CN117478474A - Channel precompensation-based antagonistic sample signal waveform generation method - Google Patents
Channel precompensation-based antagonistic sample signal waveform generation method Download PDFInfo
- Publication number
- CN117478474A CN117478474A CN202311341844.8A CN202311341844A CN117478474A CN 117478474 A CN117478474 A CN 117478474A CN 202311341844 A CN202311341844 A CN 202311341844A CN 117478474 A CN117478474 A CN 117478474A
- Authority
- CN
- China
- Prior art keywords
- signal
- training
- modulation
- network
- channel
- 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.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 57
- 230000003042 antagnostic effect Effects 0.000 title claims description 5
- 230000002708 enhancing effect Effects 0.000 claims abstract description 4
- 238000012549 training Methods 0.000 claims description 60
- 230000006870 function Effects 0.000 claims description 35
- 238000000605 extraction Methods 0.000 claims description 30
- 238000012795 verification Methods 0.000 claims description 22
- 238000004422 calculation algorithm Methods 0.000 claims description 20
- 239000013598 vector Substances 0.000 claims description 16
- 230000004913 activation Effects 0.000 claims description 10
- 230000005540 biological transmission Effects 0.000 claims description 10
- 238000012360 testing method Methods 0.000 claims description 9
- 230000000694 effects Effects 0.000 claims description 7
- 238000005457 optimization Methods 0.000 claims description 6
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 4
- 238000004364 calculation method Methods 0.000 claims description 4
- 238000011176 pooling Methods 0.000 claims description 4
- 230000009466 transformation Effects 0.000 claims description 4
- 230000009286 beneficial effect Effects 0.000 claims description 2
- 238000010276 construction Methods 0.000 claims description 2
- 239000011159 matrix material Substances 0.000 claims description 2
- 238000010606 normalization Methods 0.000 claims description 2
- 230000007704 transition Effects 0.000 claims description 2
- 230000006854 communication Effects 0.000 abstract description 36
- 238000004891 communication Methods 0.000 abstract description 32
- 238000013135 deep learning Methods 0.000 abstract description 14
- 238000005562 fading Methods 0.000 abstract description 9
- 238000012937 correction Methods 0.000 abstract description 3
- 230000008569 process Effects 0.000 description 5
- 238000013528 artificial neural network Methods 0.000 description 3
- 238000007796 conventional method Methods 0.000 description 3
- 238000013136 deep learning model Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000013473 artificial intelligence Methods 0.000 description 2
- 238000013527 convolutional neural network Methods 0.000 description 2
- 230000007123 defense Effects 0.000 description 2
- 235000000332 black box Nutrition 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000007635 classification algorithm Methods 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000011478 gradient descent method Methods 0.000 description 1
- 230000005484 gravity Effects 0.000 description 1
- 230000000116 mitigating effect Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 230000008054 signal transmission Effects 0.000 description 1
- 239000013589 supplement Substances 0.000 description 1
- 230000002194 synthesizing effect Effects 0.000 description 1
- 238000009827 uniform distribution Methods 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L27/00—Modulated-carrier systems
- H04L27/26—Systems using multi-frequency codes
- H04L27/2601—Multicarrier modulation systems
- H04L27/2626—Arrangements specific to the transmitter only
- H04L27/2627—Modulators
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B17/00—Monitoring; Testing
- H04B17/30—Monitoring; Testing of propagation channels
- H04B17/391—Modelling the propagation channel
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L27/00—Modulated-carrier systems
- H04L27/0014—Carrier regulation
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L27/00—Modulated-carrier systems
- H04L27/32—Carrier systems characterised by combinations of two or more of the types covered by groups H04L27/02, H04L27/10, H04L27/18 or H04L27/26
- H04L27/34—Amplitude- and phase-modulated carrier systems, e.g. quadrature-amplitude modulated carrier systems
- H04L27/36—Modulator circuits; Transmitter circuits
- H04L27/362—Modulation using more than one carrier, e.g. with quadrature carriers, separately amplitude modulated
- H04L27/364—Arrangements for overcoming imperfections in the modulator, e.g. quadrature error or unbalanced I and Q levels
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04W—WIRELESS COMMUNICATION NETWORKS
- H04W12/00—Security arrangements; Authentication; Protecting privacy or anonymity
- H04W12/12—Detection or prevention of fraud
- H04W12/121—Wireless intrusion detection systems [WIDS]; Wireless intrusion prevention systems [WIPS]
- H04W12/122—Counter-measures against attacks; Protection against rogue devices
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L27/00—Modulated-carrier systems
- H04L27/0014—Carrier regulation
- H04L2027/0024—Carrier regulation at the receiver end
- H04L2027/0026—Correction of carrier offset
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D30/00—Reducing energy consumption in communication networks
- Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks
Landscapes
- Engineering & Computer Science (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Physics & Mathematics (AREA)
- Electromagnetism (AREA)
- Computer Security & Cryptography (AREA)
- Digital Transmission Methods That Use Modulated Carrier Waves (AREA)
Abstract
The invention belongs to the technical field of wireless communication, and particularly relates to a channel precompensation-based anti-sample signal waveform generation method. The invention adds the correction to the multipath, the fading and the Doppler frequency of the channel in the generation of the countermeasure sample, thereby making the influence of the generated countermeasure disturbance to the multipath, the fading and the Doppler frequency of the channel more robust, and enhancing the attack performance of the countermeasure sample waveform to the modulation identifier based on the deep learning in the real wireless channel.
Description
Technical Field
The invention belongs to the technical field of wireless communication, and particularly relates to a channel precompensation-based method for generating an antagonistic sample signal waveform.
