阅读说明：本技术 低截获波形生成方法 (Low-interception waveform generation method ) 是由 向雪霜 谢海东 谭佳 张晓莹 纪楠 陈远清 廖海华 刘乃金 于 2021-08-17 设计创作，主要内容包括：本发明涉及一种低截获波形生成方法,包括：S1.生成原始信号,获取调制识别模型；S2.发送每次扰动更新前的原始信号,根据调制识别模型对原始信号的识别效果,计算更新梯度；S3.根据梯度修改原始信号,获得更新后的期望信号；S4.判断期望信号是否收敛,收敛则继续执行以下步骤,否则循环迭代S2～S4,直至收敛；S5.对期望信号进行放大、离散化处理；S6.发送期望信号,由调制识别模型输出被识别的概率；S7.将概率与预设阈值B进行比较,当概率≤阈值时,生成目标低截获信号的波形,当概率＞阈值时,循环S2～S7,直至最终获得的概率≤阈值。本发明可有效防止通信信号被识别,进而防止通信系统中传输的信息被截获和窃听,增强通信系统的安全和可靠性。(The invention relates to a low-interception waveform generation method, which comprises the following steps: s1, generating an original signal to obtain a modulation recognition model; s2, sending an original signal before each disturbance updating, and calculating an updating gradient according to the identification effect of the modulation identification model on the original signal; s3, modifying the original signal according to the gradient to obtain an updated expected signal; s4, judging whether the expected signal is converged, if so, continuing to execute the following steps, otherwise, circularly iterating from S2 to S4 until the convergence is reached; s5, amplifying and discretizing the expected signal; s6, sending an expected signal, and outputting the recognized probability by a modulation recognition model; and S7, comparing the probability with a preset threshold B, generating a waveform of the target low interception signal when the probability is less than or equal to the threshold, and circulating S2-S7 when the probability is greater than the threshold until the finally obtained probability is less than or equal to the threshold. The invention can effectively prevent the communication signal from being identified, thereby preventing the information transmitted in the communication system from being intercepted and intercepted, and enhancing the safety and reliability of the communication system.)

1. A method of low-intercept waveform generation, comprising the steps of:

s1, generating an original signal to be sent, and acquiring a modulation identification model;

s2, sending the original signal before each disturbance updating, and calculating and updating a gradient according to the recognition effect of the modulation recognition model on the original signal;

s3, modifying the original signal according to the gradient to reduce the gradient and obtain an updated expected signal;

s4, judging whether the expected signal is converged, if the expected signal is converged, continuing to execute the following steps, otherwise, circularly iterating the steps from S2 to S4 until the finally obtained expected signal is converged;

s5, amplifying the expected signal, and performing discretization processing on the expected signal;

s6, sending the expected signal processed by the S5, and outputting and obtaining the recognized probability of the expected signal after being recognized by the modulation recognition model;

and S7, comparing the probability with a preset threshold, generating a waveform of a target low interception signal when the probability is less than or equal to the threshold, and circularly repeating the steps from S2 to S7 when the probability is greater than the threshold until the finally obtained probability is less than or equal to the threshold.

2. The method of claim 1, wherein said modulation recognition model input of said S2 is said raw signal, its output is a probability that said raw signal corresponds to a classification category, and a value of said gradient is calculated from a relationship of said input to said output.

3. The low-intercept waveform generation method of claim 2 wherein the relationship of said input to said output is represented as:

y＝f(w_i)，

label＝argmax(y)，

wherein i is more than or equal to 0, y is more than or equal to 0 and less than or equal to 1, and w_iThe original signal before each disturbance update is represented, y represents the probability of each classification category, and label represents the classification label with the highest corresponding probability in the classification categories to which the original signal is identified.

