourth, the constructed end-to-end open set identification framework exhibits excellent robustness and generalization ability in heterogeneous IoT environments, and can be seamlessly applied to complex real-world scenarios such as smart cities and industrial IoT. Attached Figure Description

[0037] Figure 1 This is a flowchart of the present invention;

[0038] Figure 2 This is a schematic diagram of the deep residual shrinkage unit of the present invention;

[0039] Figure 3 This is a diagram of the dual-head deviation network framework of the present invention;

[0040] Figure 4 This is a general block diagram of the open set transmitter identification training and testing method using the present invention;

[0041] Figure 5This is a comparison chart of classification accuracy using the present invention and other existing technologies under different signal-to-noise ratio conditions;

[0042] Figure 6 This is a comparison chart of the area under the operating characteristic curve of subjects using the present invention and other existing technologies under different numbers of unknown transmitters;

[0043] Figure 7 This is a comparison chart of the area under the receiver operating characteristic curve (AUC) under different numbers of out-of-sample assisted training using the present invention and other existing technologies. Detailed Implementation

[0044] The invention will now be further described with reference to the accompanying drawings.

[0045] This invention introduces out-of-distribution data, combines a deep residual shrinking network and a dual-head bias network to construct an open-set transmitter identification framework. This significantly improves the detection capability of malicious transmitters in low signal-to-noise ratio environments, enhances generalization performance in heterogeneous IoT environments, and effectively utilizes out-of-distribution radio frequency signals to improve system robustness, thereby reducing the security threat posed by malicious transmitters to IoT systems. (See attached...) Figure 1 The specific steps of the method of the present invention are described below.

[0046] Example 1:

[0047] This embodiment presents an intelligent detection method for malicious transmitters based on out-of-distribution data. The complete technical method includes the following steps:

[0048] (1) Collect radio frequency signals within and outside the distribution to construct test set data;

[0049] (2) Construct a network based on an open set transmitter identification framework. The network based on the open set transmitter identification framework integrates a deep residual shrinking network and a dual-head bias network. The deep residual shrinking network is used to extract features of in-distribution and out-of-distribution radio frequency signals to determine the type of transmitter. The dual-head bias network detects whether the radio frequency signal belongs to a malicious transmitter based on the features.

[0050] (3) Input the test set data into the network based on the open set transmitter identification framework to obtain classification and detection results. The test set uses distributed radio frequency signals and a wider range of unauthorized transmitter signals as samples of distributed radio frequency signals for testing together.

[0051] Step 1, spectrum data acquisition.

[0052] Radio frequency signal acquisition: The signal transmitted by the transmitter is received by the receiver, and the received signal can be represented as... .in, The first part representing the transmitter signal One sampling point, Indicates the first received signal One sampling point, This indicates that the mean is 0 and the variance is . The received signal is represented as a vector composed of I / Q components using additive white Gaussian noise. In-distribution radio frequency (RF) signals include signals emitted by devices with authorized spectrum usage. These devices operate within a legal spectrum allocation framework, and their signal characteristics are within the regulatory scope of relevant authorities. RF signals from known transmitter categories (i.e., in-distribution RF signals) are collected, labeled with their category tags, and used to form a training dataset. In real-world scenarios, it is easy to obtain some unauthorized transmitter signal samples, which serve as known out-of-distribution RF signal samples. These signals are unauthorized but known. Simultaneously, out-of-distribution RF signals are collected as auxiliary data for subsequent training to enhance the detection capability of unknown transmitters. The data is stored as tensors in I / Q format. The IQ signals used are undemodulated RF signals containing multiple sampling points for feature extraction and classification.

[0053] Step 2: Construct and train a deep residual shrinkage network. Based on the I / Q data from Step 1, construct a deep residual shrinkage network, consisting of an initial convolutional layer, convolutional block attention modules, multiple channel adaptive residual shrinkage units, a global average pooling layer, and a fully connected layer. The initial convolutional layer extracts preliminary temporal features. The convolutional block attention modules optimize feature extraction through channel and spatial attention mechanisms, highlighting key RF fingerprint features. The channel adaptive residual shrinkage units employ a learnable soft thresholding mechanism to suppress noise and enhance feature robustness. The network outputs fixed-dimensional feature embeddings for subsequent classification. Training uses the Adam optimizer with a learning rate of 0.001 and a maximum training epoch of 150. The cross-entropy loss function is used to calculate the difference between the network output and the true label. During training, network parameters are randomly initialized, and I / Q data is input batch by batch to optimize the network parameters for robust feature extraction and denoising.

[0054] Step 3: Determine if the deep residual shrinkage network training has ended. Check if the network has reached the maximum training period (150 epochs) or if the loss is no longer decreasing (using an early stopping mechanism, no improvement after 10 epochs). If yes, freeze the network weights and proceed to Step 4; otherwise, increment the training iteration count and continue training the network.

[0055] Step 4: For each type of in-distribution radio frequency signal, a certain number of out-of-distribution radio frequency signals are randomly selected to form a subset. These subsets are used for training the subsequent dual-head bias network to enhance the detection capability of unknown transmitters.

[0056] Step 5: Construct and train a dual-headed bias network. Using the I / Q data from Step 1 and the out-of-distribution subset from Step 4, a dual-headed bias network is constructed, containing explicit and implicit bias branches. The explicit branch captures bias features of the class, while the implicit branch captures global bias characteristics. A bias score is generated through optimization using the bias loss function to distinguish between known and unknown transmitters. The dual-headed bias network includes explicit and implicit bias branches. The explicit bias branch is used to capture explicit bias features of the class, while the implicit bias branch captures implicit bias characteristics of the class based on residuals.

[0057] In this embodiment, using the I / Q data from step 1 and the out-of-distribution subset from step 4 as input, a bias score is generated through bias loss function optimization to distinguish between known and unknown transmitters; the bias loss function is as follows:

[0058]

[0059] in, It is a normality deviation. The bias network score is represented; the final bias loss is as follows:

[0060]

[0061] Among them, if If it is an out-of-distribution sample, then ,if If it is an in-distribution sample, then ;parameter The confidence interval threshold is used to control the degree of deviation of out-of-distribution samples; the average deviation score is calculated for the explicit bias branch and the implicit bias branch respectively.

[0062] The RMSprop optimizer was used for training with a learning rate of 0.001 and a maximum training epoch of 100. The mean squared error loss function was used to calculate the deviation between the network output and the out-of-distribution RF signal. During training, the network parameters were randomly initialized, data was input in batches, and the network was optimized to ensure convergence.

[0063] Step 6: Determine if the training of the bi-head biased network has ended. Check if the network has reached the maximum training period (100 epochs) or if the loss no longer decreases (using an early stopping mechanism, no improvement after 10 epochs). If yes, proceed to step 7; otherwise, increment the training iteration count and continue training the network.

[0064] Step 7: Integrate the Deep Residual Shrinkage Network and the B-headed Bias Network. Integrate the trained Deep Residual Shrinkage Network and the B-headed Bias Network to form an end-to-end open-set transmitter recognition framework. The Deep Residual Shrinkage Network is responsible for feature extraction and denoising, while the B-headed Bias Network generates classification and bias scores. Joint optimization is performed to achieve known transmitter classification and unknown transmitter rejection.

[0065] Step 8: Input test set data into the network. Input the test set data into the trained network. The test data includes signals of known categories and signals from unknown malicious transmitters. The format is the same as the training data, which are tensors in I / Q form.

[0066] Step 9: Output the detection results. The network outputs a classification score and a bias score, where the bias score is the average of the explicit and implicit branch scores. Calculate the bias score for each category, take the minimum of all categories as the final score, and compare it with a preset threshold. If the score is below the threshold, it is determined to be a malicious transmitter and its access is denied; otherwise, it is identified as a known transmitter and its category is output.

[0067] The effects of the present invention will be further explained below with reference to simulation experiments.

[0068] Simulation experiments and effect analysis:

[0069] 1. Simulation conditions and parameter settings.

[0070] The simulation experiments of this invention were conducted on Python 3.8 and PyTorch 2.0 platforms, with an NVIDIA 3090 GPU and an Intel Core i9 CPU. The experiments used the WiSig dataset, which contains WiFi signals collected by multiple commercial transmitters and receivers in the 2.4 GHz band at a sampling rate of 25 Msps, including raw I / Q samples. Two subsets were used: Subset 1 contained 6 transmitters and 12 receivers, with 1000 signals per pair, collected over 4 days, used to test the classifier's classification performance; Subset 2 contained 150 transmitters and 18 receivers, with 50 signals per pair, collected over 4 days. In Subset 2, 30 categories were selected as known transmitters, 30 categories as out-of-distribution samples, and another 30 categories as malicious transmitters. The training set comprised 70% of the dataset, the validation set 10%, and the test set 20%. The signal-to-noise ratio ranged from 0 dB to 20 dB with a step size of 2 dB.

