Radiation source individual feature analysis and secondary training system

Abstract

The invention discloses a radiation source individual feature analysis and secondary training system, and belongs to the field of electronic reconnaissance. The system comprises an online reconnaissance module, a signal reconnaissance equipment module, a data display module, a fingerprint feature analysis module, a database module, a secondary training data setting module, a secondary training parameter setting module, a secondary training module, a secondary training result visualization module and a model updating and pushing module. Through the multi-dimensional and multi-level signal feature extraction method, various individual features of the radiation source can be comprehensively captured, richer radiation source features are extracted, the recognition precision and robustness are improved, and the efficient recognition requirement in the complex electromagnetic environment is met. The system also supports a secondary learning function, and when a new radiation source individual appears, the system can carry out retraining on the basis of an original model through a secondary training module, so that the model can be optimized according to new data or environment change, and the long-term stability and accuracy of the system are ensured.

Description

Technical Field

[0001] The invention relates to a radiation source individual characteristic analysis and secondary training system, belonging to the field of electronic reconnaissance. Background Art

[0002] With the rapid development of wireless communication technology, radio frequency fingerprinting, as a highly effective means of communication security, has demonstrated tremendous potential in the identification and analysis of electromagnetic radiation sources. In practical applications, particularly for radio wave interference, wireless communication equipment security analysis, and spectrum management, the extraction and analysis of individual radiation source characteristics has become an important research topic.

[0003] However, existing radiation source individual identification systems lack two capabilities: more comprehensive analysis of radiation source individual characteristics and the ability to learn from them. Many existing systems rely primarily on simple time-domain, frequency-domain, or modulation features. While these features are effective in certain application scenarios, their ability to distinguish radiation source individual characteristics in complex electromagnetic environments is limited, especially when faced with new radiation source individuals or changing environmental factors. Furthermore, existing technologies often overlook more potential feature dimensions, such as statistical features and time-frequency features. This makes the analysis of radiation source individual characteristics less comprehensive in some cases, resulting in the system being unable to fully capture the complex characteristics of the radiation source. Furthermore, existing systems lack the ability to learn from them, which seriously affects the effectiveness of radiation source individual identification when faced with new radiation source individuals and when radiation source individual characteristics shift over time, thus failing to provide effective information for subsequent safety protection. Therefore, there is an urgent need for a radiation source individual characteristic analysis and secondary training system with more comprehensive feature analysis capabilities and secondary learning capabilities. Summary of the Invention

[0004] To address the existing problems in radiation source individual identification, which exist in most existing technologies, such as the lack of dynamic adaptation mechanisms and secondary learning in the extraction and analysis of radiation source individual features, the present invention aims to provide a radiation source individual feature analysis and secondary training system. Through a multi-dimensional and multi-level signal feature extraction method, it is possible to comprehensively capture various individual features of radiation sources, extract richer radiation source features, and improve the accuracy and robustness of identification. The present invention also supports secondary learning. When faced with newly emerging radiation source individuals, the system can retrain the original model through a secondary training module, allowing the model to be optimized based on new data or environmental changes, ensuring the long-term stability and accuracy of the system.

[0005] The purpose of the present invention is achieved through the following technical solutions:

[0006] The present invention discloses a radiation source individual feature analysis and secondary training system, which includes an online reconnaissance module, a signal reconnaissance equipment module, a data display module, a fingerprint feature analysis module, a database module, a secondary training data setting module, a secondary training parameter setting module, a secondary training module, a secondary training result visualization module and a model update push module.

[0007] After receiving the task assigned by the online reconnaissance module, the signal reconnaissance equipment module starts to monitor and collect the radiation source signal in real time, decodes and demodulates the radiation source signal, obtains the target signal, records and transmits the target signal to the data display module; the data display module is used to display the collected target signal waveform, spectrum diagram and analysis results of the fingerprint feature analysis module; the database module is used to store the data displayed by the data display module; the fingerprint feature analysis module extracts and analyzes the individual feature set of the radiation source from the target signal; the secondary training data setting module configures the data set for secondary training according to user needs and the data in the individual feature set, and sends it to the secondary training module; the secondary training parameter setting module configures the hyperparameters for secondary training according to user needs, and sends them to the secondary training module; after receiving the data in the secondary training data setting module and the hyperparameters in the secondary training parameter setting module, the secondary training module starts secondary training, and transmits the results of the secondary training process to the secondary training result visualization module for visualization to complete the secondary training.

