A method and model for individual identification of multi-source heterogeneous radiation sources based on adaptive fingerprint feature fusion

Abstract

The present invention discloses a method for identifying individual multi-source heterogeneous radiation sources based on adaptive fingerprint feature fusion, comprising the following steps: Step 1, using USRP to collect radio signals of different modes at different positions, including original radio frequency signals, signals after OFDM cyclic code decoding, and signals after channel equalization; Step 2, using three parallel branches of one-dimensional or two-dimensional convolutional layers, pooling layers, and fully connected layers to extract corresponding radio frequency fingerprints for three different signals respectively; Step 3, using a feature fusion module based on an attention mechanism to fuse the three radio frequency fingerprints extracted in the previous step; Step 4, using a fully connected network classifier to classify the fused radio frequency fingerprint features. The method disclosed by the present invention takes into account the data from a variety of different sensors in real scenarios, and overcomes the problem of low utilization of multi-source heterogeneous data.

Description

Technical Field

[0001] The present invention belongs to the field of radiation source individual identification, multi-source heterogeneous data processing and feature fusion, and particularly relates to a multi-source heterogeneous radiation source individual identification method and model based on adaptive fingerprint feature fusion in this field. Background Art

[0002] Radiation source individual identification is a technology that quickly identifies the identity of the source of the received signal. This is an important part of ensuring the physical layer security in the industrial Internet of Things, such as hacker intrusion detection, privacy data leakage, illegal device intervention, etc. It plays an important role in ensuring the security of the industrial Internet of Things. Traditional radiation source individual identification methods mainly include likelihood-based methods and feature-based methods. However, likelihood-based methods are difficult to obtain the position parameters of channels and noise, and feature-based methods are easily affected by noise, resulting in performance degradation. In recent years, deep learning methods have achieved great success in various pattern classification tasks, such as visual detection, visual tracking, and natural language processing. Many radiation source individual identification methods based on deep learning have been proposed one after another. However, due to the large number of parameters and hyperparameters in the model, the radiation source individual identification method based on deep learning has an inherent large amount of calculation, and generally has problems such as poor utilization of multi-source heterogeneous sensor data and high dependence on high-quality data.

[0003] In the field of radiation source individual identification, there has never been an attempt to introduce multi-source heterogeneous data. However, in real-world application scenarios, especially in industrial Internet of Things, digital twin smart parks, smart transportation and other scenarios, the perception layer contains a large number of sensors in different regions and of different types, and multi-source heterogeneous data is more common and commonly used. However, the current methods for radiation source individual identification often only select a single type of data from a single receiver for identification and authentication, resulting in low utilization of multi-source heterogeneous data and a high degree of dependence on the working stability of the receiver. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a method and model for individual identification of multi-source heterogeneous radiation sources based on adaptive fingerprint feature fusion, so as to solve the problem that the existing method for individual identification of radiation sources relies on single sensor data, resulting in low utilization rate of multi-source heterogeneous data and high dependence on the stability of the working condition of a single sensor.

[0005] The present invention adopts the following technical solution:

[0006] A method for identifying individual multi-source heterogeneous radiation sources based on adaptive fingerprint feature fusion, the improvement of which is that it includes the following steps:

[0007] Step 1: Use USRP to collect radio signals of different modes at different locations, including original RF signals, signals after OFDM cyclic code decoding, and signals after channel equalization;

[0008] Step 2, using three parallel branches of one-dimensional or two-dimensional convolutional layers, pooling layers, and fully connected layers to extract corresponding RF fingerprints for the three different signals respectively;

[0009] Step 3: Use the feature fusion module based on the attention mechanism to fuse the three RF fingerprints extracted in the previous step;

[0010] The general expression for defining the RF fingerprint feature fusion module is as follows:

[0011]

[0012] In the above formula, Γ mix Represents the fused RF fingerprint feature, Indicates the multiplication of corresponding elements in the matrix, ω ai Represents the radio frequency fingerprint feature Γ Ri and Γ R(i+1) The attention weight calculated after the attention module;