4. The method of generating low-interception waveforms of claim 3, wherein said gradient is calculated by the formula:

5. the low-interception waveform generating method according to claim 1, wherein said S3 comprises the steps of:

s31, initializing the disturbance direction of the original signal of S1, setting the disturbance direction to be 0 and the disturbance amplitude to be epsilon;

s32, carrying out normalization processing on the gradient, wherein the normalization formula is as follows:

s33, updating the disturbance direction, and carrying out disturbance updating according to the following relational expression:

wherein i is not less than 0 and delta_iIndicating the direction of the disturbance, δ, before each disturbance update_i+1Representing the disturbance direction after each disturbance update, and alpha represents the learning rate of each update;

s34, adjusting the disturbance amplitude according to the fact that whether the original signal is identified by the modulation identification model or not is judged, when the original signal is identified, adjusting the disturbance amplitude according to the following relation (a), and when the original signal is not identified, adjusting the disturbance amplitude according to the following relation (b):

ε_i+1＝(1+γ)ε_i (a)，

ε_i+1＝(1-γ)ε_i (b)，

wherein i is not less than 0, and gamma represents a shrinkage coefficient, epsilon_iRepresenting the amplitude of the disturbance, ε, before each disturbance update_i+1Representing the magnitude of the perturbation after each perturbation update;

s35, modifying and continuously updating the original signal according to the following relational expression to obtain the expected signal after each updating:

wherein, w_i+1Representing the desired signal after each update;

s36, checking the expected signal and ensuring that the amplitude range of the expected signal waveform meets the following relational expression:

w_u+1＝clip(w_i+1，-1，1)。

6. the method of claim 1, wherein said desired signal converges based on an approximation of two consecutive iterations of outputThe magnitude of the rate change being less than a threshold value, i.e. | y_i+1-y_i|＜A。

7. The method of claim 6, wherein the threshold is set to a value of 10 ═ 10^-6。

8. The low-intercept waveform generation method of claim 1 wherein said desired signal is amplified according to the following relationship:

w_i+1＝w_i+a·(w_i+1-w_i)，

where a represents a constant.

9. The method of low interception waveform generation according to claim 1, characterized in that said probability of the desired signal being identified is calculated by:

Technical Field

The invention relates to the technical field of wireless communication anti-interference, in particular to a low-interception waveform generation method.

Background

The communication system is very easy to be intercepted and intercepted due to the complex structure while generating important military and civil application values.

Thanks to the development of deep learning techniques, the strongest existing intelligent models can already identify the modulation format of a spectrum signal with extremely high accuracy. The technology has the characteristics of accuracy, high efficiency, strong generalization capability and low user deployment cost. However, for communication, the technology is developed into a double-edged sword, which brings excellent performance and provides convenience for those who use maliciously, so that the existing communication signals are exposed to severe risks of interception and eavesdropping.

The chinese patent CN112468258A discloses a full-duplex end-to-end automatic encoder communication system and an anti-eavesdropping method thereof, in a wireless channel environment with eavesdropping, a legal receiver sends a network model for resisting disturbance signals to attack an illegal eavesdropper, so that the eavesdropping capability of malicious nodes can be reduced. The prior art has not explored on the aspect of the communication signal itself to reduce the risk of eavesdropping and interception of the communication.

Disclosure of Invention

In order to prevent the signals of a communication system from being identified and further prevent transmitted information from being intercepted and intercepted, the invention provides a low-interception waveform generation method.

In order to achieve the purpose, the technical scheme of the invention is as follows:

the invention provides a low-interception waveform generation method, which comprises the following steps:

s1, generating an original signal to be sent, and acquiring a modulation identification model;

s2, sending the original signal before each disturbance updating, and calculating and updating a gradient according to the recognition effect of the modulation recognition model on the original signal;

s3, modifying the original signal according to the gradient to reduce the gradient and obtain an updated expected signal;

s4, judging whether the expected signal is converged, if the expected signal is converged, continuing to execute the following steps, otherwise, circularly iterating the steps from S2 to S4 until the finally obtained expected signal is converged;

s5, amplifying the expected signal, and performing discretization processing on the expected signal;

s6, sending the expected signal processed by the S5, and outputting and obtaining the recognized probability of the expected signal after being recognized by the modulation recognition model;

and S7, comparing the probability with a preset threshold, generating a waveform of a target low interception signal when the probability is less than or equal to the threshold, and circularly repeating the steps from S2 to S7 when the probability is greater than or equal to the threshold until the finally obtained probability is less than or equal to the threshold.