Background
For wireless communication systems, ensuring the security of their links is critical. The method for improving the communication safety comprises the following steps: 1. encrypting the transmitted data at the bit stream layer; 2. minimizing mutual information by means of physical layer security to avoid demodulating the communicated content by an eavesdropper; 3. by using the embedded communication technology, the communication signal is hidden in the signals of other systems, so that the existence of the signal detected by an eavesdropper is avoided. However, the above method may not provide guaranteed use of communication security in any scenario due to being too complex (for example, for an internet of things device, a complex encryption algorithm cannot be used due to limitation of computing resources). To further increase the security of the wireless communication link, other techniques may be used to supplement the encryption to prevent eavesdroppers from eavesdropping on the encrypted communication.
Eavesdroppers typically need to effect eavesdropping on the wireless communication link through three steps: 1) Detecting whether a signal exists or not by scanning a specific frequency band; 2) Capturing the signal by extracting the characteristics of the signal; 3) The signal is demodulated using the extracted features and a binary data stream is obtained. Breaking any of the above steps may enhance the security of the communication link. For example, encryption is focused on protecting the demodulated bit stream, while physical layer security is aimed at the third step, namely minimizing the information available to eavesdroppers. With the development of artificial intelligence technology represented by deep learning, deep learning has been widely used for detecting and recognizing communication signals, and has better adaptivity than conventional communication signal detection and recognition methods, and can detect and recognize communication signals in more complex electromagnetic environments. However, this also provides a more efficient means of eavesdropping for an illegal eavesdropper.
Through the study of the deep learning, the accuracy of the output result of the deep learning model can be greatly reduced by adding fine disturbance in the input signal of the deep learning model to generate an countermeasure sample. Since the added disturbance energy is usually weak (the energy is a fraction of the input signal), the original structure of the input signal is not destroyed, and the communication process of the communication system is not influenced. Several efforts have been proposed to improve the security of communication links using challenge-sample technology. Document one (Sadeghi, meysam, and Erik g.larsson. Universal tests on deep-learning based radio signal classification.ieee Wireless Communications Letters 8.1.8.1 (2018): 213-216.) proposes white-box and black-box resistance attacks on deep learning based modulation classifiers. This work shows that the deep learning-based wireless signal classification algorithm is extremely susceptible to resistance attacks; document two (s.kokalj-Filipovic, R.Miller, N.Chang, and c.l.lau, mitigation of adversarial examples in rf deep classifiers utilizing auto-encoder pretraining, in 2019International Conference on Military Communications and Information Systems (ICMCIS). IEEE,2019, pp.1-6.) developed an automatic encoder, and a receiver for preprocessing the modified signal. The authors disclose that pre-training a deep learning based classifier using an automatic encoder in the radio frequency domain can mitigate the deceptive effects of the resistant example; furthermore, document three (Ke, da, et al application of adversarial examples in communication modulation classification.2019international Conference on Data Mining Workshops (ICDMW) & IEEE, 2019.) considers the impact of challenge samples on the radio frequency domain and demonstrates that challenge defenses can improve the robustness of a deep learning based modulation classifier; in literature IV (Hamed, muhammad Zaid, andrasand Deniz Gündüz.The best defense is a good offense:Adversarial attacks to avoid modulation deIeee Transactions on Information Forensics and Security 16 (2020): 1074-1087.), the authors propose a method of counterattack that reduces the modulation classification accuracy of intruders while maintaining a low bit error rate for legitimate receivers; in literature five (Lin, yun, et al, university attacks in modulation recognition with convolutional neural networks, ieee Transactions on Reliability 70.1.1 (2020): 389-401.), authors verify the effectiveness of various resistance attacks by reconstructing waveforms in a modulated classification scene.
However, the signal added with the disturbance resistance is transmitted through the wireless channel, and the original disturbance resistance structure is destroyed after the effect of multipath, fading and the like of the wireless channel due to weak disturbance resistance energy. This is challenging for existing resistive perturbations to work inside an eavesdropper's system and prevent it from identifying the modulation type. To solve this problem, factors of the wireless channel must be considered in designing the countermeasure against the disturbance.
Disclosure of Invention
The invention introduces a correction term of a wireless channel by considering the influence of multipath and fading of the wireless channel when the anti-attack is designed, and provides an anti-sample signal waveform generation method based on channel pre-compensation. The invention adds the correction to the multipath, the fading and the Doppler frequency of the channel in the generation of the countermeasure sample, thereby making the influence of the generated countermeasure disturbance to the multipath, the fading and the Doppler frequency of the channel more robust, and enhancing the attack performance of the countermeasure sample waveform to the modulation identifier based on the deep learning in the real wireless channel.
The invention adopts the technical scheme that the method for generating the waveform of the countermeasure sample signal based on channel precompensation comprises the following steps:
s1 generating a transmitted signal data set
Encoding and modulating the transmitted information to generate a transmission signal s i (k) K=1, 2,..k, K represents the kth point of the transmitted signal, K represents a total of K points of the transmitted signal, K is typically 1024, 2048, 4096, etc.; i=1, 2,..n, i represents the i-th signal, N represents the total number of signals. In the process of converting the transmitted information into a transmission signal, different modulation modes can be adopted. It is generally necessary to generate signals of multiple modulation schemes for training a modulation identification network. When M modulation modes are available (namely M types of modulation signals to be classified), M is more than or equal to 2, and each transmitted signal s i (k) The corresponding modulation type is denoted by the numeral l i E {1,2,..m } is labeled, signal s is transmitted i (k) Together with the tags, constitute a dataset D. Dividing the data set D into training sets D train And verification set D valid For subsequent training of the modulation identification network.