[0013] Attention weight ω ai The calculation formula is defined as follows:

[0014]

[0015] In the above formula, σ[·] represents the activation function, ∩(·) represents the local attention operation, which consists of a convolutional layer and a batch normalization layer; ∪(·) represents the global attention operation. The two attention mechanism operations ∩(·) and ∪(·) are defined as follows:

[0016]

[0017] In the above formula, represents the convolutional layer, represents a batch normalization layer, represents the global pooling layer, ω * represents the weight corresponding to the convolutional layer, δ r represents the activation function;

[0018] Step 4: Use a fully connected network classifier to classify the fused RF fingerprint features;

[0019] Define a classifier consisting of fully connected layers as follows:

[0020]

[0021] In the above formula, represents the label output by the classifier, represents three fully connected layers, ω F represents the weight parameter corresponding to the fully connected layer. The input of the classifier is the output of step 3, which is the RF fingerprint feature after the fusion of each branch.

[0022] Step 5: Select the network optimizer as the adaptive moment estimation optimizer Adam, build the model, and start network training:

[0023] Define the network input: multi-source heterogeneous RF signal: y k-R1 ,y k-R2 ,y k-R3 ; Corresponding device label: C k ; Initial learning rate: η; Maximum number of training times: τ;

[0024] Define the output of the network: Identified device labels:

[0025] Import training data: C k ;

[0026] Initialize network model weights: ω *i ,ω Fi ;

[0027] Training process:

[0028] For each epoch:

[0029] Multi-channel parallel extraction of radio frequency fingerprint features of multi-source heterogeneous data:

[0030]

[0031] In the above formula, represents the first branch pooling layer, represents the second branch pooling layer, Represents the third branch pooling layer;

[0032] RF fingerprint feature fusion based on attention mechanism:

[0033]

[0034] Call the classifier based on the fully connected layer to perform device identification on the fused RF fingerprint:

[0035]

[0036] Calculate the loss function loss t :

[0037]

[0038] Determine whether the loss function has a convergence trend. If the loss function does not converge in the convergence direction after 20 epochs, dynamically adjust the algorithm learning rate:

[0039] η=η×0.7

[0040] Otherwise, use the Adam optimizer to dynamically update the network model weights:

[0041]

[0042] At this point, a complete network model training is completed;

[0043] Step 6: Retrain and obtain the final model:

[0044] According to the results of multiple iterative training cycles, the final radiation source individual identification model is obtained.

[0045] Furthermore, in step 3, the activation function σ[·] is selected to be a relu function or a sigmoid function.

[0046] Furthermore, in step 5, the learning rate is set to 0.0001.

[0047] A multi-source heterogeneous radiation source individual recognition model based on adaptive fingerprint feature fusion, the improvement of which lies in: the model is implemented by the above method steps.

[0048] The beneficial effects of the present invention are:

[0049] The method disclosed in the present invention takes into account the data from a variety of different sensors in real scenarios, overcomes the problem of low utilization of multi-source heterogeneous data, and realizes efficient and reliable individual identification of radiation sources through an adaptive fingerprint feature fusion module. It can also work stably in harsh environments, providing an effective solution to the physical layer security issues in the industrial Internet of Things.

[0050] The method disclosed in the present invention can extract the corresponding radio frequency fingerprint features of multi-source heterogeneous data respectively through the multi-branch parallel feature extraction module in step 2; the adaptive fingerprint feature fusion module based on the attention mechanism (Attention) designed in step 3 obtains the optimal weights of different radio frequency fingerprints, realizes the optimal fingerprint fusion, and solves the problem that traditional machine learning and deep learning cannot efficiently utilize multi-source heterogeneous data; in addition, due to the introduction of multi-source heterogeneous data, this method is more reliable than the traditional radiation source individual identification method, and has obvious stability advantages in harsh working environments.