Further, the modulation recognition model of S2 has an input of the original signal and an output of the original signal being a probability that the original signal corresponds to a classification category, and calculates a value of the gradient according to a relationship between the input and the output.

Further, the relationship of the input to the output is represented as:

y＝f(w_i)，

label＝argmax(y)，

wherein i is more than or equal to 0, y is more than or equal to 0 and less than or equal to 1, and w_iThe original signal before each disturbance update is represented, y represents the probability of each classification category, and label represents the classification label with the highest corresponding probability in the classification categories to which the original signal is identified.

Further, the gradient is calculated by the formula:

further, the S3 includes the following steps:

s31, initializing the disturbance direction of the original signal of S1, setting the disturbance direction to be 0 and the disturbance amplitude to be epsilon;

s32, carrying out normalization processing on the gradient, wherein the normalization formula is as follows:

s33, updating the disturbance direction, and carrying out disturbance updating according to the following relational expression:

wherein i is not less than 0 and delta_iIndicating the direction of the disturbance, δ, before each disturbance update_i+1Representing the disturbance direction after each disturbance update, and alpha represents the learning rate of each update;

s34, adjusting the disturbance amplitude according to the fact that whether the original signal is identified by the modulation identification model or not is judged, when the original signal is identified, adjusting the disturbance amplitude according to the following relation (a), and when the original signal is not identified, adjusting the disturbance amplitude according to the following relation (b):

ε_i+1＝(1+γ)ε_i (a)，

ε_i+1＝(1-γ)ε_i (b)，

wherein i is not less than 0, and gamma represents a shrinkage coefficient, epsilon_iRepresenting the amplitude of the disturbance, ε, before each disturbance update_i+1Representing the magnitude of the perturbation after each perturbation update;

s35, modifying and continuously updating the original signal according to the following relational expression to obtain the expected signal after each updating:

wherein, w_i+1Representing the desired signal after each update;

s36, checking the expected signal and ensuring that the amplitude range of the expected signal waveform meets the following relational expression:

w_i+1＝clip(w_i+1，-1，1)。

further, the expected signal converges according to the fact that the probability change amplitude of the two adjacent iteration outputs is smaller than a threshold value, namely y_i+1-y_i|＜A。

Further, the value of the threshold is a-10^-6。

Further, the desired signal is amplified according to the following relation:

w_i+1＝w_i+a·(w_i+1-w_i)，

where a represents a constant.

Further, the calculation formula of the probability that the desired signal is identified is:

the invention has the beneficial effects that:

the invention combines the artificial intelligence countermeasure sample technology, and reduces the gradient of the signal after being identified by the modulation identification model by optimizing, modifying and updating the communication signal, thereby reducing the probability of the signal being identified by the modulation identification model, effectively preventing the transmitted information from being intercepted and intercepted, and simultaneously not influencing the normal communication process of the communication system.

The low interception waveform generating method has the advantages that the finally generated low interception waveform signal has remarkable capability of being not accurately identified, the low interception effect is good, and the confidentiality is strong. The method shows excellent low interception effect in the actual communication environment, and has important value for the application in the aspects of communication confidentiality, communication safety and the like.

The low interception waveform generation method enhances the safety and reliability of the communication system, can be self-adapted to various application scenes of wireless communication in the using process, and has good expanded application capability. In the civil field, the method can be used for mobile communication applications such as 6G in the future and the like, and the interception resistance safety of user communication information is improved; in the military field, the method is mainly used in the fighting environment, and can meet the requirements of camouflage, deception and interception resistance of communication signals when the situation that the intelligent means is used for monitoring and intercepting the communication is faced.

Drawings

FIG. 1 schematically illustrates a flow diagram of a low-intercept waveform generation method in accordance with an embodiment of the present invention;

fig. 2 is a diagram schematically illustrating the low interception effect of a signal in an ideal state according to the low interception waveform generation method of the embodiment of the present invention;

fig. 3 is a diagram schematically illustrating the low-interception effect of a signal in a simulation environment according to a low-interception waveform generating method of an embodiment of the present invention.

Detailed Description

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the embodiments will be briefly described below. It is obvious that the drawings in the following description are only some embodiments of the invention, and that for a person skilled in the art, other drawings can be derived from them without inventive effort.

The present invention is described in detail below with reference to the drawings and the specific embodiments, which are not repeated herein, but the embodiments of the present invention are not limited to the following embodiments.

Fig. 1 schematically illustrates a flow diagram of a low-intercept waveform generation method in accordance with an embodiment of the present invention. As shown in fig. 1, in the present embodiment, after the present invention is started, S1 is executed first, and the information to be transmitted is first sent to be transmittedModulating an analog signal to be transmitted in a signal channel to generate an original signal w to be transmitted₀(t) and obtaining the original signal w for identification₀(t) to intercept or eavesdrop the original signal w₀(t) a modulation identification model of the communicated information. When the analog signal is modulated, selectable modulation modes include amplitude modulation, frequency modulation, phase modulation and the like. The modulation recognition model is typically a deep learning based neural network model. The modulation recognition model adopted by the embodiment is a ResNet neural network model trained by data in a RADIOML 2018.01A data set, and has excellent recognition capability, and is marked as f (). Wherein, the data of the radar ioml 2018.01a data set has a corresponding data length of 1024 × 2, specifically, "1024" represents the time discrete length of the signal, and "2" represents 2 polarization directions of the signal.

Then, in the present embodiment, S2 is executed, and the specific process is: the original signal w generated in S1 is transmitted first₀(t) into the channel, then in the course of signal transmission, using the trained modulation recognition model to make said original signal w₀(t) identifying the original signal w according to a modulation identification model₀(t) calculating a gradient value from the recognition effect of (t). In this embodiment, the original signal w is obtained because the trained ResNet neural network model has excellent recognition capability₀(t) can be accurately identified.

Specifically, the present embodiment first inputs the original signal w into the modulation recognition model₀(t), thereafter, the modulation recognition model outputs a raw signal w₀(t) different probabilities corresponding to all possible belonging classification categories are identified. A gradient value is calculated from the relationship of the input and output. Wherein the input and the output satisfy the following relationship:

y＝f(w₀)，

label＝argmax(y)，

where 0. ltoreq. y.ltoreq.1, y denotes the original signal w₀(t) probability corresponding to each classification category, label representing original signal w₀(t) is identified as the classification label with the highest corresponding probability in the classification category to which it belongs. Here, the original signalw₀(t) the final classification label recognized by the modulated recognition model reflects the recognition effect in S2 described above. The gradient value is calculated by the formula:

the above formula for calculating the gradient value reflects the dependency between the output probability and the input signal. The larger the gradient value, the stronger the correlation between the two, that is, the larger the gradient value, the greater the probability that the input signal is accurately identified as the corresponding belonging classification category.

Therefore, it is necessary to perform S3 by modifying the original signal w₀(t) obtaining an updated desired signal w₁(t) reducing the gradient value after passing through the modulation recognition model to reduce the original signal w₀And (t) the probability of being identified, so that the probability of information interception and eavesdropping is reduced, and the reliability and the safety of communication are ensured. By modifying the original signal w in S3₀(t) the specific process of updating the gradient includes:

s31, initializing an original signal w₀(t) setting the disturbance direction to 0 and the disturbance amplitude to epsilon; in this embodiment, let ∈ 0.01;

s32, carrying out normalization processing on the gradient of S2, wherein the normalization formula is as follows:

s33, updating original signal w₀(t) and updating the disturbance according to the following relation:

wherein, delta₀Indicating the direction of the disturbance of the initialisation, i.e. delta₀＝0，δ₁Represents the perturbation direction after the first perturbation update, and alpha representsA learning rate;

s34, according to the judgment original signal w₀(t) whether or not to adjust the disturbance amplitude by the modulation recognition model when the original signal w is recognized₀(t) is identified according to the following relational expression (a)₁) Adjusting the amplitude of the disturbance when the original signal w₀(t) is not recognized, and is expressed by the following relational expression (b)₁) Adjusting the disturbance amplitude:

ε₁＝(1+γ)ε₀ (a₁)，

ε₁＝(1-γ)ε₀ (b₁)，

wherein γ represents a shrinkage coefficient, ε₀Indicating the magnitude of the disturbance of the initialisation, i.e. epsilon₀＝ε，ε₁Representing the disturbance amplitude after the first disturbance update;

s35, the original signal w is subjected to the following relational expression₀(t) performing a first modification to obtain a first updated desired signal:

wherein, w₁Representing the desired signal w after the first update₁(t)；

S36, checking the expected signal w₁(t) and ensuring the desired signal w₁(t) the amplitude range of the waveform satisfies the following relation:

w₁＝clip(w₁，-1，1)。

next, S4 is executed to determine the expected signal w outputted from S3₁(t) if the desired signal w converges₁(t) if the convergence condition is satisfied, S5-S7 is continuously executed, if the desired signal w₁And (t) if the convergence condition is not met, continuing to circularly iterate S2-S4 until the finally obtained expected signal meets the convergence condition. Wherein, the condition of convergence is that the loss change amplitude of the expected signal output by the modulation recognition model is smaller than a threshold value A, namely the change amplitude of the probability y output by two adjacent iterations is smaller than the threshold value A, namely | y_i+1-y_i|＜A。

When the desired signal w₁If (t) satisfies the convergence condition, S5-S7 are continuously executed. Wherein S5 is the desired signal w to be output in S4₁(t) procedure of amplification and discretization, in particular of the desired signal w₁(t) amplifying according to the following relation,

w₁＝w₀+a·(w₁-w₀)，

where a represents a constant. S6 is continued after S5, and the amplified and discretized desired signal w is₁(t) transmitting, outputting and obtaining the expected signal w after being identified by a modulation identification model₁(t) probability of being identified. The formula for calculating the probability is:

finally, S7 is executed to obtain the desired signal w₁(t) comparing the identified probability with a preset threshold, and generating the waveform of the target low interception signal when the probability is less than or equal to the threshold. And when the probability is larger than the threshold value, circularly repeating S2-S7 until the finally obtained probability is smaller than or equal to the threshold value.

When the desired signal w₁And (t) continuing the loop iteration from S2 to S4 when the convergence condition is not satisfied. In the process of each loop iteration, the expected signal output in the last iteration process is used as the original signal in the next iteration process, namely, the original signal before each perturbation update. Then, the present embodiment expects the signal w₁And (t) continuously performing loop iteration from S2 to S4 as the original signal before the second iteration and the disturbance updating. Similarly, in this embodiment, S2 represents the original signal w_i(t) inputting the modulation recognition model and then outputting the original signal w_i(t) identifying different probabilities for all possible said classification categories, and then calculating gradient values. Specifically, the input and output satisfy the following relationship:

y＝f(w_i)，

label＝argmax(y)，

wherein i is more than or equal to 0, y is more than or equal to 0 and less than or equal to 1, and w_iRepresenting the original signal before each update of the perturbation, y represents the original signal w_i(t) probability corresponding to each classification category, label representing original signal w_i(t) is identified as the classification label with the highest corresponding probability in the classification category to which it belongs. The gradient value is calculated by the formula:

in the present embodiment, S3 is obtained by modifying the original signal w_i(t) obtaining an updated desired signal w_i+1(t) reducing the gradient value after passing through the modulation recognition model to reduce the original signal w_iAnd (t) the probability of being identified, so that the probability of information interception and eavesdropping is reduced, and the reliability and the safety of communication are ensured. By modifying the original signal w in S3_i(t) the specific process of updating the gradient includes:

s32, carrying out normalization processing on the gradient obtained in the step S2, wherein a normalization formula is as follows:

s33, updating original signal w_i(t) and updating the disturbance according to the following relation:

wherein i is not less than 0 and delta_iIndicating the direction of the disturbance, δ, before each disturbance update_i+1Representing the disturbance direction after each disturbance update, and alpha represents the learning rate of each update;

s34, according to the judgment original signal w_i(t) whether or not to adjust the disturbance amplitude by the modulation recognition model when the original signal w is recognized_i(t) is identified and the amplitude of the disturbance is adjusted according to the following relation (a) when the original signal w_i(t) if not identified, adjusting the disturbance amplitude according to the following relation (b)：