General training set D train The number of signals contained in the set should be greater than the verification set D valid The number of signals contained in the training set ensures that the signals contained in the training set have universality.
S2 construction modulation recognition network
S2.1, constructing a feature extraction module, wherein the feature extraction module consists of 4 convolution modules, 1 two-dimensional average pooling layer Avg_pool2D and 1 flat layer in series; the first convolution module consists of 1 two-dimensional convolution layer Conv2D, wherein the number of convolution kernels in the convolution layer is filter=64, and the size of the convolution kernels is kernel_size=1×3; the second to fourth convolution modules are composed of four two-dimensional convolution layers Conv2D, the number of convolution kernels of each convolution module is 64, 128, 512 and 256 in sequence, and the convolution kernel size kernel_size=3×3; the convolution kernel stepping stride=1 of each two-dimensional convolution layer Conv2D, and the zero Padding strategy is padding=1; the output of each two-dimensional convolution layer Conv2D is normalized by the Batch normal, so that the output of the two-dimensional convolution layer Conv2D is ensured not to be distributed in a saturation region of a next layer activation function, and gradient dissipation is avoided when a network is trained; the output of the two-dimensional convolution layer Conv2D after the Batch normal normalization is continuously input into a ReLU activation function, wherein the ReLU activation function is a nonlinear transformation used for enhancing the feature extraction capability of the feature extraction module. Compared with other activation functions (such as sigmoid function, tanh function and the like), the ReLU activation function can overcome the condition that the gradient disappears, and is beneficial to faster convergence of the network; the output of the last convolution module adopts a two-dimensional average pooling layer Avg_pool2D with the size of 1 multiplied by 3 to reduce the dimension of the characteristics obtained by the convolution module so as to reduce the operand and prevent the network from being excessively fitted; the flat layer is used as a transition from the whole feature extraction module to a full connection layer in the identification module and is used for one-dimensional serialization of the multidimensional features output by the feature extraction module; the input data dimension of the whole feature extraction module is 1 multiplied by 2 multiplied by 1024, and the output feature dimension is 1 multiplied by 16384;
s2.2, constructing an identification module, wherein the identification module consists of two full-connection layers, the input characteristic dimension is 1 multiplied by 16384, and the output characteristic dimension is 1 multiplied by M; the output of the identification module is the identification probability of which type the signal belongs to, and the highest probability is the modulation signal type identified by the network.
The feature extraction module is connected with the recognition module in series to obtain a modulation recognition network, the weight parameters of the modulation recognition network are initialized randomly, and the modulation recognition network is trained by selecting a proper optimization algorithm (the common optimization algorithm comprises an Adam algorithm, an SGD algorithm or an RMSprop algorithm and the like) in a network training link.
S3 training modulation recognition network
S3.1, initializing training configuration: initializing a learning rate lr, an upper limit of training times epoch, and an upper limit of training waiting times parameter, wherein the learning rate lr is generally set to 10 -6 ～10 -1 The epoch is generally set to be 100-10000, and the upper limit of training waiting times is smaller than the epoch and is generally set to be one fifth or one tenth of the epoch; initializing training waiting times p=0, and initially verifying loss
S3.2 feature extraction
S3.2.1 extracting artificial characteristic of signal of training set Dtrain in S1, and extracting IQ characteristic x of signal i (k) After obtaining the characteristic with strong separability, the characteristic is sent to modulationThe characteristic extraction module of the identification network: for signal s i (k) The complex space is mapped by Hilbert transformation, then the real part and the imaginary part of the complex signal are respectively extracted and spliced to obtain the IQ characteristic x of the signal i (k)：
Wherein hilbert (·) is the hilbert transform.
S3.2.2 training set D in S1 train Is used for artificially extracting IQ characteristic x i (k) After that, x is i (k) The method is input into a feature extraction module of a modulation recognition network to automatically extract more complex features, and the feature extraction module outputs features Fx with the dimension of 1X 16384 i (k)]F represents a feature extraction module.
S3.3, calculating the identification result, and obtaining the characteristic Fx of S3.2.2 i (k)]Input recognition module for outputting recognition probability vector with dimension Mp i (m) represents the probability that the i-th signal belongs to the m-th modulated signal; taking the class with the largest probability vector as a modulation recognition result I i I.e. +.>Ii∈{1,2,...,M}。
S3.4 calculating training loss, and adding I i And data label l i Comparing to obtain a vector y i ：
Define the loss function L, p i (m) and y i Calculation of training loss L by feed loss function train The method comprises the steps of carrying out a first treatment on the surface of the The loss function L typically selects the cross entropy function (cross entropy):
s3.5 optimizing the network weight, selecting a proper optimization algorithm (the common optimization algorithm comprises an Adam algorithm, an SGD algorithm or an RMSprop algorithm and the like) to optimize the network weight, so that training in S3.4 is lost L train Descending, so as to obtain an optimized modulation recognition network;
s3.6 calculating verification loss, and integrating the verification set D valid Inputting the optimized modulation recognition network, and calculating a modulation recognition result I according to the steps of S3.2-S3.3 i Then adopting the loss function L to calculate the verification lossWherein t represents the training times;
s3.7, saving the optimal network weight: judging verification lossWhether or not to descend, if->The current network weight is saved, and the steps of S3.2 to S3.6 are repeated until the training times reach the upper limit; if verify loss->And if the continuous non-descending times exceed the upper limit of the training waiting times, stopping training. And saving the network weight which minimizes the verification loss as the optimal network weight. The strategy of judging whether the training needs to be terminated in advance by monitoring the verification loss is called an early stop (early stop) strategy, so that the model overfitting can be avoided, and the generalization capability of the modulation recognition network is improved.