[0051] The model disclosed in the present invention designs a multi-channel parallel feature extraction module to extract the RF fingerprints of multi-source heterogeneous electromagnetic signal data respectively, and then fuses the signal fingerprints of multiple branches and multiple receivers through an adaptive fingerprint fusion module based on the attention mechanism, and finally uses a simple classifier to achieve fast and accurate device identification and authentication. At the same time, it has good stability, based on multi-source heterogeneous data, and gets rid of the dependence on the good working conditions of a single signal receiver, which has great practical significance in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a schematic flow diagram of the method of the present invention;

[0053] Figure 2 It is a schematic diagram of the system architecture of the model of the present invention;

[0054] Figure 3 It is a schematic diagram of the system performance of the model of the present invention when the sensor is damaged. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0056] Embodiment 1, this embodiment discloses a method for identifying individual multi-source heterogeneous radiation sources based on adaptive fingerprint feature fusion, such as Figure 1 As shown, the following steps are included:

[0057] Step 1: Use USRP (Universal Software Radio Peripheral) to collect radio signals of different modes at different locations, including original RF signals, signals after OFDM (Orthogonal Frequency Division Multiplexing) cyclic coding, and signals after channel equalization.

[0058] Select three USRP X310s and place them in different areas. Set them to different receiving modes: receiving the original RF signal, receiving the RF signal after OFDM cyclic prefix code decoding, and receiving the RF signal after channel equalization. The formats and corresponding sizes of the received signals of different receivers are shown in the following table:

[0059] Receiver type Data Format Symbols Receiver 1 Original RF signal Real number: 30720*2 <![CDATA[y R1 ]]> Receiver 2 Demodulate OFDM cyclic signal Real number: 14*2048*2 <![CDATA[y R2 ]]> Receiver 3 Signal after channel equalization Real number: 12*1200*2 <![CDATA[y R3 ]]>

[0060] Step 2, using three parallel branches of one-dimensional or two-dimensional convolutional layers, pooling layers, and fully connected layers to extract corresponding RF fingerprints for the three different signals respectively;

[0061] Design branch 1 to extract the original RF signal y collected by receiver 1 R1 The corresponding radio frequency fingerprint feature. This branch contains four one-dimensional convolutional layers, where the corresponding parameters are the number of channels filter = [16, 64, 128, 256], the convolution kernel size kernel-size = [5, 7, 5, 7]; each convolutional layer is followed by a one-dimensional maximum pooling layer with the parameter kernel-size = [4]; the fourth convolutional layer is followed by a fully connected layer with the parameter of the number of channels filter =

[256] ; that is, this branch outputs the radio frequency fingerprint feature

[0062]

[0063] Design branch 2 to extract the original RF signal y collected by receiver 2 R2 The corresponding radio frequency fingerprint feature. This branch contains four two-dimensional convolutional layers, where the corresponding parameters are the number of channels filter = [16, 64, 128, 256], the convolution kernel size kernel-size = [5, 7, 5, 7]; each convolutional layer is followed by a two-dimensional maximum pooling layer with the parameter kernel-size = [(1, 2)]; the fourth convolutional layer is followed by a fully connected layer with the parameter of the number of channels filter =

[256] ; that is, this branch outputs the radio frequency fingerprint feature

[0064]

[0065] Design branch 3 to extract the original RF signal y collected by receiver 3 R3 The corresponding radio frequency fingerprint feature. This branch contains four two-dimensional convolutional layers, where the corresponding parameters are the number of channels filter = [16, 64, 128, 256], the convolution kernel size kernel-size = [5, 7, 5, 7]; each convolutional layer is followed by a two-dimensional maximum pooling layer with the parameter kernel-size = [(1, 2)]; the fourth convolutional layer is followed by a fully connected layer with the parameter of the number of channels filter =

[256] ; that is, this branch outputs the radio frequency fingerprint feature

[0066]

[0067] Step 3: Use the feature fusion module based on the attention mechanism to fuse the three RF fingerprints extracted in the previous step;

[0068] The general expression for defining the RF fingerprint feature fusion module is as follows:

[0069]

