Multi-view feature fusion radiation source individual identification method

Abstract

The invention discloses a radiation source individual identification method based on multi-view feature fusion, which comprises the following steps of: firstly, normalizing a radiation source individual identification sample set, and dividing the radiation source individual identification sample set into a training set and a test set; constructing a multi-view feature extraction fusion model as a radiation source individual identification model, training the model by using a training set, calculating classification loss by using a cross entropy loss function, and obtaining the radiation source individual identification model through back propagation of model weight and training; and finally, inputting the test set into a trained individual identification model, outputting identification tags and identification probabilities, and taking the tag with the maximum identification probability as an identification result. According to the method, each radiation individual can be accurately identified, and a higher identification rate can be achieved under the condition of different signal-to-noise ratios, especially under the condition of low signal-to-noise ratio.

Description

Technical Field

[0001] The present invention relates to radio physical layer security technology, in particular to a method for identifying individual radiation sources by fusing multi-view features. Background Art

[0002] With the rapid development of electronic information technology, wireless communication systems have been widely used in various industries. The status of wireless communication systems has become increasingly important, leading to increasing attention to the information security of wireless communication devices. Due to the openness of electromagnetic waves, traditional wireless network security based on key security protocols has potential security risks. Physical-layer radio frequency fingerprint recognition has the advantage of difficult-to-forge features and can effectively and correctly identify the identity of radio frequency devices. It is a research hotspot in the field of wireless network security.

[0003] In recent years, artificial intelligence technologies, represented by deep learning, have experienced significant development, achieving significant success in image and speech recognition, and are rapidly being applied to radio frequency fingerprint recognition. Neural networks are artificial neural network models designed based on the structure and functional characteristics of biological neural systems, boasting strong adaptability and nonlinear mapping capabilities. The multi-layer perceptron (MLP) is one of the most basic neural network models. It is a classic feedforward neural network structure based on the universal approximation theorem, which states that a single hidden layer network with a sufficiently large number of neurons can approximate any continuous function. An MLP network consists of multiple neurons (or nodes), interconnected by weights, forming a multi-layer network structure. Each neuron receives signals from other neurons, applies a weighted linear combination to these signals, and then performs a nonlinear transformation using an activation function. The output is ultimately transmitted to the next layer of neurons, or the output layer.

[0004] Convolutional Neural Networks (CNN) is a type of feedforward neural network with a deep structure that includes convolution calculations. It is one of the representative algorithms of deep learning and has the ability to represent learning. The gated recurrent unit GRU (GRU) is a typical recurrent neural network RNN ​​(RNN) structure. When calculating the current output, it can not only use the input information of the current moment, but also use the input information of multiple moments before. It is often used for feature extraction of sequence data. Fourier Analysis Networks (FAN) are based on Fourier analysis, that is, any periodic All functions can be expressed as an infinite series of orthogonal sine or cosine basis functions. The network architecture of each layer of FAN is similar to that of MLP. The activation functions of the nodes in MLP are replaced by a combination of sine, cosine and activation functions. The combination parameters are learnable, which makes FAN have powerful periodic modeling capabilities while maintaining the versatility similar to MLP. Summary of the Invention

[0005] The purpose of this invention is to design a multi-view feature extraction and fusion model based on traditional neural network technology for individual radiation source identification. This method provides a multi-view feature fusion method for individual radiation source identification based on the source's IQ signal. This method can accurately identify individual radiation sources and achieve higher recognition rates under different signal-to-noise ratios, especially low signal-to-noise ratios.

[0006] The objects of the present invention are achieved by:

[0007] A method for identifying individual radiation sources by fusion of multi-view features includes the following steps:

[0008] 1) Signal Preprocessing: Received IQ signals were preprocessed using Matlab. Every 256 sampling points in a useful data segment were considered as a valid signal segment. 1000 valid signal segments were screened for each individual radiator. Each signal segment was then energy normalized. The IQ signal contains both in-phase and quadrature components. The processed valid signals were concatenated into a 256×2 two-dimensional signal. The data for each individual was divided into a training set and a test set at a ratio of 9:1.

