A defense method for deep receivers based on feature enhancement

Abstract

The present invention provides a defense method for a deep receiver based on feature enhancement, including constructing a data set and training a deep receiver; constructing a generative adversarial network, the constructed generative adversarial network includes a generator, a discriminator, and a demodulator; pre-training the generative adversarial network; training the generative adversarial network, and finally obtaining a generator with sufficient performance; inputting a radio signal into the generator obtained in step (4) for processing, obtaining a feature-enhanced electrical signal, and then inputting it into the deep receiver to achieve a low bit error rate for decoding the radio signal. The signal enhanced by the generator is input into the deep receiver, and the enhanced signal carries the feature vector of the clean sample, which can not only eliminate the adversarial disturbance and reduce the decoding bit error rate of the adversarial sample, but also enhance the decoding accuracy of the clean sample.

Description

Technical Field

[0001] This patent relates to the security field of artificial intelligence and wireless communications, and specifically to a defense method for deep receivers based on feature enhancement. Background Art

[0002] The wireless communication system consists of a transmitter and a receiver. The transmitter performs channel coding, modulation and pulse shaping on the signal before sending it. The receiver of the traditional wireless communication system uses processes such as carrier and symbol synchronization, channel estimation, equalization, demodulation and decoding to recover information from the received distorted signal. However, each processing module of the traditional receiver is based on a theoretically assumed scenario, and this step-by-step signal processing method of the receiver has the problem of poor signal recovery effect.

[0003] The neural network-based deep receiver model replaces the entire information recovery process at the receiver end of the traditional wireless communication system, and its input can be the signal transmitted by the transmitter end of the traditional wireless communication system, and the output is the recovered bit stream information. The model is trained based on the received IQ signal samples and can reflect the RF impairments, channel fading, noise and interference actually encountered by the communication system.

[0004] However, there are some problems with the reliability and robustness of deep receivers. One of the problems is the existence of adversaries in the deep learning training process or inference process. The broadcasting characteristics of radio determine that radio signals may be received by multiple receivers, including legitimate receivers and illegal attackers. Illegal attackers send small-power adversarial interference signals through the propagation channel and superimpose them on the normal radio signals of the transmitter. Such radio signals will increase the bit error rate of deep receivers and are difficult to detect due to their low power. Especially in some scenarios with higher security levels, such as the application of receivers based on deep neural networks in military environments, if there are security vulnerabilities and the adversarial signals are incorrectly decoded by the receiver, serious consequences may occur, such as the inability to communicate and locate in real time.

[0005] Some existing defense schemes, such as adversarial training, have low generality and limitations due to their reliance on limited adversarial examples. In addition, some schemes even have a non-negligible negative impact on the decoding accuracy of clean samples.

[0006] Based on the above considerations, this patent proposes a robust deep receiver device based on disturbance enhancement, which improves the decoding accuracy of normal samples on the basis of reducing the decoding error rate of the deep receiver for adversarial samples. Summary of the invention

[0007] The purpose of the present invention is to address the deficiencies of the prior art and provide a defense method for a deep receiver based on feature enhancement.

[0008] The object of the present invention is achieved through the following technical solution: a defense method for a deep receiver based on feature enhancement, comprising the following steps:

[0009] Step (1): Build a dataset and train a deep receiver;

[0010] Step (2): construct a generative adversarial network, which includes a generator, a discriminator, and a demodulator;

[0011] Step (3): Pre-train the generative adversarial network;

[0012] Step (4): Train the generative adversarial network and finally obtain the generator;

[0013] Step (5): The radio signal is input into the generator obtained in step (4) for processing to obtain a feature-enhanced electrical signal, and then the feature-enhanced electrical signal is input into a deep receiver to achieve a low bit error rate for decoding the radio signal.

[0014] Furthermore, the step (1) includes the following sub-steps:

[0015] (1.1) Constructing a data set: The constructed data set includes a sample signals, which are divided into b types of sample signals. C% of the sample signals are extracted from each type of sample signals as a training set. The training set participates in the subsequent model generation training. The sample signals are all clean sample signals. a, b, and c are natural numbers.