S4 adding tiny anti-disturbance in the transmitted signal
S4.1, loading the optimal network weight stored in the S3.7;
s4.2 constructing input features, and obtaining IQ features x of the emission signals according to the manner of manually extracting the features in S3.2.1 i (k)。
S4.3 modulation recognition, the IQ characteristic x of the transmitted signal i (k) Calculating the recognition probability direction according to the steps S3.2-S3.3Quantity P i And modulation recognition result I i 。
S4.4, calculating the disturbance resistance by adopting a channel precompensation method, and obtaining the recognition probability vector P by S4.3 i And modulation recognition result I i And data label l i Calculation of test loss L using loss function test Then calculate the test loss L test Gradient vector v of IQ characteristic xi (k) of signal to be transmitted x L：
The traditional method directly carries out gradient vector x L is added as an anti-disturbance to the signal to be transmitted to constitute an anti-sample signal. However, after the antipodal sample signal is transmitted over the wireless channel, the added antipodal disturbance will change with the signal, losing its performance. Thus, channel precompensation is required for the calculated disturbance cancellation.
The wireless channel is a complex system, and the signal can be subjected to effects such as Doppler shift caused by multipath caused by atmospheric attenuation and reflection of a building and relative movement between a transmitting end and a receiving end in the transmission process. Modeling wireless channels is also a complex process. The wireless channel model commonly used at present comprises a Clarke model, a Jakes model and the like. Where H (·) represents the radio channel function, a signal s is transmitted i (k) After transmission over a wireless channel, H(s) i (k) Delta) of the added challenge disturbance i (k)＝▽ x L, the transmission of the anti-disturbance through the wireless channel is then H (delta i (k) A kind of electronic device. According to the equation (5) vs- x And L performs channel precompensation to obtain a corrected anti-disturbance signal:
wherein sign (·) is a sign function, and the superscript H denotes the conjugate transpose, J H Is a jacobian matrix of the channel function H (),for the channel transmitted signal H (s i (k) IQ characteristics H (x) i (k) The gradient vector obtained through the steps S4.2 to S4.4.
S4.5 iterative optimization challenge sample
The disturbance resisting signal is generally calculated repeatedlyThe attack performance can be improved. The upper limit item_num of the iteration number is preset at first and is generally set to 10-100. Then will counter the disturbance signal->Multiplied by a perturbation factor sigma and added to the IQ signature x of the transmitted signal i (k) In (1) obtaining an challenge sample->
The disturbance coefficient sigma is set according to the following principle: so thatThe energy of (2) is less than or equal to x i (k) One percent of energy;
s4.6 an anti-sample to which a minute anti-disturbance signal is addedInputting the result into a modulation recognition network for recognition to obtain a recognition result of the countermeasure sample.
Further, the invention can feed back according to the recognition result of the challenge sample, namely repeat the steps S4.2-S4.7, so as to improve the attack effect of the challenge sample, and make the challenge sample more lifelike.
The invention is based on the following principle: the challenge sample technique is a method of constructing special input data using vulnerabilities of the deep learning algorithm, thereby spoofing the deep learning model. The inventionModifications to the communication signals are made using the anti-sample technique such that an eavesdropper cannot correctly identify the signals using the deep learning algorithm. Since the own communication system does not require modulation of the identification signal and the added anti-disturbance signalIs less than or equal to the original signal IQ characteristic x i (k) One percent of the constructed challenge sample +.>IQ features x compared to the original signal i (k) Only minor modifications are made so that the reception and interpretation of the signal by the own receiver is hardly affected.
The invention has the following technical effects:
the invention adds the tiny anti-disturbance signal into the transmitted signal, so that the transmitted signal is difficult to judge the modulation type adopted by the signal by adopting the modulation identification method based on the deep neural network after being intercepted by a non-cooperative eavesdropper, thereby organizing the eavesdropper to interpret the communication content and improving the safety and reliability of the communication. In the process of generating anti-disturbance signals, the method provided by the invention pre-compensates the influence of multipath, fading and the like in the wireless channel, improves the robustness of the anti-disturbance channel, can lighten the influence of multipath and fading when the anti-disturbance signals pass through the channel, and further improves the false recognition rate of an eavesdropper on own communication signals under the real wireless channel condition;
drawings
FIG. 1 is a flow chart of an implementation of the present invention;
FIG. 2 is a time domain waveform of a modulated signal;
FIG. 3 is a diagram of a designed modulation identification network architecture and detailed parameters;
FIG. 4 is a graph comparing the rate of fraud versus PSR for the challenge sample and the conventional method of the present invention;
FIG. 5 is a plot of the spoofing rate versus PSR for the challenge samples generated by the present invention when the channels have different Doppler frequency offsets;
fig. 6 is a plot of the spoofing rate versus PSR for the challenge samples generated by the present invention when the channel has different multipath numbers.
Detailed Description
The following describes the embodiments of the present invention further with reference to the drawings.
The invention provides a signal countermeasure sample waveform generation method based on channel precompensation, which can add tiny countermeasure disturbance in a communication signal, and prevent an illegal eavesdropper from identifying a modulation mode of the signal by using an advanced artificial intelligence method on the premise of not damaging a communication process of the eavesdropper, thereby avoiding the eavesdropper from demodulating communication contents based on the identified modulation mode.
Fig. 1 is a flowchart of an implementation of the present invention, and the present invention provides a channel precompensation-based method for generating an antagonistic sample signal waveform, which includes the following steps:
s1, generating a transmission signal data set;
s2, constructing a modulation identification network;
s3, training a modulation recognition network;
s4 adds a small countermeasure disturbance to the transmitted signal.