[0009] 2) Constructing a multi-view feature extraction fusion model: The multi-view feature extraction fusion model has an input layer, a hidden layer, and an output layer. The backbone network consists of an IQ dual-path convolutional neural network layer, a GRU layer, and three stacked FAN layers in parallel. The GRU consists of a reset gate and an update gate. Its current output is determined by the output of the previous moment and the input of the current moment. In the FAN architecture, the node is replaced by an activation function with a combination of a sine function, a cosine function, and an activation function. Different feature extraction layers have different input requirements. The IQ valid signal segment with a dimension of 256×2 is input into the IQ dual-path convolutional neural network layer and the GRU layer. The input data dimension is transformed into a valid signal segment with a dimension of 512×1 and input into the three stacked FAN layers. The forward propagation expression of the GRU is shown in formula (1):

[0010] (1),

[0011] in, Indicates the current moment input, Indicates the output at the previous moment, 、 、 is the learnable weight, [*] represents the element-wise product operation, and the final output obtained after the reset gate and the update gate is shown in formula (2):

[0012] (2),

[0013] Among them, 1 means All-1 vectors of the same length;

[0014] The output expression of the FAN layer is shown in formula (3):

[0015] (3),

[0016] in, 、 、 are learnable parameters, hyperparameters and Respectively and The first dimension of [・||・] represents the splicing along the first dimension, , represents the activation function, and then the features extracted by the three different networks are spliced ​​together. The hidden layer dimensions of the three networks are 128×1, 64×1, and 256×1 respectively. The hidden layer dimension after splicing is 448×1, and the output layer is the number of radiation source individuals.

[0017] 3) Using the training set to train the multi-view feature extraction fusion model: The training set data is fed into the IQ dual-path convolutional neural network layer, the GRU layer, and the three-stacked FAN layer. After feature extraction, the fusion is performed, and then softmax classification is performed to obtain the predicted output label, which is then compared with the actual label. The training phase is carried out using the backpropagation algorithm, with the goal of adjusting the learnable weights in the three feature extraction networks to minimize the error between the predicted output and the actual label.

[0018] 4) Use the test set to test the trained multi-view feature extraction fusion model. Input the test set data into the trained network to extract features. After softmax classification, the label with the highest output probability is used as the recognition result.

[0019] In step 3, the number of GRU hidden layers is 64, the number of FAN hidden layers is 256, the number of training iterations is 40, the batch size is 32, and the learning rate is set to 0.001.

[0020] This technical solution uses three different feature extraction networks to directly extract IQ signal correlation, IQ signal timing and IQ signal periodicity features from the original IQ sequence and fuse them. It does not require prior knowledge, is simple to implement, has low computational complexity, and is easy to deploy in practical applications.

[0021] This method can accurately identify each radiating individual and has a higher recognition rate under different signal-to-noise ratios, especially low signal-to-noise ratio conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 Schematic diagram of the process of the embodiment method;

[0023] Figure 2 Schematic diagram of the GRU structure in the embodiment;

[0024] Figure 3 Schematic diagram of the FAN structure in the embodiment;

[0025] Figure 4 This figure shows the accuracy of the multi-view feature extraction fusion model in the embodiment for identifying radiation source individuals of 6 Wi-Fi individuals compared with other commonly used models under different signal-to-noise ratios. DETAILED DESCRIPTION

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments, but the present invention is not limited thereto.

[0027] Example:

[0028] Reference Figure 1 A method for identifying individual radiation sources by fusion of multi-view features includes the following steps:

[0029] 1) Signal Preprocessing: Received IQ signals were preprocessed using Matlab. Every 256 sampling points in a useful data segment were considered as a valid signal segment. 1000 valid signal segments were screened for each individual radiator. Each signal segment was then energy normalized. The IQ signal contains both in-phase and quadrature components. The processed valid signals were concatenated into a 256×2 two-dimensional signal. The data for each individual was divided into a training set and a test set at a ratio of 9:1.

[0030] 2) Constructing a multi-view feature extraction fusion model: The multi-view feature extraction fusion model has an input layer, a hidden layer, and an output layer. The backbone network consists of an IQ dual-path convolutional neural network layer, a GRU layer, and three stacked FAN layers in parallel. Figure 2 As shown, GRU consists of a reset gate and an update gate, and its current output is determined by the output of the previous moment and the input of the current moment, as shown in Figure 3 As shown in the figure, the activation function of FAN is replaced by a combination of sine function, cosine function and activation function. The input requirements of different feature extraction layers are different. The IQ valid signal segment with dimension 256×2 is input into the IQ dual-path convolutional neural network layer and GRU layer. The input data dimension is transformed into a valid signal segment with dimension 512×1 and input into the three-stacked FAN layer. The forward propagation expression of GRU is shown in formula (1):

[0031] (1),

[0032] in, Indicates the current moment input, Indicates the output at the previous moment, 、 、 is the learnable weight, [*] represents the element-wise product operation, and the final output obtained after the reset gate and the update gate is shown in formula (2):

[0033] (2),

[0034] Among them, 1 means All-1 vectors of the same length;

[0035] The output expression of the FAN layer is shown in formula (3):

[0036] (3),

[0037] in, 、 、 are learnable parameters, hyperparameters and Respectively and The first dimension of [・||・] represents the splicing along the first dimension, , represents the activation function, and then the features extracted by the three different networks are spliced ​​together. The hidden layer dimensions of the three networks are 128×1, 64×1, and 256×1 respectively. The hidden layer dimension after splicing is 448×1, and the output layer is the number of radiation source individuals.