[0016] (1.2) Training the deep receiver:

[0017] The deep receiver is composed of a dense convolutional neural network; the hyperparameters set during the training of the deep receiver include: the number of model training rounds of the deep receiver, the number of batch sample signals of the deep receiver, the gradient update rule of the deep receiver and the training loss function of the deep receiver; the number of model training rounds of the deep receiver is Epoch dense , the number of batch sample signals of the depth receiver is M 1 , the gradient update rule of the deep receiver is stochastic gradient descent SGD, and the deep receiver training loss function Loss dense It is expressed as the following formula:

[0018]

[0019] Where superscript i represents the i-th sample signal, i = 1, 2, ..., M 1 , M 1 is the number of sample signals in each batch processing, N is the number of bits, j = 1, 2, ..., N, is the true label of the kth category of the i-th input sample signal at the j-th position, that is Indicates that the true label is 0, Indicates that the true label is 1, p j (x i ) represents the depth receiver's response to the sample signal x i The bit prediction probability of the j-th bit of the output.

[0020] Furthermore, the step (2) includes the following sub-steps:

[0021] (2.1) Build a generator:

[0022] The generator is composed of an encoder and a decoder, the encoder is composed of a convolutional neural network and a long short-term memory neural network, and the decoder is composed of a deconvolutional neural network;

[0023] (2.2) Build a discriminator:

[0024] The discriminator is composed of a convolutional neural network and a long short-term memory neural network;

[0025] (2.3) Build a demodulator;

[0026] The demodulator consists of a dense convolutional neural network.

[0027] Furthermore, the step (3) includes the following sub-steps:

[0028] (3.1) Discriminator pre-training

[0029] The discriminator is pre-trained. The training parameters set during the pre-training of the discriminator include: the number of pre-training rounds of the discriminator, the number of batch sample signals of the discriminator, the gradient update rule of the discriminator, and the pre-training loss function of the discriminator D; the number of pre-training rounds of the discriminator is Epoch k , the number of batch processing sample signals of the discriminator is M 2 , the gradient update rule of the discriminator is stochastic gradient descent SGD, and the pre-training loss function Loss of the discriminator D is D (x,ω D ) is as follows:

[0030]

[0031] where x i is a clean sample signal, the superscript i represents the i-th sample signal, i = 1, 2, ..., M 2 , M 2 is the number of sample signals in each batch processing; D(x i ;ω D ) indicates that the discriminator judges xi The probability of whether it is a clean sample signal, if it is greater than the threshold, it is considered a clean sample signal, and if it is less than the threshold, it is considered a non-clean sample signal; ω D represents the weight parameter of the discriminator; taking the log of the output probability D(·;·) is to convert the probability of event occurrence into information;

[0032] (3.2) Demodulator Pre-training

[0033] Pre-train the demodulator and set the training parameters: number of pre-training epochs k , the number of batch processing sample signals M 3 , the gradient update rule is set to stochastic gradient descent SGD, and the pre-trained loss function Loss of the demodulator C C (x,ω c ) is as follows:

[0034]

[0035] where x i is a clean sample signal, the superscript i represents the i-th clean sample signal, i = 1, 2, ..., M 3 , M 3 is the number of sample signals in each batch processing; N is the number of bits, j = 1, 2, ..., N, C j (x i ;ω C ) represents the demodulator C's response to the sample signal x i The predicted probability of the jth bit of the output is, represents the true label of the k-th category of the i-th input sample at the j-th position; ω C Represents the weight parameters of the demodulator.

[0036] Furthermore, the step (4) includes the following sub-steps:

[0037] (4.1) Generate noise sample signal: Get random noise sample signal z from Gaussian distribution 1 ,z 2 ,…,z i ,…z M4 , z i The dimension and clean sample signal x i The dimension is the same, which is 2*n;

[0038] (4.2) Substitute the noise sample signal z obtained in step (4.1) i and the clean sample signal x i Overlapping combination The dimension is 4*n, and the Input to the generator;

[0039] (4.3) The generator generates an enhanced noise sample signal: the input to the generator After being processed by the encoder and decoder in the generator, the enhanced noise sample signal v is output i ; The enhanced noise sample signal v i and the clean sample signal x i The dimensions are the same;

[0040]

[0041] Where G(;) represents the generator, ω G Represents the weight parameters of the generator.