[0008] Furthermore, after the secondary training module is completed, the model update push module encrypts the secondary trained model and pushes it to the signal reconnaissance device module; after receiving the model transmitted by the model update push module, the signal reconnaissance device module first performs model file integrity verification and decryption operations, and then replaces the original model in the signal reconnaissance device module.

[0009] Furthermore, the fingerprint feature analysis module extracts and analyzes the individual feature set of the radiation source from the target signal by:

[0010] S1: Construct a feature importance evaluation model to retain the feature combinations that contribute most to radiation source identification during feature selection;

[0011] Feature importance evaluation model:

[0012]

[0013] Where H(D) is the information entropy of data set D, D v is feature F i Take a subset of v values; H(D v ) is a symbol explanation;

[0014] By calculating the IG(F i) value, set a threshold to screen out important features and obtain a feature combination; further use a genetic algorithm to optimize the feature combination to obtain an optimized feature set F;

[0015] S2: Based on the t-SNE dimension reduction silhouette coefficient quantified feature separation, S1 is optimized twice.

[0016] The feature separation is expressed as:

[0017]

[0018] Where a(i) is the average distance of sample i to the same cluster, b(i) is the average distance to the nearest different cluster, and N is the total number of target signals;

[0019] Feature combination optimization model:

[0020]

[0021] Where F is the optimized feature set, C is the number of categories, and F k is the candidate feature set, that is, the set of individual features finally selected.

[0022] Furthermore, the secondary training parameter setting module uses the tree-structured Bayesian optimization algorithm to perform hyperparameter optimization as follows:

[0023] S1: Use hyperparameter optimization objective function to guide the tree-structured Bayesian optimization algorithm to optimize the selection of hyperparameters;

[0024] Hyperparameter optimization objective function:

[0025] min θ L(θ)=α·Loss val +β·t train (4)

[0026] Where θ represents the set of hyperparameters; Loss val is the loss function value on the validation set, t train is the time consumed by the training process, α and β are weight coefficients;

[0027] S2: Use Gaussian process regression to model the search space of hyperparameters;

[0028] Gaussian Process Regression Modeling:

[0029]

[0030] Where P(L|θ) is the distribution of the objective function L(θ) under the given hyperparameter θ, m(θ) is the mean function, and the kernel function k(θ,θ′) adopts Matern, which is expressed as follows:

[0031]

[0032] Where r = ‖θ-θ′‖ is the Euclidean distance between hyperparameters, θ and θ′ represent the currently searched hyperparameter and the hyperparameter that has been searched, respectively, and σ 2 is the variance of the signal, l is the length scale;

[0033] S3: Use the Expected Improvement acquisition function to guide the tree-based Bayesian optimization algorithm to find the best hyperparameters in the search space modeled in S2 to maximize the objective function of S1;

[0034] Expected Improvement collection function:

[0035] EI(θ)=Ε[max(f min -f(θ),0)] (7)

[0036] Among them, f min is the currently known minimum objective function value, and f(θ) is the predicted objective function value under the hyperparameter θ;

[0037] Furthermore, the secondary training module uses an incremental learning strategy to balance the learning ability of the secondary trained model for old data and new data to achieve secondary training. The method is as follows:

[0038] The old data refers to the data of the target signal collected previously and stored in the database module; the new data refers to the data of the target signal collected currently.

[0039] S1: An incremental learning loss function is used to measure the model's ability to grasp new and old data during the secondary training process;

[0040] Incremental learning loss function:

[0041] L total =λL new +(1-λ)Ε x～M [L old (x)] (8)

[0042] Among them, λ is the dynamic adjustment coefficient, L new is the cross entropy loss of new data, L old is the cross entropy loss of old data, E x～M Represents the expectation of sampling samples from the old dataset M;

[0043] Dynamic adjustment coefficient λ:

[0044]

[0045] Where γ is the attenuation coefficient, n new / n old is the ratio of the number of new and old data;

[0046] S2: The secondary training module initially has the old model model 旧 , use knowledge distillation to extract the model 旧 The ability to classify old data of the target signal and use the incremental learning loss function of S1 to guide the model 旧 Learn new data of the target signal to achieve secondary training on the new data of the target signal while retaining the classification ability of the old data, and finally obtain a new model model 新 ;