[0070] In the above formula, Γ mix Represents the fused RF fingerprint feature, Indicates the multiplication of corresponding elements in the matrix, ω ai Represents the radio frequency fingerprint feature Γ Ri and Γ R(i+1) The attention weight calculated after the attention module;

[0071] It is important to specify that the attention weight ω ai The calculation formula of , that is, the attention mechanism designed by the present invention is defined as follows:

[0072]

[0073] In the above formula, σ[·] represents the activation function, which can be either the relu function or the sigmoid function. ∩(·) represents the local attention operation, which consists of a convolutional layer and a batch normalization layer; ∪(·) represents the global attention operation, which has an additional global pooling layer compared to ∩(·). The two attention mechanism operations of ∩(·) and ∪(·) are defined as follows:

[0074]

[0075] In the above formula, represents the convolutional layer, represents a batch normalization layer, represents the global pooling layer, ω * represents the weight corresponding to the convolutional layer, δ r represents the activation function;

[0076] Step 4: Use a simple fully connected network classifier to classify the fused RF fingerprint features and finally identify different radiation source individuals;

[0077] Define a classifier consisting of fully connected layers as follows:

[0078]

[0079] In the above formula, represents the label output by the classifier, represents three fully connected layers, ω F represents the weight parameter corresponding to the fully connected layer. The input of the classifier is the output of step 3, which is the RF fingerprint feature after the fusion of each branch.

[0080] Step 5: Select the network optimizer as Adaptive Moment Estimation (Adam), set the learning rate to 0.0001, build the model, and start network training:

[0081] Define the network input: multi-source heterogeneous RF signal: y k-R1 ,y k-R2 ,y k-R3 ; Corresponding device label: C k ; Initial learning rate: η; Maximum number of training times: τ;

[0082] Define the output of the network: Identified device labels:

[0083] Import training data: C k ;

[0084] Initialize network model weights: ω *i ,ω Fi ;

[0085] Training process:

[0086] For each epoch (the number of times the entire training data set is forward propagated and back propagated through the neural network):

[0087] Multi-channel parallel extraction of radio frequency fingerprint features of multi-source heterogeneous data:

[0088]

[0089] In the above formula, represents the first branch pooling layer, represents the second branch pooling layer, Represents the third branch pooling layer;

[0090] RF fingerprint feature fusion based on attention mechanism:

[0091]

[0092] Call the classifier based on the fully connected layer to perform device identification on the fused RF fingerprint:

[0093]

[0094] Calculate the loss function loss t :

[0095]

[0096] Determine whether the loss function has a convergence trend. If the loss function does not converge in the convergence direction after 20 epochs, dynamically adjust the algorithm learning rate:

[0097] η=η×0.7

[0098] Otherwise, use the Adam optimizer to dynamically update the network model weights:

[0099]

[0100] At this point, a complete network model training is completed;

[0101] Update network weight parameters according to model classification accuracy;

[0102] Step 6: Retrain and obtain the final model:

[0103] According to the results of multiple iterative training cycles, the final radiation source individual identification model is obtained.

[0104] In order to verify the method of the present invention, we established a multi-source heterogeneous receiving platform based on a laboratory environment. The experimental results show that the method of the present invention has good recognition performance under ideal conditions. Even in the presence of noise interference, the method of the present invention can still maintain its performance advantage. Finally, we verified that the method of the present invention can still maintain stable operation in real-world scenarios where the sensor is damaged, which is a major advantage compared to traditional radiation source individual identification methods. The method of the present invention is suitable for scenarios such as industrial Internet of Things, Internet of Vehicles, digital twins, and the Internet of Everything, with the goal of further ensuring the safety of high-tech life in society.

[0105] This embodiment also discloses a multi-source heterogeneous radiation source individual recognition model based on adaptive fingerprint feature fusion, which is implemented by the above method steps, such as Figure 2 As shown in the figure, it consists of a multi-channel RF fingerprint feature extraction module, an RF fingerprint fusion module based on attention mechanism and an automatic classifier.