[0042] (4.4) Enhance the noise sample signal v i and the clean sample signal x i Add to obtain the enhanced sample signal x I :

[0043] x I =x i +v i

[0044] (4.5) Enhanced sample signal x I and the clean sample signal x i Input to the discriminator: respectively enhance the sample signal x I and the clean sample signal x i Input to the discriminator, the output of the discriminator is the probability output of whether the input sample signal is a clean sample signal, and the output interval of the discriminator ranges from [0,1]. By improving its stochastic gradient update discriminator, the loss function of discriminator training is:

[0045]

[0046] Where D(·; ·) represents the probability of the discriminator determining whether the input sample is a clean sample, M 4 is the number of training rounds for the discriminator;

[0047] The output of the discriminator D is back-propagated, that is, the output result is input into the discriminator D again;

[0048] (4.6) The discriminator output is fed back to the generator: According to the output of the discriminator in step (4.5), the generator is updated by improving the stochastic gradient; the training loss function fed back from the discriminator output to the generator for:

[0049]

[0050] in represents the loss function where the output of the discriminator is fed back to the generator, ω D represents the weight parameter of the discriminator, ω G Represents the weight parameters of the generator;

[0051] (4.7) Enhanced sample signal x I and the clean sample signal x i Input to the demodulator: The enhanced sample signal x I and the clean sample signal x i Input to the demodulator, the demodulator outputs the probability of each bit Back-propagate the output of the demodulator C and update the demodulator by improving its stochastic gradient; the training loss function of the demodulator is Loss C (x,ω c ) is as follows:

[0052]

[0053] Among them, M 4 is the number of sample signals processed by the demodulator in each batch, which is the same as the number of sample signals processed by the discriminator in each batch in step (4.5); ω C represents the weight parameter of the demodulator; N is the number of bits, j = 1, 2, ..., N, is the true label of the kth category of the i-th input sample signal at the j-th position, that is Indicates that the true label is 0, Indicates that the true label is 1;

[0054] The output of demodulator C is back-propagated, that is, the output result is input into demodulator C again;

[0055] (4.8) The demodulator output is fed back to the generator:

[0056] Feed the output of the demodulator in step (4.7) back to the generator for training; update the generator by improving the stochastic gradient, and the training loss function of the demodulator for:

[0057]

[0058] in represents the loss function of the demodulator output fed back to the generator; M 4 is the number of sample signals processed by the generator G in each batch, which is the same as the number of sample signals processed by the discriminator D in each batch in step (4.5); N is the number of bits, j = 1, 2, ..., N, is the true label of the kth category of the i-th input sample signal at the j-th position, that is Indicates that the true label is 0, Indicates that the true label is 1

[0059] Return to step (4.1) and repeat steps (4.2)-(4.8) until the generator, discriminator, and demodulator converge, and save the output to obtain the generator.

[0060] The technical concept of the present invention is: in view of the risk that a deep receiver based on deep learning may be attacked by potential illegal attackers, resulting in an increased bit error rate, the present invention provides a defense method for a deep receiver based on feature enhancement, in which the feature enhancement generated by a trained generator is added to the radio signal to achieve disturbance elimination and feature enhancement in the radio signal, which is then input into the deep receiver, ultimately achieving a low bit error rate for decoding the radio signal by the deep receiver.