[0047] The individual features of the fingerprint feature analysis module include constellation diagram and signal modulation feature class, nonlinear feature class, statistical feature class, waveform feature class, envelope feature class, transform domain feature class, spectrum diagram, constellation diagram, TSNE diagram;

[0048] The encryption implementation method in the model update push module includes the following steps:

[0049] S1: Extract model 新 Parameter dictionary, which contains the model 新 All weight data of ;

[0050] S2: Model 新 The weight data of the model is saved as byte stream data. 新 The weights are converted into binary data that can be manipulated in memory;

[0051] S3: Perform Base64 encoding on the byte stream data, that is, convert the binary data into character data that can be safely transmitted;

[0052] S4: Initialize the AES encryptor using the key vector, where the AES encryptor uses the AES encryption algorithm and the CFB model;

[0053] S5: Use the initialized AES encryptor to encrypt the model 新 Encrypt the character data to generate an encrypted data stream and get the weight 密 file and calculate the weight 密 Get the MD5 of the file 收 .

[0054] The method for implementing the model file integrity check in the signal reconnaissance device module comprises the following steps:

[0055] S1: The signal reconnaissance device module receives the file pushed by the model update push module, including the encrypted weight 密 File and weight 密 MD5 of the file 收 ;

[0056] S2: Signal reconnaissance equipment module weight 密 Calculate the MD5 value of the file to get MD5 计 ;

[0057] S3: Comparison with MD5 收 and MD5 计 ;

[0058] S4: If they are consistent, it means the received weight 密 The file is complete; otherwise, the weight 密 The file is corrupted or tampered with.

[0059] The method for implementing the decryption operation in the signal reconnaissance device module comprises the following steps:

[0060] S1: Initialize the AES decryptor using the local key. The AES decryptor uses the AES decryption algorithm and the CFB model.

[0061] S2: Use the initialized AES decryptor to decrypt the weight 密 Decrypt the file and generate a decrypted data stream;

[0062] S3: Convert the decrypted data stream into byte stream data to obtain weight 明 document.

[0063] Beneficial effects:

[0064] 1. This invention discloses a system for analyzing and retraining individual radiation source characteristics. Through a multi-dimensional, multi-layered signal feature extraction method, the system can comprehensively capture various individual characteristics of radiation sources, including constellation diagrams, signal modulation features, and nonlinear features. Unlike traditional methods that rely solely on a limited number of feature dimensions, this system can extract a richer set of radiation source characteristics, improving recognition accuracy and robustness. Through comprehensive analysis of multi-dimensional features, the system can effectively distinguish different types of radiation sources, meeting the needs of efficient identification in complex electromagnetic environments.

[0065] 2. The present invention discloses a system for analyzing and retraining individual radiation source characteristics. This system supports a secondary learning function and can quickly adapt and perform secondary training when faced with newly emerging individual radiation sources. Through the secondary training module, the system can retrain the existing model, allowing the model to be optimized based on new data or environmental changes. Unlike traditional static learning methods, incremental learning can dynamically adjust the model's parameters, avoiding the performance degradation caused by environmental changes or new signal types in traditional methods, thereby ensuring the long-term stability and accuracy of the system.

[0066] 3. This invention discloses a system for analyzing and retraining individual characteristics of radiation sources. The retraining module automatically adapts to new data, updates the model in real time, and pushes it to the terminal device. The system's flexibility lies in the ability for users to customize the data and parameters for retraining, ensuring that each training session meets the specific requirements of the current environment. The training results visualization module provides an intuitive display of the training process, helping users evaluate and monitor training results, ensuring that each training session's optimization truly reflects improved system performance.

[0067] 4. This invention discloses a system for analyzing and retraining individual radiation source characteristics. After retraining, the optimized model is automatically pushed to actual reconnaissance equipment, ensuring that the equipment always uses the most appropriate version for the current environment. This real-time update mechanism enables the system to quickly respond to new threats or changes, enhancing its adaptability in dynamic environments, providing more accurate and stable identification of individual radiation sources, and further improving the system's practical application. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 This is a module connection diagram of a radiation source individual characteristic analysis and secondary training system of the present invention;

[0069] Figure 2 These are all the features contained in the fingerprint feature analysis module of the radiation source individual feature analysis and secondary training system of the present invention;

[0070] Figure 3 This is an encrypted workflow in the update push module of the radiation source individual feature analysis and secondary training system of the present invention;

[0071] Figure 4 This is a workflow for signal reconnaissance equipment module model file integrity verification of a radiation source individual feature analysis and secondary training system of the present invention;

[0072] Figure 5 This is the workflow for the decryption operation of the signal reconnaissance equipment module of the radiation source individual characteristic analysis and secondary training system of the present invention. DETAILED DESCRIPTION

[0073] In order to better illustrate the purpose and advantages of the present invention, the invention is further described below with reference to the accompanying drawings and examples.

