CN112183352A - Communication interference method based on generation countermeasure network - Google Patents
Communication interference method based on generation countermeasure network Download PDFInfo
- Publication number
- CN112183352A CN112183352A CN202011045415.2A CN202011045415A CN112183352A CN 112183352 A CN112183352 A CN 112183352A CN 202011045415 A CN202011045415 A CN 202011045415A CN 112183352 A CN112183352 A CN 112183352A
- Authority
- CN
- China
- Prior art keywords
- interference
- network
- waveform
- training
- generated
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 230000006854 communication Effects 0.000 title claims abstract description 97
- 238000004891 communication Methods 0.000 title claims abstract description 96
- 238000000034 method Methods 0.000 title claims abstract description 56
- 238000012549 training Methods 0.000 claims abstract description 83
- 230000006870 function Effects 0.000 claims abstract description 71
- 230000008859 change Effects 0.000 claims abstract description 27
- 230000000694 effects Effects 0.000 claims abstract description 27
- 238000004519 manufacturing process Methods 0.000 claims abstract description 6
- 230000004913 activation Effects 0.000 claims description 24
- 230000008569 process Effects 0.000 claims description 19
- 230000001629 suppression Effects 0.000 claims description 6
- 238000005070 sampling Methods 0.000 claims description 5
- 230000015572 biosynthetic process Effects 0.000 claims description 3
- 230000009466 transformation Effects 0.000 claims description 2
- 238000010586 diagram Methods 0.000 description 29
- 238000002474 experimental method Methods 0.000 description 13
- 238000001228 spectrum Methods 0.000 description 13
- 238000009826 distribution Methods 0.000 description 11
- 238000013459 approach Methods 0.000 description 5
- 238000005516 engineering process Methods 0.000 description 5
- 230000002452 interceptive effect Effects 0.000 description 4
- 238000011478 gradient descent method Methods 0.000 description 3
- 238000012803 optimization experiment Methods 0.000 description 3
- 238000006243 chemical reaction Methods 0.000 description 2
- 230000000875 corresponding effect Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000004088 simulation Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 230000005540 biological transmission Effects 0.000 description 1
- 230000000903 blocking effect Effects 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 230000002596 correlated effect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000002068 genetic effect Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000010355 oscillation Effects 0.000 description 1
- 230000002688 persistence Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000002787 reinforcement Effects 0.000 description 1
- 230000001360 synchronised effect Effects 0.000 description 1
- 230000003313 weakening effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
- G06F2218/10—Feature extraction by analysing the shape of a waveform, e.g. extracting parameters relating to peaks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
- G06F2218/04—Denoising
Abstract
The invention discloses a communication interference method based on a generation countermeasure network. The method comprises the following steps: constructing a GAN network model; acquiring a target signal time domain waveform of a target communication system, and manufacturing a training set; training a GAN network model by using a training set of target signals, changing the learning rate, training for multiple times, observing the change trend of a loss function and a generated waveform under different learning rates, and fixing the learning rate when the network loss function tends to be stable within the set training times and can generate a generated waveform with the same time domain and frequency domain characteristics as the signals in the training set; training the network at the determined learning rate to generate an interference waveform with the same time domain and frequency domain characteristics as the signals in the training set; the generated interference waveform is used to apply interference to the target communication system. The method can autonomously extract the characteristics of the target signal and generate the interference waveform to implement interference, and has good application generalization while obtaining a good interference effect.
Description
Technical Field
The invention relates to the technical field of communication interference, in particular to a communication interference method based on a generation countermeasure network.
Background
The communication interference is an important research direction in the field of communication countermeasure, and on the basis of acquiring the prior information of a target signal through a communication reconnaissance means, a communication interference system suppresses and destroys the communication process of the target signal by generating an interference signal associated with the target signal, so as to finally achieve the purpose of weakening or even blocking the communication capability of the other party.
Generally speaking, the existing communication interference method depends on a communication reconnaissance means, and needs to select an interference strategy from a fixed interference strategy library or make an interference strategy by user experience on the basis of acquiring target signal prior information by the reconnaissance means, and select an interference waveform from a waveform library accumulated at ordinary times to apply effective interference to a target communication system. On the basis of obtaining the prior information of the target signal, many methods have been used to substitute for manually searching for the optimal interference strategy, such as genetic algorithm, information theory, game theory, etc. However, the method is implemented on the premise that the method has partial prior information about a communication party and an environment, and in an actual interference process, especially in the face of an unknown target signal with dynamically changed parameters generated by an intelligent communication system, the prior information is often difficult to acquire, and the methods are difficult to function.
On the premise of no prior information, a reinforcement learning algorithm is used for optimizing the interference strategy. The basic idea is that an interference strategy set is constructed by using an interference pattern, interference power and an interference persistence ratio, the feedback of a target system after being interfered is used as a reward value to adjust the interference strategy, and the optimal interference strategy is found through continuous trial and error, so that a better effect is achieved. However, this method cannot be used when the interfering party cannot obtain the feedback information of the target system.
Disclosure of Invention
The invention aims to provide a communication interference method based on a generation countermeasure network, which aims at carrying out interference on unknown target signals generated by an intelligent communication system under the condition that an interference party cannot obtain feedback of the target system.
The technical solution for realizing the purpose of the invention is as follows: a method for countering communication interference of a network based on generation, comprising the following steps:
step 4, training the network under the learning rate determined in the step 3, and generating an interference waveform with the same time domain and frequency domain characteristics as the signals in the training set;
and 5, applying the generated interference waveform to apply interference to the target communication system.
Further, the GAN network model is constructed in step 1, specifically as follows:
and (3) building the GAN network by adopting a full-connection layer activation function, wherein the generated network and the discrimination network only comprise one hidden layer, and the hidden layer and the output layer are realized by using the full-connection layer activation function.
Further, the modulation scheme of the communication system in step 2 includes BPSK, QPSK, 16QAM, and 2 FSK.
Further, the training set is manufactured in the step 2, and the specific process is as follows: each signal produced 128000 samples, with 10 symbols, 1000 sample points as one sample.
Further, the generated network input is one-dimensional random noise with the length of 100, and the random noise is converted into an interference waveform with the length of 1000 through two fully-connected layers to be output, wherein the activation function of the hidden layer is Relu, and the activation function of the output layer is Tanh.
Further, the structure of the discrimination network is opposite to that of the generation network, the input is target signal time domain sampling data with the length of 1000 or an interference waveform output by the generation network, a scalar is output after two times of conversion of the full connection layer, the scalar represents the probability that the input is the target signal, the activation function of the hidden layer is Relu, and the activation function of the output layer is Sigmoid.
Further, step 5 includes a step of comparing the interference effect with the best interference and the noise interference after applying the interference to the target communication system.
Further, the definition of the best interference is: for a given signal form and communication receiving mode, the interference pattern with the minimum suppression coefficient is required, and the suppression coefficient refers to the ratio of the interference power and the signal power required by the communication receiving end.
Further, the noise interference refers to gaussian noise, that is, the target communication system is directly interfered by the gaussian noise.
Compared with the prior art, the invention has the following remarkable advantages: (1) the method comprises the following steps of utilizing the ability that the GAN can automatically learn the potential distribution of data in a training set and generate samples with the same distribution, directly taking the acquired time domain waveform of a target signal as the training set to train the GAN, taking the generated samples with similar time domain characteristics and frequency domain characteristics to the target signal as interference waveforms, and implementing interference aiming at the channel of a target communication system: (2) the method has the advantages that prior knowledge about the characteristics of the target signal is not needed, the characteristics of the target signal can be automatically extracted, interference waveforms are generated to implement interference, the process of manual decision making is reduced, the problem of signal interference of unknown signals and dynamic parameter changes in the actual interference process is solved, and good application generalization is achieved while a good interference effect is achieved.
Drawings
FIG. 1 is a schematic diagram of the basic structure of GAN.
Fig. 2 is a basic structure diagram of a generation countermeasure network constructed by the present invention, in which (a) is a structure diagram of the generation network and (b) is a structure diagram of the discrimination network.
Fig. 3 is a flow chart of a method of the present invention based on generating interference to communications across a network.
Fig. 4 is a graph showing the change in the loss function when the data set is BPSK and the learning rates are 0.0001 and 0.001, where (a) is a graph showing the change in the loss function for discriminating the network and (b) is a graph showing the change in the loss function for generating the network.
Fig. 5 is a graph showing the change in the loss function when the data set is BPSK and the learning rate is 0.01, where (a) is a graph showing the change in the loss function for discriminating a network and (b) is a graph showing the change in the loss function for generating a network.
Fig. 6 is a time domain waveform diagram and a frequency spectrum diagram of samples generated when the data set is BPSK and different learning rates are set, wherein (a) the time domain waveform diagram of the samples generated when the data set is BPSK and the learning rates are 0.0001, 0.001 and 0.01 from top to bottom, and (b) the frequency spectrum diagram of the samples generated when the data set is BPSK and the learning rates are 0.0001, 0.001 and 0.01 from top to bottom.
Fig. 7 is a graph showing the variation of the loss function of QPSK signal, wherein (a) is a graph showing the variation of the loss function of the discrimination network, and (b) is a graph showing the variation of the loss function of the generation network.
Fig. 8 is a graph of the change of the loss function of the FSK signal of 2, wherein (a) is a graph of the change of the loss function of the discrimination network and (b) is a graph of the change of the loss function of the generation network.
Fig. 9 is a graph of the loss function change of a 16QAM signal, wherein (a) is a graph of the loss function change of a discrimination network and (b) is a graph of the loss function change of a generation network.
Fig. 10 is a time domain waveform and a frequency spectrum diagram for generating QPSK samples, where (a) is the time domain waveform diagram for generating QPSK samples, and (b) is the frequency spectrum diagram for generating QPSK samples.
Fig. 11 is a time domain waveform diagram and a spectrogram of a generated 2FSK sample, where (a) is the time domain waveform diagram of the generated 2FSK sample and (b) is the spectrogram of the generated 2FSK sample.
Fig. 12 is a time domain waveform diagram and a spectrum diagram for generating a 16QAM sample, where (a) is the time domain waveform diagram for generating the 16QAM sample, and (b) is the spectrum diagram for generating the 16QAM sample.
Fig. 13 is a graph of the error rates of BPSK, QPSK, 2FSK, 16QAM communication systems under three interferences, where (a) is a graph of the error rates of the BPSK communication systems under three interferences, (b) is a graph of the error rates of the QPSK communication systems under three interferences, (c) is a graph of the error rates of the 2FSK communication systems under three interferences, and (d) is a graph of the error rates of the 16QAM communication systems under three interferences.
Detailed Description
The existing communication interference method generally carries out interference decision based on target signal characteristics obtained in communication reconnaissance, selects a proper interference waveform to carry out interference, and is difficult to deal with the condition that the target signal characteristics are unknown or parameters dynamically change. Therefore, the invention provides a communication interference waveform generation technology based on a generation countermeasure network (GAN), which utilizes the GAN to autonomously learn the acquired target signal time domain waveform, acquire the potential distribution of the target signal, generate an interference waveform approaching the target signal distribution, and apply interference to a target communication system. On the basis of introducing the GAN principle, firstly, aiming at the characteristics of a communication signal, a network model is designed, and the learning rate is optimized, so that when training data are a communication signal time domain waveform, the network can be rapidly converged, and an interference waveform highly associated with a target signal is generated; secondly, carrying out interference waveform generation experiments on communication signals of different types and parameters to verify the generalization of the technology; finally, a performance comparison test is carried out with the best interference and noise interference effect.
Under the influence of game theory, the generation countermeasure network is composed of two networks playing mutually, the generation network tries to generate data approaching to the distribution of real samples to cheat the discrimination network, the discrimination network tries to correctly judge the source of input, and in the mutual game, the generation capability and the discrimination capability of the two networks are gradually improved. The final result of the game is that the two reach Nash equilibrium, the generated network learns the distribution of real samples in a training set, and a vivid sample cheating discrimination network can be generated. In communication interference, under the condition that no target signal prior information exists, starting from a physical layer, selecting a waveform with the same characteristics as a target signal as an interference waveform, applying interference to a target system channel, and disturbing a threshold judgment process of a target communication system demodulator, wherein the threshold judgment process is a feasible interference approach, and the higher the correlation degree between the interference waveform and the target signal waveform is, the better the interference effect is.
1 Generation of GAN-based communication interference waveforms
There are three main possible ways to interfere with digital communication systems, namely, channel interference, synchronous system interference, and transmission information interference. The interference to the channel is the interference applied to the characteristics of the demodulator, and the interference signal is required to be similar to the target signal in the signal level, namely the interference signal and the target signal are required to have characteristics of a similar time domain, a similar frequency domain and a similar modulation domain, so that the interference signal and the communication signal can enter the input end of the demodulator together, and the threshold judgment process of the demodulator is disturbed, so that the judgment is wrong, and the error rate is increased. The interference waveform generated by the invention based on the GAN is used for interfering aiming at the channel.
1.1 generating a Confrontation network
The GAN is composed of two competing networks, a generation network G and a discrimination network D, and its basic structure is shown in fig. 1. The input of G is a random noise vector z, the goal of which is to generate pseudo data G (z) whose distribution approximates the true sample distribution as closely as possible to deceive D; the input of D is the pseudo data G (z) generated by the real samples x or G in the training set, with the goal of judging the source of the input as correctly as possible. The output of D is a scalar quantity used to represent the probability of determining that the input is a true sample, and when D considers that the input is a true sample, 1 is output, and when D considers that the input is dummy data G (z) generated by G, 0 is output.
D and G can be any differentiable function, and in the actual training process, the random gradient descent method is adopted to optimize D and G alternately. When D is optimized, the parameters of G are fixed, and since the goal of D is to implement binary classification of input data, that is, when input is real data x, 1 is output, and when input is dummy data G (z), 0 is output, its loss function can be described as:
d (x) in the first term in equation (1) represents the probability that the D decision input is a true sample when the input is a true sample x, and D (G (z)) in the second term represents the probability that the D decision input is a true sample when the input is pseudo data generated by G. The process of optimizing D is the process of adjusting the network parameters of D by using a random gradient descent method so as to minimize J (D).
Similarly, when G is optimized, the parameters of D need to be fixed. The goal of G is to fool D by the generated G (z), i.e., D (G (z)) is expected to approach 1, the smaller the second term of j (D) the better, so the loss function of G can be described as:
therefore, the process of optimizing G is a process of adjusting network parameters of G by using a random gradient descent method so as to minimize j (G). Also, since there is no need to use real samples when optimizing G, the overall GAN objective function can be described as:
therefore, as can be seen from equation (3), the GAN optimization process is the process of G and D playing the infinitesimal game. During training, firstly fixing the parameters of G, and training D by taking pseudo data G (z) generated by real samples x and G as a training set to maximize an objective function V (D, G), namely enabling D to have certain discrimination capacity; and fixing the parameters of D, optimizing G, and minimizing the objective function V (D, G), namely improving the generation capability of G. The above processes are performed alternately, and finally, when the generation capacity of G is optimized to generate false data which is false and true to cheat D, D cannot judge the input source, and the output approaches a certain fixed value, G is considered to have mastered the distribution of real samples in the training set, and a 'vivid' sample can be generated.
1.2 network model design
In the experiment of the invention, the time domain waveform of a target signal is directly sampled, 1000 points are taken as one sample, and one-dimensional time sequence data with the length of 1000 are obtained and are used as real data x to be input into a discrimination network. Because the data processed by the method are one-dimensional time sequence data, the structure is simple, and compared with a two-dimensional image, the communication signal has fewer characteristics and is easier to understand, so that a convolution layer and a deconvolution layer are not adopted to extract the distribution of the characteristics and fitting data when a network is built, but the characteristics and the fitting data are directly realized by adding an activation function by a full connection layer, and the basic structure for generating the countermeasure network built by the method is shown in fig. 2. The generation network and the discrimination network only contain one hidden layer, and the hidden layer and the output layer are both realized by a full connection layer and an activation function.
As shown in fig. 2(a) is a schematic structural diagram of a generation network, an input is one-dimensional random noise with a length of 100, and the random noise is converted into an interference waveform with a length of 1000 through two fully-connected layers and output, wherein an activation function of a hidden layer is Relu, and an activation function of an output layer is Tanh. Fig. 2(b) is a schematic structural diagram of a decision network, the structure of the decision network is opposite to that of a generation network, the input is target signal time domain sampling data with a length of 1000 or an interference waveform output by the generation network, and after two times of transformation of a full connection layer, a scalar is output to represent the probability that the decision input is a target signal, wherein an activation function of a hidden layer is Relu, and an activation function of an output layer is Sigmoid. 1.3 communication interference waveform Generation
And constructing a GAN network and a communication system with BPSK, QPSK, 16QAM and 2FSK modulation modes, acquiring communication signals to make a training set, and training the GAN. In the training process, considering that the loss function can only reflect the change of the network generating capacity and the discrimination capacity and can not intuitively and accurately measure the relevance of the generated interference waveform and the time domain waveform of the target signal in the training set, the method performs Fourier transform on the generated waveform, observes and analyzes the frequency domain characteristic of the generated waveform, and is used as a standard for measuring the quality of the generated waveform together with the time domain characteristic and the loss function. The learning rate is then optimized to enable the network to converge quickly, generating an interference waveform that is highly correlated with the target signal. And finally, applying the generated interference waveform to a target communication system to apply interference, and comparing the interference effect with the interference effect of optimal interference and noise interference to verify the effectiveness of the communication interference waveform generation technology. The main algorithm architecture of the present invention is shown in fig. 3.
The invention relates to a communication interference method based on a generation countermeasure network, which comprises the following steps:
step 4, training the network under the learning rate determined in the step 3, and generating an interference waveform with the same time domain and frequency domain characteristics as the signals in the training set;
and 5, applying the generated interference waveform to apply interference to the target communication system.
Further, the GAN network model is constructed in step 1, specifically as follows:
and (3) building the GAN network by adopting a full-connection layer activation function, wherein the generated network and the discrimination network only comprise one hidden layer, and the hidden layer and the output layer are realized by using the full-connection layer activation function.
Further, the modulation scheme of the communication system in step 2 includes BPSK, QPSK, 16QAM, and 2 FSK.
Further, the training set is manufactured in the step 2, and the specific process is as follows: each signal produced 128000 samples, with 10 symbols, 1000 sample points as one sample.
Further, the generated network input is one-dimensional random noise with the length of 100, and the random noise is converted into an interference waveform with the length of 1000 through two fully-connected layers to be output, wherein the activation function of the hidden layer is Relu, and the activation function of the output layer is Tanh.
Further, the structure of the discrimination network is opposite to that of the generation network, the input is target signal time domain sampling data with the length of 1000 or an interference waveform output by the generation network, a scalar is output after two times of conversion of the full connection layer, the scalar represents the probability that the input is the target signal, the activation function of the hidden layer is Relu, and the activation function of the output layer is Sigmoid.
Further, step 5 includes a step of comparing the interference effect with the best interference and the noise interference after applying the interference to the target communication system.
Further, the definition of the best interference is: for a given signal form and communication receiving mode, the interference pattern with the minimum suppression coefficient is required, and the suppression coefficient refers to the ratio of the interference power and the signal power required by the communication receiving end.
For BPSK, QPSK, and 2FSK communication systems, the optimal interference is a communication signal having the same modulation scheme, carrier frequency, and symbol rate, so in experiments, a randomly generated baseband signal is subjected to different modulation to be used as an optimal interference waveform for BPSK, QPSK, and 2FSK communication systems. For a 16QAM communication system, theoretically, the best interference is a QPSK signal having the same carrier frequency and symbol rate as the best interference, but the effect of interfering with a 16QAM signal having the same carrier frequency and symbol rate as the target signal in the simulation is slightly different from the best interference effect, and when the interference ratio is low, the effect of interfering with the 16QAM signal is slightly better than that of the QPSK signal, so the interference effect of the 16QAM signal having the same carrier frequency and symbol rate as the target signal on the 16QAM communication system is still used as the best interference in the experiment of the present invention.
Further, the noise interference refers to gaussian noise, that is, the target communication system is directly interfered by the gaussian noise.
The invention is described in further detail below with reference to the figures and the embodiments.
The embodiment is based on a method for generating communication interference for a countermeasure network, and the steps are as follows:
And 2, respectively training the GAN by using training sets of several signals, stopping training when the variation trend of the loss function tends to be stable and the generated waveform has time domain and frequency domain characteristics similar to those of the signals in the training sets, and outputting the generated waveform.
And 3, changing the learning rate, training for multiple times, and selecting the optimal learning rate to realize the generation of the communication interference waveform.
And 4, applying the generated interference waveform to apply interference to the target communication system, and comparing the interference effect with the interference effect of the optimal interference and noise interference.
In order to verify the feasibility of the interference waveform generation technology, the optimization experiment of the learning rate, the application generalization experiment of the network and the interference effect comparison experiment are carried out.
Hardware device environment: window 10 system, NVIDIA GeForce GTX 1660Ti graphic card and TensorFlow1.14.0 framework.
Example 1 learning Rate optimization experiment
At present, GAN is widely used in the field of computer vision for processing two-dimensional picture data with complex features, whereas GAN is required to process one-dimensional time sequence data with simpler features than two-dimensional pictures. The variation of data dimensions and features causes the required learning rate to vary. Therefore, firstly, taking BPSK signals as an example, an optimization experiment of the learning rate is performed, and a proper learning rate is selected for generating communication signals.
Experiments were conducted with learning rates set to 0.0001 and 0.001, respectively, from 1000 to 90000 times, and the loss function change curves are shown in fig. 4(a) to (b). As can be seen from fig. 4, since the two networks are opposing each other, the loss functions of the two networks are in violent oscillation regardless of the general trend of change. Since the speed of increasing the generation capacity of the network is faster and better when the learning rate is 0.001 than when the learning rate is 0.0001, it is attempted to observe the change tendency of the loss function by increasing the learning rate.
When the learning rate was 0.01, the number of times of training was 1000 to 90000 times, and the loss function change curve was as shown in fig. 5. Considering that the variation trend of the loss function is greatly different when the learning rate is 0.01, the network is trained for many times, and a more typical three times is plotted as fig. 5. As can be seen from fig. 5(a), the loss function of the discrimination network rapidly approaches 0 after a short fluctuation, the gradient cannot be provided for updating the generation network, the gradient of the generation network disappears, and the generation capability cannot be improved any more. This indicates that the learning rate is too large to cause the parameters to oscillate back and forth in the optimal solution, and thus convergence is not possible.
In general, from the trend of the loss functions of the two networks, the generation capability of the network is improved quickly when the learning rate is 0.001, and the network can be converged quickly after training. And then, the quality of the interference waveform generated by the network is measured more intuitively by further analyzing time domain oscillograms and frequency spectrogram of samples generated by the network at different learning rates. Fig. 6 is a time domain waveform diagram and a frequency spectrum diagram of samples generated when the data set is BPSK and the learning rates are different, fig. 6(a) is a time domain waveform diagram of samples generated when the learning rates are 0.0001, 0.001, and 0.01 in sequence from top to bottom, the left side is a sample generated after 2000 times of training at the learning rate, the right side is a sample generated after 90000 times of training, and the corresponding position in fig. 6(b) is a frequency spectrum diagram thereof. It can be seen from the figure that when the learning rate is 0.001, the GAN network has the fastest training speed and the strongest generation capability, the samples generated by the network after 2000 times of training have a frequency domain characteristic similar to the target signal, and the samples generated after 90000 times of training have a higher degree of similarity to the time domain characteristic and the frequency domain characteristic of the target signal. When the learning rate is 0.01, although the samples generated after 2000 times of training have relatively similar frequency domain characteristics with the target signals, the generation capability of the generated network is not obviously improved, and the similarity between the samples generated after 90000 times of training and the target signals is still equivalent to that after 2000 times of training. When the learning rate is 0.0001, the speed of improving the generating capacity of the generating network is obviously slower than that of the learning rates 0.001 and 0.01, the generated sample still has no obvious characteristics after 2000 times of training, but the generating capacity is improved continuously along with the increase of the training times, the generated sample has time domain characteristics and frequency domain characteristics which are similar to the target signal after 90000 times of training, but the similarity is greatly different from that of the learning rate 0.001, which indicates that when the learning rate is 0.0001, the generating capacity of the generating network can be improved continuously along with the increase of the training times, but the improving speed is slower. In general, when the learning rate is 0.001, the network generation capability is improved quickly, and an interference waveform having similar time domain characteristics and frequency domain characteristics to the target signal can be generated quickly.
Example 2 application generalization experiments on networks
The traditional interference method selects a proper interference waveform to carry out interference based on the characteristics of a detected target signal, and the GAN-based communication interference waveform generation technology provided by the invention directly uses GAN to learn the time domain waveform of the acquired target signal, acquires the potential characteristics of the target signal, and generates a waveform with similar characteristics as an interference waveform, does not need prior knowledge about the target characteristics, and has stronger generalization compared with the traditional interference method. The following experiment researches the application effect of the GAN on different target communication signals, and verifies the generalization of the GAN.
In experiment 1, a BPSK signal time domain waveform is sampled to prepare a data set, and a network is trained. In order to show that the GAN has a good generating effect on other types of communication signals, QPSK, 2FSK and 16QAM signal time domain waveforms are taken to manufacture a data set, and the network is trained.
The learning rate was set to 0.001, the batch size was 256, the network was trained with QPSK signal data set, the number of times of training was from 1000 to 90000 times, the loss function change curves were as shown in fig. 7(a) to (b), the network was trained with 2FSK signal data set, the number of times of training was from 1000 to 90000 times, the loss function change curves were as shown in fig. 8(a) to (b), the network was trained with 16QAM signal data set, the number of times of training was from 1000 to 90000 times, and the loss function change curves were as shown in fig. 9(a) to (b). As can be seen from fig. 7, 8, and 9, the loss functions of the generation network and the discrimination network both change greatly in the initial stage of training, and after 40000 times of training, the loss functions of the generation network and the discrimination network both stabilize, and the network can converge quickly. For example, 10 symbols are used, a time domain waveform diagram and a frequency spectrum diagram of QPSK, 2FSK, 16QAM signal samples generated after 90000 times of training are shown in fig. 10, 11 and 12. Fig. 10(a) is a time domain waveform diagram for generating QPSK samples, and it can be seen that the network learns the characteristics of the phase inversion of the QPSK signals, and the phase inversion occurs at the 2, 4, 6, 8 symbol boundaries in fig. 10 (a). Fig. 10(b) is a spectrum diagram of QPSK sample generation, which is consistent with the carrier frequency of the QPSK signal used for training, and shows that the network learns the spectrum characteristics of the QPSK signal. As can be seen from fig. 11, after 2000 times of training, the generating network can generate an interference waveform containing two frequency components, and the generated sample spectrum characteristics are consistent with those of the 2FSK signal used for training, but at this time, the waveforms of the two frequencies are uniformly mixed together in the time domain, and the network is not sufficient for learning the time domain characteristics of the 2FSK signal. After 90000 times of training, the network gradually learned the time domain characteristics of the 2FSK signal, generating an interference waveform in which the two frequency waveforms do not appear at the same time. Fig. 12(a) is a time domain waveform diagram for generating 16QAM samples, and it can be seen that, every 4 symbols, the phase or amplitude of the waveform changes, and the network learns the characteristics of 16QAM phase inversion and amplitude change. Fig. 12(b) is a spectrum diagram of a generated 16QAM sample, corresponding to the carrier frequency of the 16QAM signal used for training.
Example 3 comparative experiment of interference Effect
The final purpose of the generation of the communication interference waveform is to realize effective interference, and in order to illustrate the effectiveness and the superiority of the interference waveform generation method compared with the traditional method, the interference effect comparison test is carried out on the method, the optimal interference and Gaussian noise interference.
In the experiment, in order to sufficiently observe the interference effect, 10000 symbol lengths are taken as an example, the interference waveform generated after training is carried out for 80000 times, the optimal interference waveform and the Gaussian noise respectively interfere the target communication system, the interference is repeated for 100 times under different interference-to-signal ratios, the average error rate is taken, and an interference effect comparison graph is drawn. In the case of BPSK and QPSK communication systems, the optimal interference is a communication signal having the same modulation scheme, carrier frequency, and symbol rate as the BPSK and QPSK communication systems, and therefore, in experiments, a baseband signal generated at random is subjected to different modulation and then is used as an optimal interference waveform for the BPSK and QPSK communication systems. The interference effect pair is shown in fig. 13. Fig. 13(a) is the error rate of a BPSK communication system under three kinds of interference, fig. 13(b) is the error rate of a QPSK communication system under three kinds of interference, fig. 13(c) is the error rate of a 2FSK communication system under three kinds of interference, and fig. 13(d) is the error rate of a 16QAM communication system under three kinds of interference. As can be seen from fig. 13, when the interference-to-signal ratio is the same, the error rate of the generated interference waveform is much higher than that of gaussian noise, that is, the interference effect is much more significant than that of gaussian noise interference. When the interference-to-signal ratio is small, the interference effect of generating the interference waveform is better than that of the optimal interference, which also verifies from one side that the so-called optimal interference waveform is also within a certain interference-to-signal ratio range; when the interference-signal ratio is larger, the error rate of the generated interference waveform is lower than that of the optimal interference, but with the increase of the interference-signal ratio, the generated waveform error rate curve can gradually approach to that of the optimal interference. According to the combination of the variation trend of the loss functions of the four signals and the simulation result of the generated sample, when the error rate is 0.001, the network is trained by using the training set of the four signals, after the training times are 40000 times, the variation trend of the loss functions of the network can be gradually reduced, the loss value tends to be stable, the generated network can generate a sample which is close to the distribution of the real sample to cheat the discriminant network, and a good interference effect is generated on a target communication system.
Claims (9)
1. A method for generating interference to communication of a countermeasure network, comprising the steps of:
step 1, constructing a GAN network model;
step 2, acquiring a target signal time domain waveform of a target communication system, and manufacturing a training set;
step 3, training the GAN network model by using a training set of target signals, changing the learning rate, training for multiple times, observing the change trend of the loss function and the generated waveform under different learning rates, and fixing the learning rate when the network loss function tends to be stable within the set training times and can generate the generated waveform with the same time domain and frequency domain characteristics as the signals in the training set;
step 4, training the network under the learning rate determined in the step 3, and generating an interference waveform with the same time domain and frequency domain characteristics as the signals in the training set;
and 5, applying the generated interference waveform to apply interference to the target communication system.
2. The method for generating communication interference based on countermeasure network according to claim 1, wherein the GAN network model is constructed in step 1 as follows:
and (3) building the GAN network by adopting a full-connection layer activation function, wherein the generated network and the discrimination network only comprise one hidden layer, and the hidden layer and the output layer are realized by using the full-connection layer activation function.
3. The method according to claim 1, wherein the modulation scheme of the communication system in step 2 includes BPSK, QPSK, 16QAM, 2 FSK.
4. The method for generating communication interference based on the countermeasure network according to claim 1, wherein the training set is created in step 2 by the following specific processes: each signal produced 128000 samples, with 10 symbols, 1000 sample points as one sample.
5. The method for generating communication interference for the countermeasure network according to claim 2, wherein the generated network input is one-dimensional random noise with a length of 100, and is converted into an interference waveform output with a length of 1000 through two fully-connected layers, wherein an activation function of the hidden layer is Relu, and an activation function of the output layer is Tanh.
6. The communication interference method based on the generation countermeasure network of claim 2, characterized in that the structure of the discrimination network is opposite to that of the generation network, the input is time domain sampling data of a target signal with the length of 1000 or an interference waveform of the generation network output, and after two times of transformation of a full connection layer, a scalar is output to represent the probability that the discrimination input is the target signal, wherein the activation function of the hidden layer is Relu, and the activation function of the output layer is Sigmoid.
7. The method for generating communication interference for countermeasure network according to any of claims 1 to 6, wherein the step 5 further comprises a step of comparing the interference effect with the best interference and the noise interference after applying interference to the target communication system.
8. The method of claim 7, wherein the definition of the best interference is: for a given signal form and communication receiving mode, the interference pattern with the minimum suppression coefficient is required, and the suppression coefficient refers to the ratio of the interference power and the signal power required by the communication receiving end.
9. The method of claim 7, wherein the noise interference refers to gaussian noise, that is, the gaussian noise is directly applied to interfere with the target communication system.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011045415.2A CN112183352A (en) | 2020-09-29 | 2020-09-29 | Communication interference method based on generation countermeasure network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011045415.2A CN112183352A (en) | 2020-09-29 | 2020-09-29 | Communication interference method based on generation countermeasure network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN112183352A true CN112183352A (en) | 2021-01-05 |
Family
ID=73946460
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011045415.2A Pending CN112183352A (en) | 2020-09-29 | 2020-09-29 | Communication interference method based on generation countermeasure network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112183352A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113055107A (en) * | 2021-02-23 | 2021-06-29 | 电子科技大学 | Interference strategy generation method for radio station with unknown communication mode |
| CN113114408A (en) * | 2021-04-06 | 2021-07-13 | 电子科技大学 | Method and device for generating and evaluating generated deception jamming signal |
| CN113726350A (en) * | 2021-08-09 | 2021-11-30 | 哈尔滨工程大学 | Deep neural network-based strong correlation self-interference cancellation method |
| CN114513278A (en) * | 2022-04-18 | 2022-05-17 | 南京邮电大学 | Intelligent interference method, device and system based on electromagnetic spectrum characteristic cognition |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107220600A (en) * | 2017-05-17 | 2017-09-29 | 清华大学深圳研究生院 | A kind of Picture Generation Method and generation confrontation network based on deep learning |
| CN109286466A (en) * | 2018-11-19 | 2019-01-29 | 浙江嘉科电子有限公司 | A kind of portable integrated frequency interferometer of novel on-vehicle |
| CN109998500A (en) * | 2019-04-30 | 2019-07-12 | 陕西师范大学 | A kind of pulse signal generation method and system based on production confrontation network |
| CN110988818A (en) * | 2019-12-09 | 2020-04-10 | 西安电子科技大学 | Cheating interference template generation method for countermeasure network based on condition generation formula |
-
2020
- 2020-09-29 CN CN202011045415.2A patent/CN112183352A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107220600A (en) * | 2017-05-17 | 2017-09-29 | 清华大学深圳研究生院 | A kind of Picture Generation Method and generation confrontation network based on deep learning |
| CN109286466A (en) * | 2018-11-19 | 2019-01-29 | 浙江嘉科电子有限公司 | A kind of portable integrated frequency interferometer of novel on-vehicle |
| CN109998500A (en) * | 2019-04-30 | 2019-07-12 | 陕西师范大学 | A kind of pulse signal generation method and system based on production confrontation network |
| CN110988818A (en) * | 2019-12-09 | 2020-04-10 | 西安电子科技大学 | Cheating interference template generation method for countermeasure network based on condition generation formula |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113055107A (en) * | 2021-02-23 | 2021-06-29 | 电子科技大学 | Interference strategy generation method for radio station with unknown communication mode |
| CN113114408A (en) * | 2021-04-06 | 2021-07-13 | 电子科技大学 | Method and device for generating and evaluating generated deception jamming signal |
| CN113726350A (en) * | 2021-08-09 | 2021-11-30 | 哈尔滨工程大学 | Deep neural network-based strong correlation self-interference cancellation method |
| CN114513278A (en) * | 2022-04-18 | 2022-05-17 | 南京邮电大学 | Intelligent interference method, device and system based on electromagnetic spectrum characteristic cognition |
| CN114513278B (en) * | 2022-04-18 | 2022-08-05 | 南京邮电大学 | Intelligent interference method, device and system based on electromagnetic spectrum characteristic cognition |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112183352A (en) | Communication interference method based on generation countermeasure network | |
| CN109274621B (en) | Communication protocol signal identification method based on depth residual error network | |
| CN111314257B (en) | Modulation mode identification method based on complex value neural network | |
| CN110086737A (en) | A kind of recognition methods of the modulation mode of communication signal based on figure neural network | |
| JP4270934B2 (en) | Method for detecting a plurality of symbols transmitted by or towards a plurality of users | |
| Lin et al. | DL-CFAR: A novel CFAR target detection method based on deep learning | |
| Ali et al. | An M-QAM signal modulation recognition algorithm in AWGN channel | |
| CN114745233B (en) | Joint channel estimation method and device based on pilot frequency design | |
| Ji et al. | Data-limited modulation classification with a CVAE-enhanced learning model | |
| Ding et al. | Online regularization of complex-valued neural networks for structure optimization in wireless-communication channel prediction | |
| Mahmutoglu et al. | Particle swarm optimization algorithm based decision feedback equalizer for underwater acoustic communication | |
| An et al. | Robust automatic modulation classification in low signal to noise ratio | |
| Varadan et al. | Design of piecewise maps for chaotic spread-spectrum communications using genetic programming | |
| US20020126773A1 (en) | Simplified method of detection by spheres when there is a low signal to noise ratio | |
| Zhou et al. | Wireless signal enhancement based on generative adversarial networks | |
| Parmar et al. | Dual-stream CNN-BiLSTM model with attention layer for automatic modulation classification | |
| CN108900454A (en) | A kind of modulation system parameter blind detection method and device based on machine learning | |
| Yıldırım et al. | Deep receiver design for multi-carrier waveforms using cnns | |
| CN113905383B (en) | IFF signal identification method, device and medium based on radio frequency fingerprint | |
| CN115913849A (en) | Electromagnetic signal identification method based on one-dimensional complex value residual error network | |
| CN113242201B (en) | Wireless signal enhanced demodulation method and system based on generation classification network | |
| Chakraborty et al. | Domain knowledge aided neural network for wireless channel estimation | |
| Zhang et al. | HpGAN: Sequence search with generative adversarial networks | |
| CN114785649B (en) | Satellite communication signal identification method based on multiport neural network | |
| Liu et al. | Automatic radar waveform recognition based on neural network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20210105 |
|
| RJ01 | Rejection of invention patent application after publication |