The method comprises the following steps:
first, the invention specifically produces 10 differently modulated signals s i (k) The total number of signals is N=96000, the length K of each signal is 1024, and the modulation mode of each signal is marked with a label l i E {1, 2..10 }. To ensure the diversity of signal conditions, the signal-to-noise ratio of the signal covers 10dB to 20dB, and the code rate is 0.1Mbps. Then according to 7: the ratio of 3 divides 96000 signals into a training set and a validation set;
then, artificial feature extraction is performed for each signal, where IQ features are extracted. Each signal length is 1024, so that the dimension of the IQ characteristic obtained after the artificial characteristic extraction is 2×1024, and the input dimension of the characteristic extraction module of the modulation recognition network is 2×1024. The identification module is redesigned, and since there are 10 types of modulation in total, the output dimension of the identification module is 10. And connecting the two modules in series and randomly initializing the weight of the network to form an initial modulation identification network.
The modulation identification network is then trained. The training parameters are first initialized, here specifically the learning rate lr=10 -6 Upper limit of training period epoch=200, upper limit of training waiting period parameter=20, initial training waiting period p=0, initial verification lossAnd then inputting the constructed training set into a constructed modulation recognition network, and sequentially carrying out feature extraction and modulation recognition. Comparing the calculated modulation recognition result with the data label, and calculating the training loss L train The method comprises the steps of carrying out a first treatment on the surface of the The Adam algorithm is used to update the weights of the network. And inputting the signals of the verification set into the modulation recognition network after updating the weight, calculating the verification loss of the current training times, judging whether the verification loss is reduced, repeating the process if the verification loss is reduced until the continuous non-reduction times of the verification loss exceed the parameters or the training times exceed the upper limit epoch of the training times, stopping training, and storing the optimal network weight.
And finally, when signal transmission is required, loading the optimal modulation recognition network weight, then manually extracting the characteristics of the transmitted signal to obtain the IQ characteristics of the transmitted signal, inputting the IQ characteristics of the transmitted signal into the optimal modulation recognition network, and calculating a modulation recognition result. An upper limit iter_num=10 for the number of iterations is preset. And calculating gradient vectors according to the modulation recognition result and performing channel precompensation. The invention takes Jakes model as an example to realize modeling of the wireless channel.
The channel response of the Jakes model can be described as h (t):
wherein N is a multipath number, N 0 = (N/2-1)/2 is the number of sine waves required for synthesizing the channel model. θ n ＝2πn/N，n＝1,2,…,N 0 。w n ＝w d cosθ n ，w d ＝2πf m 。φ n Is the initial phase of the nth Doppler-shifted sine wave signal, phi N Is subjected to the maximum Doppler shift f m Is a primary phase of a sinusoidal signal. The setting of the initial phase must make the phase of the fading channel obey uniform distribution, and the initial phase can be set as follows:
the compensated gradient vector is multiplied by a small perturbation coefficient σ=0.001 to obtain the counterperturbation. Adding the challenge disturbance to the transmitted signal results in a signal challenge sample. And after a plurality of iterations until reaching the upper limit item_num of the iteration times, outputting a final signal waveform countermeasure sample. The transmitted communication signal is difficult to judge the modulation mode of the signal by an eavesdropper by using a deep learning method.
Fig. 2 is a time domain waveform of a modulated signal. Through step S1, 10 kinds of modulation are obtained, the total number of signals is 96000, the signal to noise ratio is 10dB to 20dB, and the code rate of the signals is 0.1Mbps.
FIG. 3 shows the basic structure and network parameters of the modulation and identification network used in the present invention, with input dimensions of [2,1024 ]]Where 2 represents the input as two components of the signal, quadrature and in-phase, and 1024 represents that each signal contains 1024 point samples. The output dimension is 10, indicating a total of 10 classes of modulation to be identified. The network has 17 layers of convolutional neural networks and 2 layers of fully-connected neural networks. The iteration times are set to be 200 times, the loss function adopts a cross entropy loss function, and the learning rate is 10 -5 The network was trained using a batch gradient descent method with a number of samples of 64 per batch. Adopts an early stop (early stop) training strategy, namely, if the verification loss of training exceeds a certain iteration number and does not drop, the training is stopped in advance, the strategy canTo avoid overfitting of the neural network.
Fig. 4 is a graph comparing the fraud rate versus PSR for the proposed algorithm and the existing mainstream challenge sample algorithm of the present invention. The invention adds disturbance-resistant to the communication signal to prevent the eavesdropper from identifying own communication modulation mode on the basis of not damaging the communication process of the eavesdropper as much as possible. Thus, the two most important indicators evaluating the new method proposed by the present invention are the Fraud Rate (FR) and the disturbance-to-signal power ratio (PSR). Wherein FR is defined as:
wherein the method comprises the steps ofTo indicate a function, if the formula in brackets is true, the value is 1, otherwise the value is 0.f (·) represents a deep learning based modulation identifier for an eavesdropper. X is x i To modulate the input features of the identifier delta i To combat disturbances. n is the number of samples that are correctly identified by the modulation identifier. To sum up, FR means: sample x, which was originally correctly identified i Adding disturbance delta i The number of samples that are later misidentified is the specific gravity of the total number of samples. The larger the FR, the more resistant the anti-sample algorithm is to the eavesdropper's ability to recognize the own communication signal.
PSR is defined as:
wherein P is δ Representing the power against disturbance, P x Representing the power of the communication signal to be transmitted. The smaller the PSR, the less impact the challenge disturbance has on the own communication process.