[0038] 3) Use the training set to train the multi-view feature extraction fusion model: The training set data is fed into an IQ dual-path convolutional neural network layer, a GRU layer, and a three-stacked FAN layer. After feature extraction, the layers are fused and softmax classified to obtain the predicted output labels, which are then compared with the actual labels. The training phase is performed using a backpropagation algorithm, aiming to adjust the learnable weights in the three feature extraction networks to minimize the error between the predicted output and the actual labels. In this example, the GRU has 64 hidden layers, the FAN has 256 hidden layers, the training iterations are 40, the batch size is 32, and the learning rate is set to 0.001.

[0039] 4) Use the test set to test the trained multi-view feature extraction fusion model. Input the test set data into the trained network to extract features. After softmax classification, the label with the highest output probability is used as the recognition result.

[0040] like Figure 4 As shown in the figure, the accuracy of this method in identifying 6 Wi-Fi individual radiators compared with other commonly used models under different signal-to-noise ratios is shown in the figure. When the signal-to-noise ratio is greater than 18dB, the recognition accuracy of this method reaches more than 98.5%, indicating that the model can accurately identify each radiator. When the signal-to-noise ratio is 10dB, the recognition accuracy of this method reaches 97%, indicating that the model can achieve a higher recognition rate under different signal-to-noise ratios, especially under low signal-to-noise ratio conditions.

Claims

1. A method for identifying individual radiation sources by fusion of multi-view features, characterized in that: The steps include: 1) Signal Preprocessing: Received IQ signals were preprocessed using Matlab. Every 256 sampling points in a useful data segment were considered as a valid signal segment. 1000 valid signal segments were screened for each individual radiator. Each signal segment was then energy normalized. The IQ signal contains both in-phase and quadrature components. The processed valid signals were concatenated into a 256×2 two-dimensional signal. The data for each individual was divided into a training set and a test set at a ratio of 9:

1. 2) Constructing a multi-view feature extraction fusion model: The multi-view feature extraction fusion model has an input layer, a hidden layer, and an output layer. The backbone network consists of an IQ dual-path convolutional neural network layer, a GRU layer, and three stacked FAN layers in parallel. The GRU consists of a reset gate and an update gate. The output at the current moment is determined by the output at the previous moment and the input at the current moment. In the FAN architecture, the node activation function is replaced by a combination of a sine function, a cosine function, and an activation function. Different feature extraction layers have different input requirements. The IQ valid signal segment with a dimension of 256×2 is input into the IQ dual-path convolutional neural network layer and the GRU layer. The input data dimension is transformed into a valid signal segment with a dimension of 512×1 and input into the three stacked FAN layers. The forward propagation expression of the GRU is shown in formula (1): （1）， in, Indicates the current moment input, Indicates the output at the previous moment, 、 、 is the learnable weight, [*] represents the element-wise product operation, and the final output obtained after the reset gate and the update gate is shown in formula (2): （2）， Among them, 1 means All-1 vectors of the same length; The output expression of the FAN layer is shown in formula (3): （3）， in, 、 、 are learnable parameters, hyperparameters and Respectively and The first dimension of [・||・] represents the splicing along the first dimension, , represents the activation function, and then the features extracted by the three different networks are spliced ​​together. The hidden layer dimensions of the three networks are 128×1, 64×1, and 256×1 respectively. The hidden layer dimension after splicing is 448×1, and the output layer is the number of radiation source individuals. 3) Using the training set to train the multi-view feature extraction fusion model: The training set data is fed into the IQ dual-path convolutional neural network layer, the GRU layer, and the three-stacked FAN layer. After feature extraction, the fusion is performed, and then softmax classification is performed to obtain the predicted output label, which is then compared with the actual label. The training phase is carried out using the backpropagation algorithm, with the goal of adjusting the learnable weights in the three feature extraction networks to minimize the error between the predicted output and the actual label. 4) Use the test set to test the trained multi-view feature extraction fusion model. Input the test set data into the trained network to extract features. After softmax classification, the label with the highest output probability is used as the recognition result.

2. The radiation source individual identification method based on multi-view feature fusion according to claim 1 is characterized in that: In step 3, the number of GRU hidden layers is 64, the number of FAN hidden layers is 256, the number of training iterations is 40, the batch size is 32, and the learning rate is set to 0.001.

Images