[0061] The beneficial effects of the present invention are as follows: 1. After the radio signal is de-perturbed and feature-enhanced by a trained generator, it is then input into a deep receiver, which not only eliminates adversarial disturbances but also reduces the decoding bit error rate of the radio signal;

[0062] 2. Compared with common training methods, the present invention only needs to use clean samples for training during the training process, which is more robust and generalizable than common training methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0064] Figure 1 A flowchart of a defense method for deep receivers based on feature enhancement;

[0065] Figure 2 Diagram of the process of training a generative adversarial network. DETAILED DESCRIPTION

[0066] 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 implementation methods described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0067] Reference Figure 1-2 , a defense method based on a deep receiver with feature enhancement, comprising the following steps:

[0068] Step (1): Build a dataset and train a deep receiver;

[0069] Step (1) is implemented through the following sub-steps:

[0070] (1.1) Constructing a data set: The constructed data set includes a sample signals, which are divided into b categories of sample signals. C% of the sample signals are extracted from each category of sample signals as the training set. The training set participates in the subsequent model generation training. The sample signals are all clean sample signals, and a, b, and c are natural numbers.

[0071] (1.2) Training the deep receiver:

[0072] The deep receiver is composed of a dense convolutional neural network; the hyperparameters set during the training of the deep receiver include: the number of model training rounds of the deep receiver, the number of batch sample signals of the deep receiver, the gradient update rule of the deep receiver and the training loss function of the deep receiver; the number of model training rounds of the deep receiver is Epoch dense , the number of batch sample signals of the depth receiver is M 1 , the gradient update rule of the deep receiver is stochastic gradient descent SGD, and the deep receiver training loss function Loss dense It is expressed as the following formula:

[0073]

[0074] Where superscript i represents the i-th sample signal, i = 1, 2, ..., M 1 , M 1 is the number of sample signals in each batch processing, N is the number of bits, j = 1, 2, ..., N, is the true label of the kth category of the i-th input sample signal at the j-th position, that is Indicates that the true label is 0, Indicates that the true label is 1, p j (x i ) represents the depth receiver's response to the sample signal x i The bit prediction probability of the jth bit of the output of

[0075] In the embodiment, the hyperparameters of the training setting are: Model training round number Epoch dense =10, the number of sample signals in each batch processing is M 1 =32, the gradient update rule is set to stochastic gradient descent SGD.

[0076] Step (2): construct a generative adversarial network (GAN), which includes a generator G, a discriminator D, and a demodulator C;

[0077] Step (2) is implemented through the following sub-steps:

[0078] (2.1) Build a generator G;

[0079] The generator G is composed of an encoder and a decoder, wherein the encoder is composed of a convolutional neural network (CNN) and a long short-term memory neural network (LSTM), and the decoder is composed of a deconvolutional neural network (DCNN). Combining LSTM and CNN enables the generator G to extract the time domain features of the input signal.

[0080] (2.2) Build a discriminator D;

[0081] The discriminator D is composed of a convolutional neural network (CNN) and a long short-term memory neural network (LSTM). The discriminator is used to extract the features of the time series signal as much as possible to determine the similarity between the input sample and the clean sample;

[0082] (2.3) Build a demodulator C;

[0083] The demodulator C is composed of a dense convolutional neural network. The demodulator C and the target model have the same model structure but different weight parameters. The demodulator C is mixed trained on clean samples and samples generated from the generator G, while the deep receiver is only trained on clean samples.

[0084] Step (3): Pre-training the Generative Adversarial Network (GAN);

[0085] Step (3) is implemented through the following sub-steps:

[0086] (3.1) Discriminator D pre-training

[0087] A certain amount of sample signal x i The input is input into the discriminator D, and the discriminator D is pre-trained so that the discriminator D has a certain ability to distinguish whether the sample signal is a clean sample signal. The training parameters set during the discriminator pre-training include: the number of pre-training rounds of the discriminator, the number of batch sample signals of the discriminator, the gradient update rule of the discriminator, and the pre-training loss function of the discriminator D; the number of pre-training rounds of the discriminator is Epoch k , the number of batch processing sample signals of the discriminator is M 2 , the gradient update rule of the discriminator is stochastic gradient descent SGD, and the pre-training loss function Loss of the discriminator D is D (x,ω D ) is as follows:

[0088]

[0089] where x i is a clean sample signal, the superscript i represents the i-th sample signal, i = 1, 2, ..., M 2 , M 2is the number of sample signals in each batch processing; D(x i ;ω D ) indicates that the discriminator D judges the sample signal x i The probability of whether it is a clean sample signal, if it is greater than the threshold, it is considered a clean sample signal, and if it is less than the threshold, it is considered a non-clean sample signal; ω D Represents the weight parameter of the discriminator D; taking the log of the output probability D(·; ·) is to convert the probability of an event into information, which is a common setting in the loss formula.