ε_i+1＝(1+γ)ε_i (a)，

ε_i+1＝(1-γ)ε_i (b)，

Wherein i is not less than 0, and gamma represents a shrinkage coefficient, epsilon_iRepresenting the amplitude of the disturbance, ε, before each disturbance update_i+1Representing the magnitude of the perturbation after each perturbation update;

s35, modifying and continuously updating the original signal w according to the following relational expression_i(t), obtaining the expected signal after each update:

wherein, w_i+1Representing the desired signal w after each update_i+1(t)；

S36, checking the expected signal w_i+1(t) and ensuring the desired signal w_i+1(t) the amplitude range of the waveform satisfies the following relation:

w_i+1＝clip(w_i+1，-1，1)。

in this embodiment, the whole process of the S3 gradient update utilizes a typical countermeasure sample optimization technique in deep learning, specifically embodies the idea of the DDN countermeasure sample optimization technique, has a high intelligence level, and can adapt to various application scenarios. In addition, the present embodiment may use a countermeasure sample generation concept such as FGSM, PGD, depfol, CW, or the like.

In the present embodiment, S4 determines desired signal w_i+1(t) if the desired signal w converges_i+1(t) if the convergence condition is satisfied, S5-S7 is continuously executed, if the desired signal w_i+1(t) the convergence condition is not satisfied, the loop iteration S2-S4 is continued until the finally obtained desired signal satisfies the convergence condition. Preferably, in the present embodiment, the threshold a in the convergence condition determination is 10^-6. That is, when the probabilities obtained by two adjacent iterations are very close, and the probability change values of the two adjacent outputs tend to be closeTowards 10^-6The low interception performance of the desired signal obtained at this time has reached a limit, substantially optimal. Even if the probability change value of two adjacent outputs is made further small, i.e., less than 10^-6The above-mentioned low interception performance of the desired signal changes very little.

When the desired signal w_i+1(t) when the convergence condition is satisfied, S5 is executed to apply the desired signal w_i+1(t) the amplification is performed according to the following relational expression, and then the discretization process is performed.

w_i+1＝w_i+a·(w_i+1-w_i)，

Where a represents a constant. Continuing to S6, the amplified and discretized desired signal w is_i+1(t) transmitting, outputting and obtaining the expected signal w after being identified by a modulation identification model_i+1(t) probability of being identified. Finally, S7 is executed to obtain the desired signal w_i+1(t) comparing the identified probability with a preset threshold, and generating the waveform of the target low interception signal when the probability is less than or equal to the threshold. And when the probability is larger than the threshold value, circularly repeating S2-S7 until the finally obtained probability is smaller than or equal to the threshold value. When the threshold is set to different values, the low-interception waveform generation method of the embodiment can be applied to different scenes with interception prevention and interception prevention requirements.

Fig. 2 and 3 schematically show graphs of low-interception effect of a signal in an ideal state and in a simulated environment, respectively, according to a low-interception waveform generation method of an embodiment of the invention.

As shown in fig. 2, in an ideal state, the signal w is compared with the original signal w transmitted₀(t) compared with the prior art, no matter how the signal-to-noise ratio changes, the probability that the target low interception signal finally generated by the low interception waveform generation method of the embodiment is identified by the modulation identification model is 0, and at the moment, the low interception effect of the information transmitted in the communication system is the best, so that the information can be effectively prevented from being intercepted and intercepted.

As shown in fig. 3, in a complex channel environment of communication simulation, the accuracy of the target low-interception signal finally generated and output by the low-interception waveform generation method of the present embodiment is maintained below 40% by the modulation recognition model. Express mailWhen the noise ratio is 0, the accuracy rate is close to 40%; when the signal-to-noise ratio is below 0 or above 0, the accuracy is below 40%, but the signal is identified with a lower accuracy when the signal-to-noise ratio is below 0 than when the signal-to-noise ratio is above 0. Therefore, by modifying the original signal w₀And (t), the probability of interception and eavesdropping of the communication information can be effectively reduced by increasing the noise interference mode, and the safety and the reliability of the communication system are increased.

The above description is only one embodiment of the present invention, and is not intended to limit the present invention, and it is apparent to those skilled in the art that various modifications and variations can be made in the present invention. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.