[0074] Example 1:

[0075] like Figure 1 As shown, the present embodiment discloses a radiation source individual feature analysis and secondary training system, including an online reconnaissance module, a signal reconnaissance equipment module, a data display module, a fingerprint feature analysis module, a database module, a secondary training data setting module, a secondary training parameter setting module, a secondary training module, a secondary training result visualization module and a model update push module.

[0076] After receiving the task assigned by the online reconnaissance module, the signal reconnaissance equipment module starts to monitor and collect the radiation source signal in real time, decodes and demodulates the radiation source signal, obtains the target signal, records and transmits the target signal to the data display module; the data display module is used to display the collected target signal waveform, spectrum diagram and analysis results of the fingerprint feature analysis module; the database module is used to store the data displayed by the data display module; the fingerprint feature analysis module extracts and analyzes the individual feature set of the radiation source from the target signal; the secondary training data setting module configures the data set for secondary training according to user needs and the data in the individual feature set, and sends it to the secondary training module; the secondary training parameter setting module configures the hyperparameters for secondary training according to user needs, and sends it to the secondary training module; after receiving the data in the secondary training data setting module and the hyperparameters in the secondary training parameter setting module, the secondary training module starts secondary training and transmits the results of the secondary training process to the secondary training result visualization module for visualization; after the secondary training module ends, the model update push module encrypts the model and pushes it to the signal reconnaissance equipment module.

[0077] The online reconnaissance module is used to send real-time monitoring and data collection instructions for radiation source signals from the radio 900MHz-2.0GHz frequency band to the signal reconnaissance equipment module;

[0078] The signal reconnaissance equipment module includes hardware for signal acquisition and processing, such as RF front-ends and signal modulation and demodulation modules. Through precise signal decoding and processing, the signal reconnaissance equipment module can capture and record the characteristics of target signals and supports high-frequency signal detection. Through this module, the system can acquire target signals in complex electromagnetic environments, providing clear, interference-free raw data for subsequent data analysis and feature extraction.

[0079] The data display module is used to display the collected target signal waveform, spectrum diagram, and analysis results of the fingerprint feature analysis module, including signal visualization and recognition results. In addition, this module also provides data interaction functions. Users can modify analysis parameters and adjust the visual display content through the settings interface, so as to flexibly view data according to different needs;

[0080] The database module is used to store the data displayed by the data display module, including original signal data, extracted features, training models, and related log information. The database module uses the SQLite framework and establishes a data table to record the target signal data and extracted features, a model table to save model updates and storage location information, and a log table to record related log information. Through the database module, the system can efficiently access historical data and support long-term storage and fast retrieval of data. This module plays a key role in the system, ensuring the security, integrity, and traceability of the data. It also provides powerful data processing capabilities for the entire system and supports the storage and management of large amounts of signal data.

[0081] Fingerprint feature analysis module, which is responsible for extracting and analyzing the individual feature set of the radiation source from the target signal, including constellation diagram and signal modulation feature class, nonlinear feature class, statistical feature class, waveform feature class, envelope feature class, transform domain feature class, spectrum diagram, constellation diagram, TSNE diagram, and its details are as follows Figure 2 As shown. Among them, the constellation diagram and signal modulation feature class include constellation Figure 4Polygonal feature image, constellation cross angle feature, constellation gain imbalance, constellation phase imbalance, gain imbalance, phase imbalance, nonlinear feature class includes third-order nonlinear feature, second-order nonlinear feature image, statistical feature class includes information dimension feature, fisher kurtosis factor, skewness factor, waveform factor, kurtosis factor, pulse factor, segmented fisher kurtosis factor, segmented skewness factor, segmented waveform factor, segmented kurtosis factor, segmented pulse factor, high-order cumulant c20, high-order cumulant c21, high-order cumulant c40, high-order cumulant c41, high-order cumulant c42, high-order cumulant c60, high The first-order cumulant c63 and the high-order cumulant c80 are included. The waveform feature class includes the M-th power instantaneous phase extraction image, fractal dimension image, frequency deviation symmetry, carrier frequency, error vector magnitude, box dimension feature, shaping parameter, phase difference image, and symbol rate. The envelope feature class includes the envelope line image, envelope rising edge image, envelope leading edge high-order moment, envelope high-order kurtosis, envelope high-order skewness, and envelope variance. The transform domain feature class includes the average integral bispectrum image, axial integral bispectrum image, rectangular integral bispectrum, radial integral bispectrum image, circular integral bispectrum, time-frequency singular value image, wavelet transform image, and fractional Fourier transform (Frft) image. The method for extracting and analyzing the individual feature set of the radiation source from the target signal is as follows:

[0082] S1: Construct a feature importance evaluation model to retain the feature combinations that contribute most to radiation source identification during feature selection;

[0083] Feature importance evaluation model:

[0084]

[0085] Where H(D) is the information entropy of data set D, D v is feature F i Take a subset of v values; H(D v ) is the dataset D v Information entropy;

[0086] By calculating the IG(F i ) value, set the threshold to filter out important features and obtain the feature combination; further use the genetic algorithm to optimize the feature combination to obtain the optimized feature set F; this formula measures the feature F i The degree of uncertainty reduction of the classification information of dataset D. A higher information gain indicates that the feature contributes more to the recognition process.

[0087] In order to improve the efficiency of feature selection, a genetic algorithm is used to optimize the feature combination. The encoding method is represented by binary genetic coding; the fitness function uses the recognition accuracy of the feature combination; on this basis, the fitness function also adds a feature importance evaluation model IG (F i ) to ensure that the selected feature combination has high classification performance; the crossover probability that controls the frequency of genetic information exchange between parent genomes is set to 0.3 to increase the diversity of the population; the mutation probability that controls the probability of gene mutation is set to 0.1 to prevent the algorithm from falling into a local optimal solution; the termination condition of the genetic algorithm is that the maximum number of iterations reaches 100 or the change in the objective function value is less than the preset threshold value of 0.5 (that is, when the optimization stops, the performance of the model has stabilized). At this time, the obtained F includes constellation gain imbalance, constellation phase imbalance, gain imbalance, Fisher kurtosis factor, skewness factor, waveform factor, kurtosis factor, impulse factor, segmented Fisher kurtosis factor, segmented skewness factor, segmented waveform factor, segmented kurtosis factor, segmented impulse factor, high-order cumulant c20, high-order cumulant c21, and high-order cumulant c40;

[0088] S2: Based on the t-SNE dimensionality reduction silhouette coefficient to quantify the feature separation, S1 is optimized twice to improve the feature discrimination and recognition accuracy. By reducing the dimensionality of the target signal, t-SNE can effectively preserve the local structural information between samples, facilitating visual analysis in low-dimensional space.

[0089] The feature separation is expressed as:

[0090]

[0091] Where a(i) is the average distance from sample i to the same cluster, b(i) is the average distance to the nearest different cluster, and N is the total number of target signals. This indicator measures the separation between different clusters. The distinguishability of features is evaluated by calculating the distance relationship between samples. Based on the feature separation indicator, a threshold of 0.6 is set to determine whether the feature combination has sufficient distinguishing ability.

[0092] Feature combination optimization model:

[0093]

[0094] Where F is the optimized feature set, C is the number of categories, and F k is the candidate feature set, that is, the set of individual features finally selected. The optimization objective is to maximize the feature subset F kThe separation degree is 1, so that the distinguishing ability of each feature subset is optimized, thereby improving the recognition accuracy. At this time, the final selected individual feature set includes the plot gain imbalance, Fisher kurtosis factor, skewness factor, waveform factor, kurtosis factor, impulse factor, segmented skewness factor, segmented waveform factor, segmented kurtosis factor, segmented impulse factor, and high-order cumulative amount c20;

[0095] The secondary training data setup module is responsible for configuring the dataset for secondary training based on user needs and the data in the individual feature set. By adjusting the amount of data for different target signals, the model's learning ability is controlled so that the model can further improve its ability to identify new individual emitters of interest. This module can adjust and optimize the dataset according to different training requirements to ensure the quality and diversity of the training data.

[0096] The secondary training parameter setting module uses the tree-structured Bayesian optimization algorithm to optimize hyperparameters:

[0097] S1: Use hyperparameter optimization objective function to guide the tree-structured Bayesian optimization algorithm to optimize the selection of hyperparameters;

[0098] Hyperparameter optimization objective function:

[0099] min θ L(θ)=α·Loss val +β·t train (4)

[0100] Where θ represents the set of hyperparameters; Loss val is the loss function value on the validation set, t train is the time consumed by the training process, ɑ is 0.7, and β is 0.3. This objective function can effectively select the optimal hyperparameter configuration by minimizing the weighted sum of validation loss and training time, thereby improving the training performance and practical application effect of the model;

[0101] S2: In order to efficiently search for the optimal solution in the hyperparameter space, a Gaussian Process Regression (GPR) model is used to model the objective function. This model can predict the objective function value without an explicit expression and provide prediction uncertainty.

[0102] Gaussian Process Regression Modeling:

[0103]

[0104] Where P(L|θ) is the distribution of the objective function L(θ) under the given hyperparameter θ, m(θ) is the mean function, and the kernel function k(θ,θ′) uses Matern 5 / 2, which is expressed as follows:

[0105]

[0106] Where r = ‖θ-θ′‖ is the Euclidean distance between hyperparameters, θ and θ′ represent the currently searched hyperparameter and the hyperparameter that has been searched, respectively, and σ 2 is the variance of the signal, l is 5;

[0107] S3: Use the Expected Improvement acquisition function to guide the tree-based Bayesian optimization algorithm to find the best hyperparameters in the search space modeled in S2 to maximize the objective function of S1;

[0108] Expected Improvement collection function:

[0109] EI(θ)=Ε[max(f min -f(θ),0)] (7)

[0110] Among them, f min is the currently known minimum objective function value, and f(θ) is the predicted objective function value under the hyperparameter θ;

[0111] The secondary training module uses an incremental learning strategy to balance the new model's learning ability for old data (referring to the previously collected target signal data stored in the database module) and new data (referring to the currently collected target signal data) to achieve secondary training. The method is as follows:

[0112] S1: The basic network architecture used in the secondary training module is AAResNeXt, and its specific architecture is shown in the figure below:

[0113]

[0114] S2: An incremental learning loss function is used to measure the model's ability to grasp new and old data during the secondary training process;

[0115] Incremental learning loss function:

[0116] L total =λL new +(1-λ)Ε x～M [L old (x)] (8)

[0117] Among them, λ is the dynamic adjustment coefficient, L new is the cross entropy loss of new data, L oldis the cross entropy loss of old data, E x～M Represents the expectation of sampling samples from the old dataset M;

[0118] Dynamic adjustment coefficient λ:

[0119]

[0120] Where γ is 0.5, n new / n old is the ratio of the number of new and old data;

[0121] S3: The secondary training module initially has the old model model 旧 , use knowledge distillation to extract the model 旧 The ability to classify old data of the target signal and use the incremental learning loss function of S1 to guide the model 旧 Learn new data of the target signal to achieve secondary training on the new data of the target signal while retaining the classification ability of the old data, and finally obtain a new model model 新 ;

[0122] The secondary training results visualization module is responsible for displaying the results of the secondary training module, including metrics such as training loss and accuracy, as well as the model's learning progress. Real-time visualization charts allow users to intuitively see the training results, evaluate model performance, and make corresponding adjustments. This module also supports tracking training progress, helping users understand the model's performance in each round of training and ensuring a transparent and controllable training process.

[0123] The model update push module is responsible for updating the model after the secondary training module 新 The model is pushed to the signal reconnaissance equipment module. Through this module, the latest model 新 The model is encrypted and uploaded to the signal reconnaissance device module to improve performance in real-world applications. The model update push module ensures continuous model updates and upgrades, supports automated push, and ensures that the signal reconnaissance device module uses the latest and most accurate recognition model, providing stronger support for subsequent security protection.

[0124] The encrypted workflow in the model update push module is as follows Figure 3 As shown, specifically including:

[0125] S1: Extract model 新 Parameter dictionary, which contains the model 新 All weight data of ;

[0126] S2: Model 新 The weight data of the model is saved as byte stream data.新 The weights are converted into binary data that can be manipulated in memory;

[0127] S3: Perform Base64 encoding on the byte stream data, that is, convert the binary data into character data that can be safely transmitted;

[0128] S4: Initialize the AES encryptor using the key vector, where the AES encryptor uses the AES encryption algorithm and the CFB model;

[0129] S5: Use the initialized AES encryptor to encrypt the model 新 Encrypt the character data to generate an encrypted data stream and get the weight 密 file and calculate the weight 密 Get the MD5 of the file 收 .

[0130] The workflow of signal reconnaissance equipment module model file integrity check is as follows: Figure 4 As shown, specifically including:

[0131] S1: The signal reconnaissance device module receives the file pushed by the model update push module, including the encrypted weight 密 File and weight 密 MD5 of the file 收 ;

[0132] S2: Signal reconnaissance equipment module weight 密 Calculate the MD5 value of the file to get MD5 计 ;

[0133] S3: Comparison with MD5 收 and MD5 计 ;

[0134] S4: If they are consistent, it means the received weight 密 The file is complete; otherwise, the weight 密 The file is corrupted or tampered with.

[0135] The workflow of the signal reconnaissance equipment module decryption operation is as follows: Figure 5 As shown, specifically including:

[0136] S1: Initialize the AES decryptor using the local key. The AES decryptor uses the AES decryption algorithm and the CFB model.

[0137] S2: Use the initialized AES decryptor to decrypt the weight 密 Decrypt the file and generate a decrypted data stream;

[0138] S3: Convert the decrypted data stream into byte stream data to obtain weight 明 document.

[0139] The above specific description further illustrates the purpose, technical solutions and beneficial effects of the invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A radiation source individual feature analysis and secondary training system, characterized by: It includes online reconnaissance module, signal reconnaissance equipment module, data display module, fingerprint feature analysis module, database module, secondary training data setting module, secondary training parameter setting module, secondary training module, secondary training result visualization module and model update push module; After receiving the task from the online reconnaissance module, the signal reconnaissance equipment module starts to monitor and collect the radiation source signal in real time, decodes and demodulates the radiation source signal, obtains the target signal, records it, and transmits the target signal to the data display module; The data display module is used to display the collected target signal waveform, spectrum diagram and analysis results of the fingerprint feature analysis module; The database module is used to store the data displayed by the data display module; The fingerprint feature analysis module extracts and analyzes the individual feature set of the radiation source from the target signal; the secondary training data setting module configures the data set for secondary training based on user requirements and the data in the individual feature set, and sends it to the secondary training module; the secondary training parameter setting module configures the hyperparameters for secondary training based on user requirements and sends them to the secondary training module; After receiving the data in the secondary training data setting module and the hyperparameters in the secondary training parameter setting module, the secondary training module starts secondary training and transmits the results of the secondary training process to the secondary training result visualization module for visualization to complete the secondary training.

2. The radiation source individual characteristic analysis and secondary training system according to claim 1, characterized in that: After the secondary training module is completed, the model update push module encrypts the secondary trained model and pushes it to the signal reconnaissance device module; after receiving the model transmitted by the model update push module, the signal reconnaissance device module first performs model file integrity verification and decryption operations, and then replaces the original model in the signal reconnaissance device module.

3. The radiation source individual characteristic analysis and secondary training system according to claim 1, characterized in that: The fingerprint feature analysis module extracts and analyzes the individual feature set of the radiation source from the target signal, including the following steps: S1: Construct a feature importance evaluation model to retain the feature combinations that contribute most to radiation source identification during feature selection; Feature importance evaluation model: Where H(D) is the information entropy of the data set D, D v is feature F i Take a subset of v values; H(D v ) is a symbol explanation; By calculating the IG(F i ) value, set a threshold to screen out important features and obtain a feature combination; further use a genetic algorithm to optimize the feature combination to obtain an optimized feature set F; S2: Based on the t-SNE dimension reduction silhouette coefficient quantified feature separation, S1 is optimized twice. The feature separation is expressed as: Where a(i) is the average distance of sample i to the same cluster, b(i) is the average distance to the nearest different cluster, and N is the total number of target signals; Feature combination optimization model: Where F is the optimized feature set, C is the number of categories, and F k is the candidate feature set, that is, the set of individual features finally selected.

4. The radiation source individual characteristic analysis and secondary training system according to claim 1, characterized in that: Furthermore, the secondary training parameter setting module uses a tree-structured Bayesian optimization algorithm to perform a hyperparameter optimization method, including the following steps: S1: Use hyperparameter optimization objective function to guide the tree-structured Bayesian optimization algorithm to optimize the selection of hyperparameters; Hyperparameter optimization objective function: minutes θ L(θ)=α·Loss val +β·t train (4) Where θ represents the set of hyperparameters; Loss val is the loss function value on the validation set, t train is the time consumed by the training process, α and β are weight coefficients; S2: Use Gaussian process regression to model the search space of hyperparameters; Gaussian Process Regression Modeling: Where P(L|θ) is the distribution of the objective function L(θ) under the given hyperparameter θ, m(θ) is the mean function, and the kernel function k(θ, θ′) adopts Matern, which is expressed as follows: Where r = ||θ-θ′|| is the Euclidean distance between hyperparameters, θ and θ′ represent the currently searched hyperparameter and the previously searched hyperparameter, respectively, and σ 2 is the variance of the signal, l is the length scale; S3: Use the Expected Improvement acquisition function to guide the tree-based Bayesian optimization algorithm to find the best hyperparameters in the search space modeled in S2 to maximize the objective function of S1; Expected Improvement collection function: EI(θ)=E[max(f min -f(θ),0)] (7) Among them, f min is the currently known minimum objective function value, and f(θ) is the predicted objective function value under the hyperparameter θ.

5. The radiation source individual characteristic analysis and secondary training system according to claim 1, characterized in that: The secondary training module uses an incremental learning strategy to balance the learning ability of the secondary trained model on old and new data to achieve secondary training. The method is as follows: S1: An incremental learning loss function is used to measure the model's ability to grasp new and old data during the secondary training process; Incremental learning loss function: Among them, λ is the dynamic adjustment coefficient, L new is the cross entropy loss of new data, L old is the cross entropy loss of old data, Represents the expectation of sampling samples from the old dataset M; Dynamic adjustment coefficient λ: Where γ is the attenuation coefficient, n new / n old is the ratio of the number of new and old data; S2: The secondary training module initially has the old model Extraction using knowledge distillation The old data classification ability of the target signal is guided by the incremental learning loss function of S1 Learn new data of the target signal to achieve secondary training on the new data of the target signal while retaining the classification ability of the old data, and finally obtain a new model model 新 ; The old data refers to the data of the target signal collected previously and stored in the database module; the new data refers to the data of the target signal collected currently.

6. The radiation source individual characteristic analysis and secondary training system according to claim 2, characterized in that: The encryption implementation method in the model update push module includes the following steps: S1: Extract model 新 Parameter dictionary, which contains the model 新 All weight data of ; S2: Model 新 The weight data of the model is saved as byte stream data. 新 The weights are converted into binary data that can be manipulated in memory; S3: Perform Base64 encoding on the byte stream data, that is, convert the binary data into character data that can be safely transmitted; S4: Initialize the AES encryptor using the key vector, where the AES encryptor uses the AES encryption algorithm and the CFB model; S5: Use the initialized AES encryptor to encrypt the model 新 Encrypt the character data to generate an encrypted data stream and get the weight 密 File, and calculate weight 密 Get the MD5 of the file 收 .

7. The radiation source individual characteristic analysis and secondary training system according to claim 3, characterized in that: The individual features of the fingerprint feature analysis module include constellation diagram and signal modulation feature class, nonlinear feature class, statistical feature class, waveform feature class, envelope feature class, transform domain feature class, spectrum diagram, constellation diagram, and TSNE diagram.

8. The radiation source individual characteristic analysis and secondary training system according to claim 4, characterized in that: The method for implementing the model file integrity check in the signal reconnaissance device module comprises the following steps: S1: The signal reconnaissance device module receives the file pushed by the model update push module, including the encrypted weight 密 File and weight 密 MD5 of the file 收 ; S2: Signal reconnaissance equipment module weight 密 Calculate the MD5 value of the file to get MD5 计 ; S3: Comparison with MD5 收 and MD5 计 ; S4: If they are consistent, it means the received weighht 密 The file is complete; otherwise, the weight 密 The file is corrupted or tampered with.

9. The radiation source individual characteristic analysis and secondary training system according to claim 4, characterized in that: The method for implementing the decryption operation in the signal reconnaissance device module comprises the following steps: S1: Initialize the AES decryptor using the local key. The AES decryptor uses the AES decryption algorithm and the CFB model. S2: Use the initialized AES decryptor to decrypt weight 密 Decrypt the file and generate a decrypted data stream; S3: Convert the decrypted data stream into byte stream data to obtain weight 明 document.

Images