[0090] In this embodiment, set the training parameters: number of pre-training rounds Epoch k =10, the number of batch processing sample signals M 2 =32, and the gradient update rule is set to stochastic gradient descent SGD.

[0091] After pre-training, the discriminator D has a certain ability to distinguish whether it is a clean sample signal.

[0092] (3.2) Demodulator C Pre-training

[0093] A certain amount of clean sample signal x i The demodulator C is inputted and pre-trained to enable the demodulator C to have a certain demodulation capability for the clean sample signal. In this embodiment, the training parameters set during the demodulator pre-training include: the number of pre-training rounds of the demodulator, the number of batch sample signals of the demodulator, the gradient update rule of the demodulator, and the pre-training loss function of the demodulator C; the number of pre-training rounds of the demodulator is Epoch k , the number of batch processing sample signals of the demodulator is M 3 , the gradient update rule of the demodulator is stochastic gradient descent SGD. The demodulator C only inputs clean signal samples x i , back-propagate the output result, and pre-train the demodulator C by improving its stochastic gradient. The pre-training loss function of the demodulator C is Loss C (x,ω c ) is as follows:

[0094]

[0095] where x i is a clean sample signal, the superscript i represents the i-th clean sample signal, i = 1, 2, ..., M 3 , M 3 is the number of sample signals in each batch processing; N is the number of bits, j = 1, 2, ..., N, C j (x i ;ω C ) represents the demodulator C for the clean sample signal x iThe predicted probability of the jth bit of the output is, represents the true label of the k-th category of the i-th input sample at the j-th position; ω C Represents the weight parameter of demodulator C.

[0096] In this embodiment, the number of pre-training rounds is set to Epoch k =10, the number of batch samples M 3 =32, and the gradient update rule is set to stochastic gradient descent SGD.

[0097] After the demodulator C is pre-trained, the demodulator C has a certain demodulation capability for the clean sample signal.

[0098] Step (4): Train the Generative Adversarial Network (GAN) to finally obtain the generator;

[0099] Step (4) is implemented through the following sub-steps:

[0100] (4.1) Generate noise sample signal: Get random noise sample signal z from Gaussian distribution 1 ,z 2 ,…,z i ,…z M4 , z i The dimension and clean sample signal x i The dimensions are the same, both are 2*n;

[0101] (4.2) Substitute the noise sample signal z obtained in step (4.1) i and the clean sample signal x i Overlapping combination The dimension is 4*n, and the Input into generator G;

[0102] (4.3) Generator G generates an enhanced noise sample signal: After being processed by the encoder and decoder in the generator G, the enhanced noise sample signal v is output i ; The enhanced noise sample signal v i and the clean sample signal x i The dimension is the same, which is 2*n;

[0103]

[0104] Where G(;) represents the generator G, ω G Represents the weight parameters of the generator G.

[0105] (4.4) Enhance the noise sample signal v i and the clean sample signal xi Add to obtain the enhanced sample signal x I :

[0106] x I =x i +v i

[0107] (4.5) Enhanced sample signal x I and the clean sample signal x i Input to the discriminator D: respectively enhance the sample signal x I and the clean sample signal x i Input to the discriminator D, the output of the discriminator D is the probability output of whether the input sample signal is a clean sample signal, and the output interval of the discriminator D ranges from [0,1]. By improving its stochastic gradient update discriminator D, the loss function of the discriminator D training is:

[0108]

[0109] Where D(·; ·) represents the probability of the discriminator D to determine whether the input sample is a clean sample; M 4 is the number of sample signals processed by the discriminator D in each batch.