The invention provides 3 kinds of channel robust countermeasure sample waveform generating methods based on a formula (5), namely Ch-FGSM, ch-PGD and Ch-CW methods, and 4 kinds of classical mainstream countermeasure samplesThe present waveform generation method (WGN, FGSM, PGD and CW, respectively) performed a performance comparison. Experimental setup of Doppler frequency offset f for Wireless channels d ＝926Hz,N 0 =8. As can be seen from fig. 4, after the communication signal added with the challenge sample passes through the wireless channel, the performance of the 3 methods proposed by the present invention is fully due to the comparative 4 conventional methods. While the conventional method can realize the spoofing rate exceeding 80% only when the PSR is about-4 dB, the Ch-FGSM method provided by the invention can realize the spoofing rate exceeding 80% only by the PSR of-24 dB, and the Ch-PGD and Ch-CW methods can realize the spoofing rate exceeding 80% when the PSR is = -22 dB. When PSR reaches-14 dB, the 3 methods provided by the invention can realize the deception rate of 100%, namely the modulation recognition process of an eavesdropper can be completely destroyed, and the added disturbance energy is only a few tenths of the original signal.
Fig. 5 is a plot of FR versus PSR for 3 methods proposed in the present invention when the doppler frequency offset of the actual channel is changed. Since the wireless channel is changing in real time, it is often difficult to estimate accurately. Therefore, there is often a case where the channel parameters set at the time of channel compensation are inconsistent with the channel parameters actually faced. Here the Doppler shift f of the channel compensation is set d The doppler shifts of the actual channels are 726Hz, 826Hz, 926Hz, 1026Hz, 1126Hz, respectively, =926 Hz. Three method pairs f can be seen d The variation of (a) has certain robustness, wherein the robustness of the Ch-FGSM method is the strongest, and the attenuation of FR is within 10% when the Doppler frequency offset of an actual channel is +/-200 Hz (726 Hz and 126 Hz).
Fig. 6 is a plot of FR versus PSR for 3 methods proposed by the present invention when the number of multipath of the actual channel is changed. The multipath number N of the channel compensation is set 0 =8, the multipath number of the actual channel is 1, 8, 16. Three methods can be seen for N 0 The increase in (2) is more robust when N 0 At=16, FR losses for all three methods are within 2%.
Claims (8)
1. The channel precompensation-based method for generating the waveform of the countersample signal is characterized by comprising the following steps:
s1 generating a transmitted signal data set
Encoding and modulating the transmitted information to generate a transmission signal s i (k) K=1, 2,..k, K represents the kth point of the transmitted signal, K represents a total of K points of the transmitted signal, i=1, 2,..n, i represents the ith signal, N represents the total number of signals; signals of various modulation modes are required to be generated for training the modulation recognition network; when M available modulation modes are available, M is more than or equal to 2, and each transmitted signal s i (k) The corresponding modulation type is denoted by the numeral l i E {1,2,..m } is labeled, signal s is transmitted i (k) Together with the tag, form a dataset D; dividing the data set D into training sets D train And verification set D valid The modulation recognition network is used for subsequent training;
s2 construction modulation recognition network
S2.1, constructing a feature extraction module, wherein the feature extraction module consists of 4 convolution modules, 1 two-dimensional average pooling layer Avg_pool2D and 1 flat layer in series; the first convolution module consists of 1 two-dimensional convolution layer Conv2D, wherein the number of convolution kernels in the convolution layer is filter=64, and the size of the convolution kernels is kernel_size=1×3; the second to fourth convolution modules are composed of four two-dimensional convolution layers Conv2D, the number of convolution kernels of each convolution module is 64, 128, 512 and 256 in sequence, and the convolution kernel size kernel_size=3×3; the convolution kernel stepping stride=1 of each two-dimensional convolution layer Conv2D, and the zero Padding strategy is padding=1; the output of each two-dimensional convolution layer Conv2D is normalized by the Batch normal, so that the output of the two-dimensional convolution layer Conv2D is ensured not to be distributed in a saturation region of a next layer activation function, and gradient dissipation is avoided when a network is trained; the output of the two-dimensional convolution layer Conv2D subjected to the Batch normal normalization is continuously input into a ReLU activation function, wherein the ReLU activation function is nonlinear transformation and is used for enhancing the feature extraction capability of a feature extraction module; compared with other activation functions (such as sigmoid function, tanh function and the like), the ReLU activation function can overcome the condition that the gradient disappears, and is beneficial to faster convergence of the network; the output of the last convolution module adopts a two-dimensional average pooling layer Avg_pool2D with the size of 1 multiplied by 3 to reduce the dimension of the characteristics obtained by the convolution module so as to reduce the operand and prevent the network from being excessively fitted; the flat layer is used as a transition from the whole feature extraction module to a full connection layer in the identification module and is used for one-dimensional serialization of the multidimensional features output by the feature extraction module; the input data dimension of the whole feature extraction module is 1 multiplied by 2 multiplied by 1024, and the output feature dimension is 1 multiplied by 16384;
s2.2, constructing an identification module, wherein the identification module consists of two full-connection layers, the input characteristic dimension is 1 multiplied by 16384, and the output characteristic dimension is 1 multiplied by M; the output of the identification module is the identification probability of which type the signal belongs to, and the highest probability is the modulation signal type identified by the network;
the characteristic extraction module is connected with the identification module in series to obtain a modulation identification network;
s3 training modulation recognition network