[0106] like Figure 3 As shown, the x-axis is the mask ratio, which indicates the proportion of the original signal covered; the y-axis indicates the recognition accuracy of different algorithms in this experiment. The larger the mask ratio, the longer it takes for the sensor to be damaged or broken. The red solid line represents the performance curve of the present invention, while the cyan and green dotted lines represent the performance of the comparison algorithm multi-scale convolutional neural network and convolutional neural network, respectively. (a), (b), (c) represent the damage or downtime of receiver 1, receiver 2, and receiver 3, respectively. The experimental results show that under complex working conditions, the model of the present invention can maintain stable recognition performance even when one receiver is interrupted.

Claims

1. A method for identifying individual multi-source heterogeneous radiation sources based on adaptive fingerprint feature fusion, characterized in that: The steps include: Step 1: Use USRP to collect radio signals of different modes at different locations, including original RF signals, signals after OFDM cyclic code decoding, and signals after channel equalization; Step 2, using three parallel branches of one-dimensional or two-dimensional convolutional layers, pooling layers, and fully connected layers to extract corresponding RF fingerprints for the three different signals respectively; Step 3: Use the feature fusion module based on the attention mechanism to fuse the three RF fingerprints extracted in the previous step; The general expression for defining the RF fingerprint feature fusion module is as follows: In the above formula, Γ mix Represents the fused RF fingerprint feature, Indicates the multiplication of corresponding elements in the matrix, ω ai Represents the radio frequency fingerprint feature Γ Ri and Γ R(i+1) The attention weight calculated after the attention module; Attention weight ω ai The calculation formula is defined as follows: In the above formula, σ[·] represents the activation function, ∩(·) represents the local attention operation, which consists of a convolutional layer and a batch normalization layer; ∪(·) represents the global attention operation. The two attention mechanism operations ∩(·) and ∪(·) are defined as follows: In the above formula, represents the convolutional layer, represents a batch normalization layer, represents the global pooling layer, ω * represents the weight corresponding to the convolutional layer, δ r represents the activation function; Step 4: Use a fully connected network classifier to classify the fused RF fingerprint features; Define a classifier consisting of fully connected layers as follows: In the above formula, represents the label output by the classifier, represents three fully connected layers, ω F represents the weight parameter corresponding to the fully connected layer. The input of the classifier is the output of step 3, which is the RF fingerprint feature after the fusion of each branch. Step 5: Select the network optimizer as the adaptive moment estimation optimizer Adam, build the model, and start network training: Define the network input: multi-source heterogeneous RF signal: y k-R1 ,y k-R2 ,y k-R3 ; Corresponding device label: C k ; Initial learning rate: η; Maximum number of training times: τ; Define the output of the network: Identified device labels: Import training data: Initialize network model weights: ω *i ,ω Fi ; Training process: For each epoch: Multi-channel parallel extraction of radio frequency fingerprint features of multi-source heterogeneous data: In the above formula, represents the first branch pooling layer, represents the second branch pooling layer, Represents the third branch pooling layer; RF fingerprint feature fusion based on attention mechanism: Call the classifier based on the fully connected layer to perform device identification on the fused RF fingerprint: Calculate the loss function loss t : Determine whether the loss function has a convergence trend. If the loss function does not converge in the convergence direction after 20 epochs, dynamically adjust the algorithm learning rate: η=η×0.7 Otherwise, use the Adam optimizer to dynamically update the network model weights: At this point, a complete network model training is completed; Step 6: Retrain and obtain the final model: According to the results of multiple iterative training cycles, the final radiation source individual recognition model is obtained.

2. According to claim 1, the method for identifying individual multi-source heterogeneous radiation sources based on adaptive fingerprint feature fusion is characterized by: In step 3, the activation function σ[·] is selected as the relu function or the sigmoid function.

3. According to claim 1, the method for identifying individual multi-source heterogeneous radiation sources based on adaptive fingerprint feature fusion is characterized by: In step 5, the learning rate is set to 0.0001.

Images