[0110] Under the premise that the discriminator D has been pre-trained, if the discriminant output value of the enhanced sample signal is close to 0.5, it means that the distribution of the enhanced sample signal is close to the clean sample signal. The output of the discriminator D is back-propagated, that is, the output result is input into the discriminator D again.

[0111] (4.6) The output of the discriminator D is fed back to the generator G: According to the output of the discriminator D in step (4.5), the generator G is updated by boosting the stochastic gradient. The training loss function fed back from the output of the discriminator D to the generator G is for:

[0112]

[0113] in Denotes the loss function of the output of the discriminator D fed back to the generator G, ω D represents the weight parameter of the discriminator D, ω G represents the weight parameter of the generator G, M 4 is the number of training rounds for the discriminator;

[0114] (4.7) Enhanced sample signal x I and the clean sample signal x i Input to demodulator C: respectively enhance the sample signal x I and the clean sample signal x i Input demodulator C, demodulator C outputs the probability of each bit Back propagate the output of demodulator C and update demodulator C by improving its stochastic gradient. The training loss function of demodulator C is C (x,ω c ) is as follows:

[0115]

[0116] Among them, M 4 is the number of sample signals processed in each batch by the demodulator C, which is the same as the number of sample signals processed in each batch by the discriminator D in step (4.5); ω C represents the weight parameter of the demodulator C; N is the number of bits, j = 1, 2, ..., N, is the true label of the kth category of the i-th input sample signal at the j-th position, that is Indicates that the true label is 0, Indicates that the true label is 1.

[0117] The output of demodulator C is back-propagated, that is, the output result is input into demodulator C again.

[0118] (4.8) The output of the demodulator C is fed back to the generator G:

[0119] Feed the output of the demodulator C in step (4.7) back to the generator G for training; update the generator G by improving the stochastic gradient. The training loss function of the demodulator C is for:

[0120]

[0121] in represents the loss function of the demodulator C output fed back to the generator G; M 4 is the number of sample signals processed by the generator G in each batch, which is the same as the number of sample signals processed by the discriminator D in each batch in step (4.5); N is the number of bits, j = 1, 2, ..., N, is the true label of the kth category of the i-th input sample signal at the j-th position, that is Indicates that the true label is 0, Indicates that the true label is 1.

[0122] Return to step (4.1) and repeat steps (4.2)-(4.8) until the generator G, discriminator D, and demodulator C all converge, and the output is saved to obtain the generator G.

[0123] The training process of the generative adversarial network is as follows Figure 1As shown, the generative adversarial network (GAN) constructed by the present invention is different from the common generative adversarial network structure. The generative adversarial network of the present invention is composed of a generator, a discriminator and a demodulator. In the training stage, the generator is responsible for outputting an enhanced noise sample signal. The output enhanced noise sample signal and the clean sample signal are superimposed to obtain an enhanced sample signal, and the enhanced sample signal and the clean sample signal are respectively input into the discriminator and the demodulator. The outputs of the discriminator and the demodulator are then fed back to their respective inputs and the generator. The present invention constructs a three-person game GAN neural network structure to obtain a generator with better performance. Finally, the generator is combined with a deep receiver to enhance the defense effect of the deep receiver.

[0124] Step (5): The radio signal is input into the generator obtained in step (4) for processing to obtain a feature-enhanced electrical signal, and then the feature-enhanced electrical signal is input into a deep receiver to achieve a low bit error rate for decoding the radio signal.

[0125] In the present invention, the deep receiver uses the bit error rate (SER) as an evaluation index of defense performance. The bit error rate is an indicator to measure the accuracy of data transmission within a specified time. The lower the bit error rate, the higher the demodulation accuracy. For a single bit stream information, the bit error rate formula is calculated as follows:

[0126]

[0127] Where N represents the total number of bits in the bit stream of a piece of information, and Ne represents the number of erroneous bits in the bit stream of a piece of information. In this embodiment, N=32.