S3.1, initializing training configuration: initializing a learning rate lr, a training frequency upper limit epoch, a training waiting frequency upper limit parameter, and setting the learning rate lr to 10 -6 ～10 -1 Setting the epoch to be 100-10000, wherein the upper limit of training waiting times is smaller than the epoch and is set to be one fifth or one tenth of the epoch; initializing training waiting times p=0, and initially verifying loss
S3.2 feature extraction
S3.2.1 extracting artificial characteristic of signal of training set Dtrain in S1, and extracting IQ characteristic x of signal i (k) The method comprises the steps of obtaining the characteristic with strong separability and then sending the characteristic to a characteristic extraction module of a modulation recognition network: for signal s i (k) The complex space is mapped by Hilbert transformation, then the real part and the imaginary part of the complex signal are respectively extracted and spliced to obtain the IQ characteristic x of the signal i (k)：
Wherein hilbert (·) is the hilbert transform;
s3.2.2 training set D in S1 train Is used for artificially extracting IQ characteristic x i (k) After that, x is i (k) The method is input into a feature extraction module of a modulation recognition network to automatically extract more complex features, and the feature extraction module outputs features Fx with the dimension of 1X 16384 i (k)]F represents a feature extraction module;
s3.3, calculating the identification result, and obtaining the characteristic Fx of S3.2.2 i (k)]Input recognition module for outputting recognition probability vector with dimension Mp i (m) represents the probability that the i-th signal belongs to the m-th modulated signal; taking the class with the largest probability vector as a modulation recognition result I i I.e. +.>Ii∈{1,2,...,M}；
S3.4 calculating training loss, and adding I i And data label l i Comparing to obtain a vector y i ：
Define the loss function L, p i (m) and y i Calculation of training loss L by feed loss function train ；
S3.5 optimizing the network weight, selecting an Adam algorithm to optimize the network weight, and enabling training in S3.4 to lose L train Descending, so as to obtain an optimized modulation recognition network;
s3.6 calculating verification loss, and integrating the verification set D valid Inputting the optimized modulation recognition network, and calculating a modulation recognition result I according to the steps of S3.2-S3.3 i Then adopting the loss function L to calculate the verification lossWherein t represents the training times;
s3.7, saving the optimal network weight: judging verification lossWhether or not to descend, if->The current network weight is saved, and the steps of S3.2 to S3.6 are repeated until the training times reach the upper limit; if verify loss->If the continuous non-descending times exceeds the upper limit of the training waiting times, stopping training; saving the network weight which minimizes the verification loss as the optimal network weight;
s4 adding tiny anti-disturbance in the transmitted signal
S4.1, loading the optimal network weight stored in the S3.7;
s4.2 constructing input features, and obtaining IQ features x of the emission signals according to the manner of manually extracting the features in S3.2.1 i (k)；
S4.3 modulation recognition, the IQ characteristic x of the transmitted signal i (k) The recognition probability vector P is calculated according to the steps S3.2-S3.3 i And modulation recognition result I i ；
S4.4, calculating the disturbance resistance by adopting a channel precompensation method, and obtaining the recognition probability vector P by S4.3 i And modulation recognition result I i And data label l i Calculation of test loss L using loss function test Then calculate the test loss L test IQ characteristics x of signal to be transmitted i (k) Gradient vector of (a)
The radio channel function is denoted by H (,) and the signal s is transmitted i (k) After transmission over a wireless channel, H(s) i (k) To add against disturbance asThe transmission over the radio channel against the disturbance is then H (delta i (k) A) is provided; according to the formula (4)Channel precompensation is carried out to obtain a corrected anti-disturbance signal:
wherein sign (·) is a sign function, and the superscript H denotes the conjugate transpose, J H Is a jacobian matrix of the channel function H (),for the channel transmitted signal H (s i (k) IQ characteristics H (x) i (k) The gradient vector obtained through the steps S4.2 to S4.4;
s4.5 iterative optimization challenge sample
The upper iteration number limit iter_num is preset and then the disturbance signal is to be counteractedMultiplied by a perturbation factor sigma and added to the IQ signature x of the transmitted signal i (k) In (1) obtaining an challenge sample->
S4.6 an anti-sample to which a minute anti-disturbance signal is addedInputting the result into a modulation recognition network for recognition to obtain a recognition result of the countermeasure sample.
2. A channel precompensation based method of generating an anti-sample signal waveform according to claim 1, characterized by: in S1, the total number K of points of the transmission signal is 1024, 2048, 4096.
3. A channel precompensation based method of generating an anti-sample signal waveform according to claim 1, characterized by: s1, training set D train The number of signals contained in the set should be greater than the verification set D valid The number of signals contained in the training set ensures that the signals contained in the training set have universality.
4. A channel precompensation based method of generating an anti-sample signal waveform according to claim 1, characterized by: in S3.4, the loss function L selects the cross entropy function:
5. a channel precompensation based method of generating an anti-sample signal waveform according to claim 1, characterized by: and S3.5, selecting an SGD algorithm or an RMSprop algorithm to optimize the network weight.
6. A channel precompensation based method of generating an anti-sample signal waveform according to claim 1, characterized by: in S4.5, the upper limit of the iteration number iterNum is set to 10-100.
7. A channel precompensation based method of generating an anti-sample signal waveform according to claim 1, characterized by: in S4.5, the disturbance coefficient σ is set according to the following principle: so thatThe energy of (2) is less than or equal to x i (k) One percent of the energy.