[0071] 2. Simulation content.

[0072] Figure 5This paper presents a comparison of the classification accuracy of the method of this invention with other methods under different signal-to-noise ratio (SNR) conditions. The horizontal axis represents the SNR in dB, ranging from 0 dB to 20 dB in 2 dB increments; the vertical axis represents the classification accuracy. The line graph marked with dots represents the classification accuracy curve of the method of this invention; the line graph marked with squares represents the classification accuracy curve of the deep residual shrinkage network-independent channel method; the line graph marked with equilateral triangles represents the classification accuracy curve of the deep residual shrinkage network-shared channel method; and the line graph marked with diamonds represents the classification accuracy curve of the deep residual network method. Experimental results show that the method of this invention performs excellently at low SNR conditions. At an SNR of 0 dB, the classification accuracy of the method of this invention is approximately 71.2%, which is about 2.9% higher than the deep residual shrinkage network-independent channel thresholding method and about 9.1% higher than the deep residual network method. At a signal-to-noise ratio (SNR) of 4 dB, the classification accuracy of the method of this invention is approximately 83.3%, which is about 4.9% higher than the deep residual shrinkage network-independent channel method and about 9.8% higher than the deep residual network method. When the SNR reaches 20 dB, the classification accuracy of the method of this invention approaches saturation, at approximately 97.9%, which is significantly better than other methods, demonstrating its stable performance at high SNR.

[0073] Figure 6 This paper compares the performance of the method of this invention with other methods under different numbers of unknown transmitters, focusing on the area under the receiver operating characteristic (AUC) curve. The horizontal axis represents the number of unknown transmitters, ranging from 10 to 90, with intervals of 10; the vertical axis represents the AUC value. The performance curves of our network method are represented by dotted lines, the biased network method by square lines, the support vector data description method by equilateral triangle lines, and the distance-based method by diamond lines. Experimental results show that the deep fusion detection network method exhibits strong detection capabilities under different numbers of unknown transmitters. When the number of unknown transmitters is 10, the AUC of the deep fusion detection network method is approximately 0.952, which is about 6.9% higher than the support vector data description method and about 10% higher than the distance-based method. When the number of unknown transmitters is 90, the area under the curve of the deep hybrid detection network method is approximately 0.566, which is about 2.2% higher than the support vector data description method and about 2.4% higher than the distance-based method, demonstrating its robustness in scenarios with a high number of unknown transmitters.

[0074] Figure 7This paper presents a performance comparison of the area under the receiver operating characteristic (AUC) curve between the method of this invention and other methods under different numbers of out-of-distribution samples. The horizontal axis represents the number of out-of-distribution samples for each class, with values ​​of 1, 5, 10, 20, 50, and 100; the vertical axis represents the AUC value. The performance curves of the deep mixture detection network method are represented by dotted lines, the performance curves of the biased network method are represented by square lines, the performance curves of the improved support vector data description method are represented by equilateral triangle lines, the baseline value of the AUC for the support vector data description method is represented by a horizontal dashed line, and the baseline value of the AUC for the distance-based method is represented by a horizontal dotted line. Experimental results show that the deep mixture detection network method significantly improves performance with increasing out-of-distribution sample count. When the number of out-of-distribution samples is 100, the area under the curve of the bi-head bias network method is approximately 0.885, which is about 3.3% higher than the bias network method, about 10.1% higher than the support vector data description benchmark, and about 12.2% higher than the distance-based benchmark. This demonstrates the key role of out-of-distribution samples in enhancing malicious transmitter detection and the superiority of our proposed bi-head bias network over other techniques.

[0075] 3. Experimental conclusions.

[0076] Based on the simulation results and analysis above, the malicious transmitter intelligent detection method based on out-of-distribution data proposed in this invention significantly improves classification accuracy under low signal-to-noise ratio conditions, enhances the detection capability of unknown transmitters, reduces the false alarm rate, and exhibits excellent robustness in heterogeneous IoT environments. Compared with other current technologies, this invention has significant advantages in classification accuracy and detection capability of unknown transmitters under low signal-to-noise ratio conditions, making it suitable for practical scenarios such as smart cities and industrial IoT.

[0077] This invention has many specific applications. The above description is only a preferred embodiment of this invention. It should be noted that for those skilled in the art, several improvements can be made without departing from the principle of this invention, and these improvements should also be considered within the scope of protection of this invention.

Claims

1. A method for intelligent detection of malicious transmitters based on out-of-distribution data, characterized in that, The method includes the following steps: (1) collecting radio frequency signals within and outside the distribution to construct test set data; (2) Construct a network based on an open set transmitter identification framework. The network integrates a deep residual shrinking network and a dual-head bias network. The deep residual shrinking network is used to extract features of in-distribution and out-of-distribution radio frequency signals to determine the transmitter category. The dual-head bias network detects whether the radio frequency signal belongs to a malicious transmitter based on the features. Specifically, the dual-head bias network detects whether the radio frequency signal belongs to a malicious transmitter based on the features as follows: Based on the aforementioned characteristics, the correspondence between in-distribution and out-of-distribution radio frequency signals is determined. For each type of in-distribution radio frequency signal, several out-of-distribution radio frequency signals are randomly selected to form a subset. A dual-head bias network is constructed and trained. The dual-head bias network outputs a bias score to identify malicious transmitters. Determine whether the training of the deep residual shrinking network has ended. If yes, obtain the trained bi-head bias network. If no, increment the number of training iterations by one and continue training the deep residual shrinking network. The dual-head bias network comprises an explicit bias branch and an implicit bias branch. The explicit bias branch captures explicit bias characteristics of both in-distribution and out-of-distribution RF signal categories. The implicit bias branch captures implicit bias characteristics of both in-distribution and out-of-distribution RF signal categories based on residuals. A bias score is generated through bias loss function optimization to distinguish between known and unknown transmitters, where the unknown transmitter is a malicious transmitter. The bias loss function includes: ； in, It is a normality deviation. Indicates the bias network score. To represent the mean of the reference scores, Is with The associated standard deviation measures the range of the deviation scores. The deviation loss is as follows: ； Where z is the radio frequency signal, and y is the indicator bit for the radio frequency signal type, if If it is an external radio frequency signal sample, then ,if If it is an external radio frequency signal sample, then ;parameter This represents the confidence interval threshold, controlling the degree of deviation of out-of-distribution samples; Calculate the average deviation score for the explicit deviation branch and the implicit deviation branch respectively; (3) Input the test set data into the network based on the open set transmitter recognition framework to obtain classification and detection results.

2. The intelligent detection method for malicious transmitters based on out-of-distribution data according to claim 1, characterized in that, The deep residual shrinkage network is used to extract features of both in-distribution and out-of-distribution radio frequency signals, specifically as follows: Based on the distributed radio frequency signals in step (1), a deep residual shrinkage network is constructed and trained. The deep residual shrinkage network classifies the distributed radio frequency signals to obtain the transmitter category to which each type of distributed radio frequency signal belongs. It is then determined whether the training of the deep residual shrinkage network has ended. If yes, the trained deep residual shrinkage network is obtained. If not, the number of training iterations is incremented by one and the deep residual shrinkage network is trained again. The trained deep residual shrinkage network extractor is used to extract features of both in-distribution and out-of-distribution radio frequency signals.

3. The intelligent detection method for malicious transmitters based on out-of-distribution data according to claim 2, characterized in that, The deep residual shrinkage network uses cross-entropy loss to extract radio frequency fingerprint features and trains a classifier to classify the latent spatial features of the transmitter corresponding to the radio frequency signal within the distribution. The deep residual shrinkage network includes an initial convolutional layer, a convolutional block attention module, multiple channel adaptive residual shrinkage units, a global average pooling layer, and a fully connected layer. The initial convolutional layer is used to extract preliminary temporal features, and the convolutional block attention module optimizes the extraction of temporal features through channel and spatial attention mechanisms to obtain key radio frequency fingerprint features; Multiple channel adaptive residual shrinkage units employ a learnable soft thresholding mechanism to suppress noise and enhance the robustness of the key RF fingerprint features; The global average pooling layer is used to average the output of different channels.

4. The intelligent detection method for malicious transmitters based on out-of-distribution data according to claim 1, characterized in that, The specific steps (3) are as follows: The network based on the open set transmitter recognition framework outputs a classification score and a bias score. The bias score is the average of the explicit and implicit branch scores. For each category, the bias score is calculated, and the minimum value among all categories is taken as the final score, which is then compared with a preset bias score threshold. If the score is below the threshold, it is determined to be an unknown transmitter and its access is denied. Otherwise, identify it as a known transmitter and output its category: ； in It is a judgment label. It is the deviation score. It was judged as the first The probability of each category It is the total number of categories. It is the deviation score threshold.

Images