[0128] The illegal attacker sends a small-power counter-interference signal through the propagation channel and superimposes it on the normal radio signal of the transmitter to obtain a radio signal. The radio signal is different from the clean sample signal. If the radio signal is directly input into the deep receiver, there will be a high bit error rate. However, the radio signal is processed by the generator obtained in step (4) to obtain a feature-enhanced electrical signal, which is then input into the deep receiver to achieve a low bit error rate for decoding the radio signal.

Claims

1. A defense method for deep receivers based on feature enhancement, characterized in that The following steps are involved: Step (1): Build a dataset and train a deep receiver; The step (1) includes the following sub-steps: (1.1) Constructing a dataset: The constructed dataset includes a sample signal, which is divided into b categories of sample signals. C% of the sample signals are extracted from each category of sample signals as the training set. The training set participates in the subsequent model generation training. The sample signals are all clean sample signals, and a, b, and c are natural numbers; (1.2) Training the depth receiver: The deep receiver is composed of a dense convolutional neural network; the hyper parameters set during the training of the deep receiver include: the number of model training rounds of the deep receiver, the number of batch sample signals of the deep receiver, the gradient update rule of the deep receiver and the training loss function of the deep receiver; the number of model training rounds of the deep receiver is Epoch dense , the number of batch sample signals of the depth receiver is M1, the gradient update rule of the depth receiver is stochastic gradient descent SGD, and the training loss function of the depth receiver is Loss dense It is expressed as the following formula: Wherein, superscript i represents the i-th sample signal, i=1,2,…,M1, M1 is the number of sample signals in each batch, N is the number of bits, j=1,2,…,N, is the true label of the kth category of the i-th input sample signal at the j-th position, that is Indicates that the true label is 0, Indicates that the true label is 1, p j (x i ) represents the depth receiver's response to the sample signal x i The bit prediction probability of the jth bit of the output; Step (2): Construct a generative adversarial network, which includes a generator, a discriminator, and a demodulator; The step (2) includes the following sub-steps: (2.1) Build a generator: The generator consists of an encoder and a decoder, the encoder consists of a convolutional neural network and a long short-term memory neural network, and the decoder consists of a deconvolutional neural network; (2.2) Build a discriminator: The discriminator is composed of a convolutional neural network and a long short-term memory neural network; (2.3) Build a demodulator; The demodulator consists of a dense convolutional neural network; Step (3): Pre-train the generative adversarial network; The step (3) includes the following sub-steps: (3.1) Discriminator pre-training The discriminator is pre-trained. The training parameters set during the discriminator pre-training include: the number of pre-training rounds of the discriminator, the number of batch sample signals of the discriminator, the gradient update rule of the discriminator and the pre-training loss function of the discriminator D; the number of pre-training rounds of the discriminator is Epoch k , the number of batch sample signals of the discriminator is M2, the gradient update rule of the discriminator is stochastic gradient descent SGD, and the pre-training loss function of the discriminator D is Loss D (x,ω D ) is as follows: where x i is a clean sample signal, the superscript i represents the i-th sample signal, i = 1, 2, ..., M2, M2 is the number of sample signals in each batch; D(x i ;ω D ) indicates that the discriminator judges x i The probability of whether it is a clean sample signal, if it is greater than the threshold, it is considered a clean sample signal, and if it is less than the threshold, it is considered a non-clean sample signal; ω D represents the weight parameter of the discriminator; taking the log of the output probability D(·;·) is to convert the probability of event occurrence into information; (3.2) Demodulator Pre-training The demodulator is pre-trained. The training parameters set during the demodulator pre-training include: the number of pre-training rounds of the demodulator, the number of batch sample signals of the demodulator, the gradient update rule of the demodulator, and the pre-training loss function of the demodulator C; the number of pre-training rounds of the demodulator is Epoch k , the number of batch sample signals of the demodulator is M3, the gradient update rule of the demodulator is stochastic gradient descent SGD, and the pre-training loss function of the demodulator C is Loss C (x,ω c ) is as follows: where x i is a clean sample signal, superscript i represents the i-th clean sample signal, i = 1, 2, ..., M3, M3 is the number of sample signals in each batch; N is the number of bits, j = 1, 2, ..., N, C j (x i ;ω C ) represents the demodulator C's response to the sample signal x i The predicted probability of the j-th bit of the output is, represents the true label of the k-th category of the i-th input sample at the j-th position; ω C represents the weight parameter of the demodulator; Step (4): Train the generative adversarial network and finally obtain the generator; Step (5): The radio signal is input into the generator obtained in step (4) for processing to obtain a feature-enhanced electrical signal, and then the feature-enhanced electrical signal is input into the depth receiver to achieve a low bit error rate for decoding the radio signal.

2. The defense method of a deep receiver based on feature enhancement according to claim 1, characterized in that: The step (4) includes the following sub-steps: (4.1) Generate noise sample signal: Get random noise sample signal z from Gaussian distribution 1 ,z 2 ,…,z i ,…z M4 , z i The dimension and clean sample signal x i The dimension is the same, which is 2*n; (4.2) The noise sample signal z obtained in step (4.1) i and the clean sample signal x i Overlapping combination The dimension is 4*n, and the Input to the generator; (4.3) The generator generates an enhanced noise sample signal: the signal input to the generator After being processed by the encoder and decoder in the generator, the enhanced noise sample signal v is output i ; The enhanced noise sample signal v i and the clean sample signal x i The dimensions are the same; Where G(;) represents the generator, ω G Represents the weight parameters of the generator; (4.4) Enhance the noise sample signal v i and the clean sample signal x i Add to obtain enhanced sample signal x I : x I =x i +v i (4.5) Enhanced sample signal x I and the clean sample signal x i Input to the discriminator: respectively enhance the sample signal x I and the clean sample signal x i Input to the discriminator, the output of the discriminator is the probability output of whether the input sample signal is a clean sample signal, and the output interval of the discriminator ranges from [0,1]. By improving its stochastic gradient update, the loss function of the discriminator training is: Where D(·; ·) represents the probability of the discriminator to determine whether the input sample is a clean sample, and M4 is the number of training rounds of the discriminator; The output of the discriminator D is back-propagated, that is, the output result is input into the discriminator D again; (4.6) The discriminator output is fed back to the generator: According to the output of the discriminator in step (4.5), the generator is updated by improving the stochastic gradient; the training loss function fed back from the discriminator output to the generator for: in represents the loss function where the output of the discriminator is fed back to the generator, ω D represents the weight parameter of the discriminator, ω G Represents the weight parameters of the generator; (4.7) Enhanced sample signal x I and the clean sample signal x i Input to the demodulator: respectively enhance the sample signal x I and the clean sample signal x i Input demodulator, demodulator outputs the probability of each bit Backpropagate the output of the demodulator C and update the demodulator by improving its stochastic gradient; the training loss function of the demodulator Loss C (x,ω c ) is as follows: Where M4 is the number of sample signals processed in each batch by the demodulator, which is the same as the number of sample signals processed in each batch by the discriminator in step (4.5); ω C represents the weight parameter of the demodulator; N is the number of bits, j=1,2,…,N, is the true label of the kth category of the i-th input sample signal at the j-th position, that is Indicates that the true label is 0, Indicates that the true label is 1; The output of demodulator C is back-propagated, that is, the output result is input into demodulator C again; (4.8) The demodulator output is fed back to the generator: Feed the output of the demodulator in step (4.7) back to the generator for training; update the generator by improving the stochastic gradient, and the training loss function of the demodulator for: in represents the loss function of the demodulator output fed back to the generator; M4 is the number of sample signals processed in each batch by the generator G, which is the same as the number of sample signals processed in each batch by the discriminator D in step (4.5); N is the number of bits, j = 1, 2, ..., N, is the true label of the kth category of the i-th input sample signal at the j-th position, that is Indicates that the true label is 0, Indicates that the true label is 1; Return to step (4.1) and repeat steps (4.2)-(4.8) until the generator, discriminator, and demodulator converge, and save the output to obtain the generator.

Images