8. A channel precompensation based method of generating an antagonistic sample signal waveform according to any of the claims 1 to 7, characterised by: the feedback can be performed according to the recognition result of the challenge sample, that is, the steps S4.2 to S4.7 are repeated, so as to improve the attack effect of the challenge sample, and make the challenge sample more lifelike.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311341844.8A CN117478474B (en) | 2023-10-16 | 2023-10-16 | Channel precompensation-based antagonistic sample signal waveform generation method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311341844.8A CN117478474B (en) | 2023-10-16 | 2023-10-16 | Channel precompensation-based antagonistic sample signal waveform generation method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117478474A true CN117478474A (en) | 2024-01-30 |
| CN117478474B CN117478474B (en) | 2024-04-19 |
Family
ID=89633926
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311341844.8A Active CN117478474B (en) | 2023-10-16 | 2023-10-16 | Channel precompensation-based antagonistic sample signal waveform generation method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117478474B (en) |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20200410228A1 (en) * | 2019-06-28 | 2020-12-31 | Baidu Usa Llc | Systems and methods for fast training of more robust models against adversarial attacks |
| CN113642378A (en) * | 2021-05-14 | 2021-11-12 | 浙江工业大学 | Signal countermeasure sample detector design method and system based on N +1 type countermeasure training |
| US20230156473A1 (en) * | 2019-12-23 | 2023-05-18 | Northeastern University | Neural Network for Adversarial Deep Learning in Wireless Systems |
| CN116304618A (en) * | 2023-02-17 | 2023-06-23 | 中国电子科技集团公司第三十六研究所 | Universal anti-disturbance generation method and device applied to electromagnetic signal modulation recognition |
| CN116600267A (en) * | 2023-04-19 | 2023-08-15 | 东南大学 | Doppler resistance method based on deep reinforcement learning in high-speed rail honeycomb-free system |
| US11783037B1 (en) * | 2022-10-27 | 2023-10-10 | Quanzhou equipment manufacturing research institute | Defense method of deep learning model aiming at adversarial attacks |
-
2023
- 2023-10-16 CN CN202311341844.8A patent/CN117478474B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20200410228A1 (en) * | 2019-06-28 | 2020-12-31 | Baidu Usa Llc | Systems and methods for fast training of more robust models against adversarial attacks |
| US20230156473A1 (en) * | 2019-12-23 | 2023-05-18 | Northeastern University | Neural Network for Adversarial Deep Learning in Wireless Systems |
| CN113642378A (en) * | 2021-05-14 | 2021-11-12 | 浙江工业大学 | Signal countermeasure sample detector design method and system based on N +1 type countermeasure training |
| US11783037B1 (en) * | 2022-10-27 | 2023-10-10 | Quanzhou equipment manufacturing research institute | Defense method of deep learning model aiming at adversarial attacks |
| CN116304618A (en) * | 2023-02-17 | 2023-06-23 | 中国电子科技集团公司第三十六研究所 | Universal anti-disturbance generation method and device applied to electromagnetic signal modulation recognition |
| CN116600267A (en) * | 2023-04-19 | 2023-08-15 | 东南大学 | Doppler resistance method based on deep reinforcement learning in high-speed rail honeycomb-free system |
Non-Patent Citations (1)
| Title |
|---|
| LILAPATI WAIKHOM等: "GNN-Adv: Defence Strategy from Adversarial Attack for Graph Neural Network", IEEE, 31 January 2023 (2023-01-31) * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117478474B (en) | 2024-04-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Jagannath et al. | A comprehensive survey on radio frequency (RF) fingerprinting: Traditional approaches, deep learning, and open challenges | |
| Erpek et al. | Deep learning for wireless communications | |
| Adesina et al. | Adversarial machine learning in wireless communications using RF data: A review | |
| CN110113288B (en) | Design and demodulation method of OFDM demodulator based on machine learning | |
| Sahay et al. | A deep ensemble-based wireless receiver architecture for mitigating adversarial attacks in automatic modulation classification | |
| CN111130802A (en) | Physical layer security authentication algorithm based on physical layer excitation-response mechanism | |
| Comert et al. | Analysis of augmentation methods for RF fingerprinting under impaired channels | |
| Catak et al. | Defensive distillation-based adversarial attack mitigation method for channel estimation using deep learning models in next-generation wireless networks | |
| Sagduyu et al. | Is semantic communication secure? a tale of multi-domain adversarial attacks | |
| Ali et al. | Algorithm for automatic recognition of PSK and QAM with unique classifier based on features and threshold levels | |
| Zhang et al. | Spectrum focused frequency adversarial attacks for automatic modulation classification | |
| He et al. | Adversary detection for cognitive radio networks | |
| CN117478474B (en) | Channel precompensation-based antagonistic sample signal waveform generation method | |
| Liu et al. | A robust few-shot sei method using class-reconstruction and adversarial training | |
| Zhang et al. | Attacking Modulation Recognition with Adversarial Federated Learning in Cognitive Radio-Enabled IoT | |
| Wu et al. | Deep learning aided cyclostationary feature analysis for blind modulation recognition in massive MIMO systems | |
| Nair et al. | Rigorous analysis of data orthogonalization for self-organizing maps in machine learning cyber intrusion detection for LoRa sensors | |
| Cominelli et al. | On the properties of device-free multi-point CSI localization and its obfuscation | |
| Xiao et al. | Over-the-air adversarial attacks on deep learning Wi-Fi fingerprinting | |
| CN116634437B (en) | Frequency selection-based antagonistic sample signal waveform generation method | |
| CN115499071A (en) | Online confrontation learning test system combining frequency spectrum and intelligence | |
| CN109787996A (en) | A kind of spoof attack detection method based on DQL algorithm in mist calculating | |
| Chen et al. | AIR: Threats of Adversarial Attacks on Deep Learning-Based Information Recovery | |
| Tang et al. | Defending AI-based automatic modulation recognition models against adversarial attacks | |
| Huang et al. | Hidden backdoor attack against deep learning-based wireless signal modulation classifiers |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |