CN114553274A - Security self-precoding machine optimization method based on antagonistic learning - Google Patents

Security self-precoding machine optimization method based on antagonistic learning Download PDF

Info

Publication number
CN114553274A
CN114553274A CN202210112026.XA CN202210112026A CN114553274A CN 114553274 A CN114553274 A CN 114553274A CN 202210112026 A CN202210112026 A CN 202210112026A CN 114553274 A CN114553274 A CN 114553274A
Authority
CN
China
Prior art keywords
training
eve
parameters
self
precoding
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
Application number
CN202210112026.XA
Other languages
Chinese (zh)
Other versions
CN114553274B (en
Inventor
郑重
王新尧
费泽松
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Institute of Technology BIT
Original Assignee
Beijing Institute of Technology BIT
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Beijing Institute of Technology BIT filed Critical Beijing Institute of Technology BIT
Priority to CN202210112026.XA priority Critical patent/CN114553274B/en
Publication of CN114553274A publication Critical patent/CN114553274A/en
Application granted granted Critical
Publication of CN114553274B publication Critical patent/CN114553274B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B7/00—Radio transmission systems, i.e. using radiation field
    • H04B7/02—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas
    • H04B7/04—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas
    • H04B7/0413—MIMO systems
    • H04B7/0456—Selection of precoding matrices or codebooks, e.g. using matrices antenna weighting
    • 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
    • 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)
  • Mobile Radio Communication Systems (AREA)

Abstract

The invention discloses a security self-precoding machine optimization method based on countermeasure learning, and belongs to the technical field of wireless communication physical layer security. Aiming at the technical problems of non-convex safety rate optimization and high complexity of a large-scale multi-antenna system, the invention designs and constructs a safety transceiver training frame based on a self-encoder and supporting multi-user, multi-antenna and multi-data stream transmission, and performs joint optimization on signal modulation and a space precoder at a transmitting end, so that a legal user receiving end demodulates at an extremely low symbol error rate to recover correct confidential information, and an eavesdropping user receiving end cannot demodulate correctly and only can obtain wrong confidential information. The trained safety transmitter can improve the communication reliability of a legal user and simultaneously remarkably reduce the reliability of the eavesdropping user, thereby realizing safe transmission. In addition, the invention reduces the model convergence time, reduces the space signal processing complexity and improves the safe transmission efficiency by introducing the counterstudy strategy.

Description

Security self-precoding machine optimization method based on antagonistic learning
Technical Field
The invention relates to a security self-precoding machine optimization method based on countermeasure learning, and belongs to the technical field of wireless communication physical layer security.
Background
A large-scale Multiple-Input Multiple-Output (MIMO) system is a key enabling technology of the next generation mobile communication B5G/6G, can provide higher-speed physical layer data transmission efficiency, can remarkably improve the spatial freedom degree of signal processing by increasing the number of antenna arrays, and brings greater potential for precoding-based physical layer safety design and reliability gain. However, a large-scale antenna often needs to integrate dozens to hundreds of antenna units on an antenna panel, and especially in a spatial multi-stream transmission scenario, the number of radio frequency links increases with the increase of the number of antennas, which further increases the signal processing complexity of a transmitting end and the hardware processing overhead of power amplification and the like. Meanwhile, end-to-end transceiver design based on deep learning draws wide attention in academia and industry in recent years, the network of a transceiving end is trained jointly by using measured channel environment data and priori knowledge, the method is different from the prior communication system design and is realized by adopting a mode of cascading functional modules, the modules are independently optimized, and the performance of the designed communication system is not optimal under the condition. And an end-to-end design based on an Automatic Encoder (AE) for deep learning can design a unified target function for a plurality of cascade modules, perform multi-module joint optimization and realize integral optimization. In addition, deep learning can be achieved by unloading large-scale array signal processing calculation overhead to an offline training stage and guiding model training in a data-driven mode, so that the signal calculation time of the online stage is shortened, and especially for the conditions of high-order signal modulation and a larger number of antenna systems, the method has the advantages that the traditional algorithm is incomparable.
Currently, depth autocoders are beginning to be used to study physical layer secure communications. Some existing work provides some solutions to the problem, and the X.L.Zhang adopts a safety precoding design method based on supervised learning, firstly, an iterative optimization and water injection algorithm is utilized to perform combined optimization on a precoding direction vector and a power distribution vector, and a suboptimal solution of a signal covariance matrix under an MIMO channel is solved. However, the safety design method based on supervised learning has the advantage that although the computational complexity is reduced compared with the traditional scheme, the achievable practical upper limit of performance is always limited by the traditional scheme. Therefore, c.h.lin studied an end-to-end physical layer security scheme based on a variational self-encoder, in which an objective function that directs model update was designed as a sum of three parts, directing optimization of communication rate, security performance, and noise adaptation performance, respectively. Wherein the security part is realized by minimizing mutual information between the secret information and the precoded symbols, wherein the mutual information is characterized by a correlation function. Li has studied a Mutual Information Neural Estimation (MINE) based network model that can approximate the Mutual Information size of input and output distributions of a Neural network, and opens the door to design Information theory based physical layer secure communication using deep learning. In addition, r.fritschek studied an end-to-end secure self-encoder scheme based on user error rate, which designed a secure transmitter based on a neural network by designing an objective function including maximizing the error rate of an eavesdropping user and minimizing the error rate of a legitimate user.
Most of physical layer security technologies designed for MIMO channels in the scheme are based on supervised learning methods, and the problem that the security rate is limited by the traditional method is difficult to overcome; the secure self-coding machine scheme based on unsupervised learning is designed for secure channel coding or a high-level symmetric encryption algorithm, and is not designed for large-scale array signal processing layers such as a secure constellation diagram and secure pre-coding, and most system simulations are performed under the condition of a small-scale antenna array or a single-antenna system. In addition, with the close combination of artificial intelligence and mobile communication, an illegal eavesdropping user also has the ability to obtain the priori knowledge of the transmitter through blind modulation identification, transmitter fingerprint identification and the like, so that the receiving and cracking capabilities of own confidential information are improved, and the risk that a legal system is eavesdropped is further increased.
Disclosure of Invention
Aiming at the problems of non-convex safety rate optimization problem and high complexity of a future large-scale multi-antenna system, the invention provides a safety self-precoding machine optimization method based on countermeasure learning. By designing and constructing a safety transceiver training frame based on a self-encoder and simultaneously supporting multi-user, multi-antenna and multi-data stream transmission, signal modulation and space precoder are jointly optimized at a transmitting end, so that a legal user receiving end demodulates at an extremely low symbol error rate and recovers correct confidential information, while an eavesdropping user receiving end cannot demodulate correctly and only can obtain wrong confidential information. The trained safety transmitter can greatly reduce the reliability aiming at eavesdropping users while improving the communication reliability of legal users, thereby realizing safe transmission.
The purpose of the invention is realized by the following technical scheme:
aiming at the technical problems of non-convex safety rate optimization problem and high complexity in a large-scale MIMO system, the safety self-precoding machine is trained based on antagonistic learning. A modulation module and a space pre-coding module which are cascaded are designed in a combined optimization mode, a safe transmitter constellation diagram and a full-digital beam forming vector are designed, and a safe self pre-coder (SAP) is obtained through training. Meanwhile, an iterative countermeasure learning training framework is introduced, an eavesdropping receiver with better symbol detection capability is developed under the condition that parameters of a Security transmitter are known, and based on the eavesdropping receiver evolved, an countermeasure Security self-pre-coder (ASAP) with higher legal user information reliability is trained in an countermeasure mode. The trained safety transmitter can greatly reduce the reliability aiming at eavesdropping users while improving the communication reliability of legal users, thereby realizing safe transmission.
The invention discloses a security self-precoding machine optimization method based on antagonistic learning, which comprises the following steps:
step one, setting system parameters of an MIMO communication system based on a self-encoder frame, wherein the system parameters comprise: the number of antennas M, N of the transmitter Alice, the legal user Bob and the eavesdropping user EveBAnd NEBit information R of each symbol, length J of symbol sequence, channel multipath quantity L and channel parameter alphal，θl，
Figure BDA0003492490380000021
Distribution of (d), transmit power constraint p, signal-to-noise ratio SNR; setting a neural network model structure, training/testing dataset parameters, and training hyper-parameters, the training hyper-parameters comprising: selected optimizer, training round Epoch, sample length per round Batch Size.
Step two, building an MIMO communication system supporting multi-user, large-scale multi-antenna and multi-stream data transmission based on a training framework of a depth self-coding machine, designing a multi-user average SER as a loss function, and training a transmitter network model formed by cascading a modulation module and a pre-coding module; the trained model can realize the reliable transmission of multiple users in the system under the limited transmitting power through space beam forming.
Step 2.1: and designing a transmitting terminal neural network, wherein the network structure comprises a signal modulation module and a space pre-coding module.
The transmitting symbol corresponding to the jth secret information after the antenna mapping of the transmitting end is Xj：
Figure BDA0003492490380000031
Wherein the content of the first and second substances,
Figure BDA0003492490380000032
and
Figure BDA0003492490380000033
respectively representing a modulation module and a precoding module; m isjIs a secret information to be transmitted, from a predetermined limited set of secret information
Figure BDA0003492490380000034
Is obtained in (1). The modulation symbol is output after passing through the modulation neural network module
Figure BDA0003492490380000035
sjEstimating channel parameters with the transmitting end
Figure BDA0003492490380000036
Combining and inputting precoding neural network module together
Figure BDA0003492490380000037
Carrying out precoding operation on the modulation symbols to obtain a precoded signal Xj. Likewise, J XjThe method can process and send in parallel to realize multi-stream signal transmission of the MIMO system. All parameters in the training and testing data sets, including channel parameters and signal parameters, are expressed in a manner that real parts and imaginary parts are separated, that is, all channels and signals in the system are characterized as a real matrix.
All transmitting end networks adopt a Fully-Connected Neural Network (FCNN) and a modulation module Neural Network
Figure BDA0003492490380000038
The calculation process for the secret information sequence m is represented as follows:
Figure BDA0003492490380000039
wherein, the first and the second end of the pipe are connected with each other,
Figure BDA00034924903800000310
and
Figure BDA00034924903800000311
separately representing modulation module neural networks
Figure BDA00034924903800000312
The activation function, the weight vector and the offset vector of the g-th network; then all symbol sequences s after modulation and the estimated channel parameters estimated by the transmitting end
Figure BDA00034924903800000313
Combining to obtain new training sample
Figure BDA00034924903800000314
AsSpatial precoding module neural network
Figure BDA00034924903800000315
The input of (1); spatial precoding module neural network
Figure BDA00034924903800000316
The calculation process of the matrix U after combining the modulation symbols and the channel is expressed as follows:
Figure BDA00034924903800000317
wherein the content of the first and second substances,
Figure BDA0003492490380000041
and
Figure BDA0003492490380000042
separately representing modulation module neural networks
Figure BDA0003492490380000043
T ofthAn activation function, a weight vector and an offset vector of a layer network; to limit the transmit signal power | X | ≦ p,
Figure BDA0003492490380000044
t ofthThe layer is designed as a power constraint layer and adopts a self-defined activation function
Figure BDA0003492490380000045
The following were used:
Figure BDA0003492490380000046
wherein | X | represents the F-norm of matrix X, and p represents the maximum transmit power; thus, a normalized signal to be transmitted mapped to the antenna port is obtained via step 2.1.
Step 2.2: and designing a receiving end neural network, which comprises a received signal detection module and a probability mapping module.
The legal user Bob and the eavesdropping user Eve are regarded as two legal users in the step, and two receiver models with the same network structure are built. J ththSignal X with normalized powerjRespectively reaches a receiving end through respective MIMO channels of Bob and Eve, and the j-th channel received by Bob and EvethA signal YB,jAnd YE,jRespectively, as follows:
YB,j＝HBXj+nB (5)
YE,j＝HEXj+nE (6)
wherein the content of the first and second substances,
Figure BDA0003492490380000047
and
Figure BDA0003492490380000048
representing additive white gaussian noise.
At the receiving end, the receiver networks of Bob and Eve adopt the same network structure, which is respectively expressed as:
Figure BDA0003492490380000049
and
Figure BDA00034924903800000410
j-th of receiver recovery for Bob and EvethThe secret information is represented as:
Figure BDA00034924903800000411
Figure BDA00034924903800000412
wherein alpha isB,θB,
Figure BDA00034924903800000413
Individual watchShowing channel parameters estimated by a Bob receiving end; alpha is alphaE,θE,
Figure BDA00034924903800000414
Respectively representing the channel parameters estimated by an Eve receiving end;
the last layer of the receiving end neural network adopts a Softmax activation function to respectively output prediction probability vectors P of Bob and EveBAnd PE(ii) a The probability vector represents
Figure BDA00034924903800000415
And
Figure BDA00034924903800000416
the predicted secret information is a set of secret information
Figure BDA00034924903800000417
The probability corresponding to a certain secret information.
Step 2.3: designing an average cross-entropy loss function for multiple users according to the classified cross-entropy loss function
Figure BDA0003492490380000051
And updating the parameters of the self-precoding machine model by adopting a reverse gradient descent strategy.
Average cross entropy loss function of legal user Bob and eavesdropping user Eve
Figure BDA0003492490380000052
The design is as follows:
Figure BDA0003492490380000053
wherein, Pj,mOne-hot encoding matrix representing transmitted secret information sequence m
Figure BDA0003492490380000054
Row j, column m;
Figure BDA0003492490380000055
and
Figure BDA0003492490380000056
probability prediction matrices for receiver networks representing Bob and Eve, respectively
Figure BDA0003492490380000057
And
Figure BDA0003492490380000058
row j, column m;
Figure BDA0003492490380000059
expressed in discrete parameter sets
Figure BDA00034924903800000510
Next, the average of the loss function calculated under the number of batch size data samples, the batch size representing the length of each batch of training samples sent into the neural network.
Optimizer based on Tensorflow deep learning framework, minimizing the above-mentioned average cross entropy loss function
Figure BDA00034924903800000511
The confidential information recovered by the receiver sides of Bob and Eve is processed through an end-to-end neural network, an unsupervised training process is achieved, signals recovered by a receiving end are consistent with signals of a sending end, reliable transmission of the confidential information is achieved, both Eve and Bob can obtain an optimal receiver in the channel scene, meanwhile, the trained Eve is considered to be an optimal eavesdropper based on self-encoding training, and subsequent safety design is conducted on the basis of the eavesdropper of the user Eve.
Step three, designing a new safety loss function by introducing a fuzzy matrix P aiming at the eavesdropping user according to the multi-user and multi-stream MIMO self-precoding machine model built in the step two
Figure BDA00034924903800000512
Guiding models with new safety loss functionsTraining, giving safety attributes to the self-precoding machine, generating a new safety constellation diagram, and ensuring that a receiving end of a legal user Bob can complete symbol detection, and a receiving end of an eavesdropping user Eve cannot correctly complete symbol detection.
Step 3.1: similar to the second step, in order to realize the safe transmission of the physical layer signal, a new safety loss function is designed aiming at the fuzzy matrix P of the eavesdropping user
Figure BDA00034924903800000513
Is represented as follows:
Figure BDA00034924903800000514
wherein, Pj,m，
Figure BDA00034924903800000515
And
Figure BDA00034924903800000516
the meaning is in accordance with formula (9); introducing an ambiguity matrix P to confuse eavesdropping subscriber receivers, Pj,mThe elements representing the jth row and mth column of the blur matrix P, which is written as follows:
Figure BDA0003492490380000061
according to the principle of a cross entropy loss function, the prediction probability matrix of the eavesdropping user Eve receiver is closer to the fuzzy matrix P along with the progress of the training process, so that the probabilities of Eve judging that the received symbols belong to a certain class are consistent, the symbols cannot be obtained for detection, and Bob can correctly detect the received symbols, so that the secure transmission of confidential information is realized.
Step 3.2: based on the new safety loss function designed in step 3.1
Figure BDA0003492490380000062
On fixed eavesdropping user interfaceAnd carrying out safety training under the condition that the receiver parameters are not changed, and training to obtain the SAP.
Step 3.2.1: firstly, determining the total training times N, and initializing N to be 1; reading the parameters of the pre-trained self-encoder model in the step two, including the initialization network parameter phi of the transmitter AliceAAnd receiver initialization network parameters of Bob and Eve
Figure BDA0003492490380000063
And
Figure BDA0003492490380000064
step 3.2.2: initializing channel parameters of training samples and corresponding one-hot coded labels, and reading a training data set;
step 3.2.3: determining a training hyperparameter: the method comprises the steps of optimizing the learning rate of an optimizer, training turns, the length of each batch of samples, and the division ratio of a training data set and a verification data set;
step 3.2.4: starting training, updating all network model parameters phi by using an Adam optimizer based on a loss function (10)A′，
Figure BDA0003492490380000065
And
Figure BDA0003492490380000066
step 3.2.5: n is n + 1; and ending the training until N is equal to N.
Step four: an antagonistic learning mechanism is introduced, and a target loss function aiming at the eavesdropping user is designed by combining the safety loss function of the step 3.1
Figure BDA0003492490380000067
Dividing the whole self-precoding machine into two parts of links, namely a legal link Main Chain and an eavesdropping link Eve Chain, wherein the Main Chain comprises an Alice transmitter and a receiver network of Bob, the Eve Chain comprises an Eve receiver network, designing two parts of iterative confrontation training algorithms based on the pre-training model in the second step, and obtaining the confrontation safety self-coding algorithmThe coder model ASAP.
Step 4.1: an antagonistic learning mechanism is introduced, and a target loss function aiming at the eavesdropping user is designed by combining the safety loss function of the step 3.1
Figure BDA0003492490380000068
Figure BDA0003492490380000071
Wherein, Pj,m，
Figure BDA0003492490380000072
Is identical to the expression in formula (10); the purpose of this loss function is to continue to optimize the virtual eavesdropping receiver for the secure transmitter to obtain lower SER values after the secure transmitter training of step 3.2 is completed.
Step 4.2: the whole self-precoding machine is divided into two parts, namely a legal link Main Chain and an eavesdropping link Eve Chain, the Main Chain comprises an Alice transmitter and a receiver network of Bob, the Eve Chain comprises an Eve receiver network, and based on the pre-training model in the step two, an iterative confrontation training algorithm of the two parts is designed to obtain the confrontation safety self-precoding machine ASAP.
Step 4.2.1: firstly, determining the total iteration number N; determining a training hyperparameter: the method comprises the steps of optimizing the learning rate of an optimizer, training round, the length of a sample of each batch, and the division ratio of a training data set and a verification data set;
step 4.2.2: is iteration turn n? (ii) a If n is 1, reading the model parameters of the pre-trained self-precoding machine in the step two, including the initialized network parameter phi of the transmitter AliceAAnd receiver initialization network parameters of Bob and Eve
Figure BDA0003492490380000073
And
Figure BDA0003492490380000074
if n ≠ 1, read n-Model parameters of 1 round update, ΦA＝Φ′A,
Figure BDA0003492490380000075
Step 4.2.3: initializing channel parameters of training samples
Figure BDA0003492490380000076
Reading a training data set with the corresponding one-hot coded label P;
step 4.2.4: setting the training times epoch _1, freezing the network parameters of Eve Chain, training the network model of Main Chain according to the loss function (10), and updating the parameter phiA＝Φ′A,
Figure BDA0003492490380000077
Step 4.2.5: setting the training times epoch _2, freezing the network parameters of the Main Chain, training the network model of the Eve Chain according to the loss function (12), and updating the parameters
Figure BDA0003492490380000078
Step 4.2.6: n is n + 1; returning to step 4.2.2, 4.2.3, 4.2.4 and 4.2.5 are continuously executed until N is equal to N, and the training is finished.
Step five: according to the countermeasure security self-precoding machine ASAP obtained by training in the step four, under a new security transmission scene, by collecting a small amount of channel samples, the step four is continuously executed to finely adjust the model, and then the security self-precoding machine model with updated parameters is used for carrying out combined optimization of modulation and precoding on confidential information, so that a signal to be transmitted with confidential property for a target eavesdropping user is obtained. And the interception user Eve can only obtain the symbol detection performance of the blind guess level while the legal user Bob has high reliability, so that the safe transmission is realized.
And finally, completing the whole process of the safety self-precoding machine optimization method based on the counterstudy through the steps from the first step to the fifth step.
Has the advantages that:
1. the invention discloses a security self-precoding machine optimization method based on antagonistic learning, which can jointly optimize a signal modulation module and a space precoding module of a transmitting terminal under the condition of meeting the maximum transmitting power constraint, optimize a brand-new security receiving constellation map, obviously reduce the information reliability of a receiving terminal of an eavesdropping user while improving the information decoding reliability of the receiving terminal of a legal user, and realize the security transmission of confidential information under any receiving-transmitting antenna relationship.
2. The invention discloses a security self-precoding machine optimization method based on antagonistic learning, which can be used for alternately and iteratively training a legal link network and an eavesdropping link network by introducing an antagonistic learning strategy on the basis of a security self-precoding machine, continuously improving the security information transmission reliability and security of a legal user even if an eavesdropper has active learning capacity, and simultaneously reducing the model convergence time, the space signal processing complexity and the security transmission efficiency compared with the security self-precoding machine optimization method without antagonistic learning.
3. The safety self-precoding machine optimization method based on the countermeasure learning disclosed by the invention can effectively improve the communication transmission throughput under the limited transmitting power by designing reasonable receiving end constellation deployment while ensuring the information safety and reliability, and realizes the balance of the communication safety and effectiveness.
Drawings
FIG. 1 is a flowchart of an overall training framework of a countermeasure safety self-precoding machine in an optimization method and an embodiment of the countermeasure safety self-precoding machine based on countermeasure learning according to the present invention;
FIG. 2 is a schematic diagram of a method and an embodiment of a joint modulation constellation based on a deep self-coder and a secure self-precoding machine optimization according to the present invention;
FIG. 3 is a schematic diagram of an iterative training algorithm of a security transmitter and a wiretap receiver based on countermeasure learning in the optimization method of the security self-precoding machine based on countermeasure learning and the embodiment of the invention;
FIG. 4 is a schematic diagram showing the comparison of SER vs SNR performances of a legal user Bob and an eavesdropping user Eve under the training frames of the security self-precoding machine SAP and the countermeasure security self-precoding machine ASAP in the countermeasure learning-based security self-precoding machine optimization method and embodiment of the invention;
fig. 5 is a schematic diagram of comparison of signal constellations of receiving ends of a legal user Bob and an eavesdropping user Eve under two safety training frames in the method for optimizing the safety self-precoding machine based on countermeasures learning and the embodiment of the invention.
Detailed Description
The present invention will be described in detail below with reference to the accompanying drawings and examples. The technical problems and the advantages solved by the technical solutions of the present invention are also described, and it should be noted that the described embodiments are only intended to facilitate the understanding of the present invention, and do not have any limiting effect.
This embodiment details the steps of the security self-precoding machine optimization method based on counterstudy when the method is implemented specifically under specific system parameter configuration, self-coding machine network parameter configuration, and training hyper-parameter configuration.
In the embodiment, a three-node MIMO interception channel scene is considered, and the whole communication system is built based on the architecture of a depth self-encoder; in the system, Alice represents a safety transmitter trained based on a neural network model, and Bob represents a legal user receiver obtained by end-to-end training; eve represents an eavesdropping user receiver which is also trained end to end, and after model pre-training is finished, Eve can be trained to obtain a receiver with basically consistent receiving and demodulating performances with Bob.
The channel conforms to the 5G millimeter wave channel model as follows: (the invention and the channel model itself have no constraint relation, and only exemplifies the algorithm implementation process)
Figure BDA0003492490380000091
Wherein, L represents the number of scattering diameters; alpha is alphalRepresenting the complex channel gain in the first path; a isr() represents a receive antenna array response vector; a ist(. represents hair)A transmit antenna array response vector;
Figure BDA0003492490380000092
and thetalRespectively representing an arrival angle and an emission angle; with a Uniform Linear Array (ULA) in the K dimension, the array response vector of the antenna is expressed as follows:
Figure BDA0003492490380000093
wherein T represents the number of transmitting antennas, d represents the distance between two adjacent antennas on the antenna panel, and λ is the electromagnetic wavelength; meanwhile, the perfect channel information of Bob and Eve is known by Alice
Figure BDA0003492490380000094
The same is true for the receiving end.
As shown in fig. 1, the secure self-precoding machine optimization method based on counterlearning disclosed in this embodiment includes the following specific implementation steps:
step one, setting system parameters of an MIMO communication system based on a self-encoder frame, wherein the system parameters comprise: the number of antennas M, N of the transmitter Alice, the legal user Bob and the eavesdropping user EveBAnd NEBit information R of each symbol, length J of symbol sequence, channel multipath quantity L and channel parameter alphal，θl，
Figure BDA0003492490380000095
Distribution of (d), transmit power constraint p, signal-to-noise ratio SNR; setting a neural network model structure, training/testing dataset parameters, and training hyper-parameters, the training hyper-parameters comprising: selected optimizer, training round Epoch, sample length per round Batch Size. Regarding the system configuration parameters, the number of antennas of the transmitter Alice is set to M-64; the number of antennae of legal user Bob is NB2; the number of the antennas for intercepting Eve of the user is NE2; transmitting a symbol sequence length J of 1, wherein each symbol contains information of R of 4 bits; the relative positions of the channel multipath numbers L being 1, Alice, Bob and Eve are random along withVaries with variations in discrete channel parameters, and the channel parameter distribution is:
Figure BDA0003492490380000101
the maximum transmitting power is p-1; meanwhile, in order to adapt to a plurality of signal-to-noise ratio scenes, a data enhancement method is adopted, and the SNR range in the training data set is set to obey uniform distribution, namely
Figure BDA0003492490380000102
Regarding the model configuration hyper-parameters, all the Neural Network modules adopt a Fully-Connected Neural Network (FCNN); 5 layers of FCNN are adopted for the modulation module, and the number of neuron nodes of each layer is 512,256,128,32,2](ii) a Adopting 5 layers of FCNN for the precoding module, wherein the number of neuron nodes of each layer is [512,512,256,256,128 ]](ii) a Receiver modules for Bob and Eve respectively adopt 4 layers of FCNN, and the number of neuron nodes in each layer is [128, 64 and 16 ]](ii) a In addition, the activation function of each layer adopts a Rectified Linear Unit (ReLU); updating model parameters by adopting an Adam optimizer; in addition, 10 is generated based on the system configuration parameters described above6A model data sample including discrete channel parameters
Figure BDA0003492490380000103
Sending secret information sequence m ═ m corresponding to each channel realization1,m2,…mJAnd a One-hot encoder (OE) matrix corresponding to each secret information sequence; this matrix is used as a label for each sample at the receiving end. The number of times of iterative training is set to 50, the number of times of model training Epoch in each iteration is set to 30, and the number of samples Batch Size in each input network is set to 512.
And step two, constructing an MIMO communication system supporting multi-user, large-scale multi-antenna and multi-stream data transmission based on a model architecture and a training method of the deep self-coding machine. Designing a transmitter network model in which a signal modulation module and a space pre-coding module are cascaded, designing a multi-user average SER loss function, and realizing the reliable transmission of multiple users in the system through space beam forming.
Step 2.1: and designing a transmitting end neural network, which comprises a cascaded modulation module and a spatial pre-coding module.
The sending symbol corresponding to the jth secret information after the antenna mapping of the sending end is Xj：
Figure BDA0003492490380000104
Wherein the content of the first and second substances,
Figure BDA0003492490380000105
and
Figure BDA0003492490380000106
respectively representing a modulation module and a precoding module; m is a unit ofjIs secret information to be sent, outputs modulation symbols after passing through a modulation neural network module
Figure BDA0003492490380000107
sjCo-discrete channel parameters
Figure BDA0003492490380000108
Combining and inputting precoding neural network module together
Figure BDA0003492490380000109
Multiple XjThe processing and transmission can be performed in parallel to achieve multi-stream signaling for a MIMO system. It should be noted that, currently, the mainstream deep learning framework based on the tensoflow cannot directly represent complex numbers, and therefore, all parameters in the data set, including channel parameters and signal parameters, are represented in a manner that real parts and imaginary parts are separated, that is, all signals and channels are represented as real matrixes.
All network structures of the transmitting and receiving ends adopt FCNN, and the modulation module neural network
Figure BDA0003492490380000111
The calculation process for the secret information sequence m is represented as follows:
Figure BDA0003492490380000112
wherein the content of the first and second substances,
Figure BDA0003492490380000113
and
Figure BDA0003492490380000114
separately representing modulation module neural networks
Figure BDA0003492490380000115
The activation function, the weight vector and the offset vector of the g-th layer network; then all symbol sequences s after modulation and the estimated sparse channel parameters
Figure BDA0003492490380000116
Combining to obtain new training sample
Figure BDA0003492490380000117
Neural network as spatial precoding module
Figure BDA0003492490380000118
The input of (1); therefore, the modulation symbol sequence is mapped to the antenna in the precoding module with the aid of the channel prior information, and the signal matrix to be transmitted is obtained as follows:
Figure BDA0003492490380000119
wherein the content of the first and second substances,
Figure BDA00034924903800001110
and
Figure BDA00034924903800001111
are respectively provided withRepresentation modulation module neural network
Figure BDA00034924903800001112
T ofthAn activation function, a weight vector and an offset vector of a layer network; in particular, to limit the transmit signal power | X | ≦ p,
Figure BDA00034924903800001113
t ofthThe layer is designed as a power constraint layer and adopts a self-defined activation function
Figure BDA00034924903800001114
The following were used:
Figure BDA00034924903800001115
wherein | X | represents the F-norm of matrix X, and p represents the maximum transmit power constraint;
step 2.2: and designing a receiving end neural network, which comprises a received signal detection module and a probability mapping module.
J (th) after power normalizationthA signal XjAnd respectively reaching a receiving end through respective MIMO channels of Bob and Eve, wherein the channel passing process is represented as follows:
YB,j＝HBXj+nB (19)
YE,j＝HEXj+nE (20)
wherein the content of the first and second substances,
Figure BDA00034924903800001116
and
Figure BDA00034924903800001117
representing additive white gaussian noise.
At the receiving end, the receiver networks of Bob and Eve adopt the same network structure, which is respectively expressed as:
Figure BDA0003492490380000121
and
Figure BDA0003492490380000122
j-th of receiver recovery for Bob and EvethThe secret information is represented as:
Figure BDA0003492490380000123
Figure BDA0003492490380000124
wherein alpha isB,θB,
Figure BDA0003492490380000125
Respectively representing the channel parameters estimated by the Bob receiving end; alpha is alphaE,θE,
Figure BDA0003492490380000126
Respectively representing the channel parameters estimated by an Eve receiving end; according to the illustration in fig. 2, the last layer of the receiving end neural network adopts the Softmax activation function to output the prediction probability vectors P of Bob and Eve respectivelyBAnd PE(ii) a The probability vector represents
Figure BDA0003492490380000127
And
Figure BDA0003492490380000128
the predicted secret information is a set of secret information
Figure BDA0003492490380000129
The probability corresponding to a certain secret information.
Step 2.3: designing an average cross-entropy loss function for multiple users according to the classified cross-entropy loss function
Figure BDA00034924903800001210
Directing self-precodersAnd (5) network training.
Average cross entropy loss function of Bob and Eve
Figure BDA00034924903800001211
The design is as follows:
Figure BDA00034924903800001212
wherein, Pj,mOne-hot encoding matrix representing transmitted secret information sequence m
Figure BDA00034924903800001213
Row j, column m;
Figure BDA00034924903800001214
and
Figure BDA00034924903800001215
probability prediction matrices for receiver networks representing Bob and Eve, respectively
Figure BDA00034924903800001216
And
Figure BDA00034924903800001217
row j, column m;
Figure BDA00034924903800001218
expressed in discrete parameter sets
Figure BDA00034924903800001219
Next, the average of the loss function calculated under the number of batch size data samples, the batch size representing the length of each batch of training samples sent into the neural network.
Adam optimizer self-contained with Tensorflow deep learning framework is adopted to minimize the above average cross entropy loss function
Figure BDA00034924903800001220
Implementation ofUnsupervised training process.
Step three, designing a new safety loss function by introducing a fuzzy matrix P aiming at the eavesdropping user according to the multi-user and multi-stream MIMO self-precoding machine model built in the step two
Figure BDA00034924903800001221
And guiding model training by using a new safety loss function, giving a self-precoding machine safety attribute, generating a new safety constellation diagram, and ensuring that a receiving end of a legal user Bob can finish symbol detection, and a receiving end of an eavesdropping user Eve cannot finish symbol detection correctly.
Step 3.1: similar to the second step, in order to realize the safe transmission of the physical layer signal, a new safety loss function designed for the fuzzy matrix P of the eavesdropping user is introduced
Figure BDA0003492490380000131
Is represented as follows:
Figure BDA0003492490380000132
wherein, Pj,m，
Figure BDA0003492490380000133
And
Figure BDA0003492490380000134
the meaning is consistent with that in formula (23); introducing an ambiguity matrix P to confuse eavesdropping subscriber receivers, Pj,mThe elements representing the jth row and mth column of the obfuscation matrix P, which is written in the form:
Figure BDA0003492490380000135
according to the principle of a cross entropy loss function, along with the progress of a training process, a prediction probability matrix of an Eve receiver of an eavesdropping user is closer to a confusion matrix P, so that the probabilities of Eve judging that received symbols belong to a certain class are consistent, therefore, symbol detection cannot be performed in a striving way, Bob can perform correct detection, and the secure transmission of confidential information is realized.
Step 3.2: based on the new safety loss function designed in step 3.1
Figure BDA0003492490380000136
And carrying out safety training under the condition that the parameters of the fixed eavesdropping user receiver are not changed, and training to obtain the SAP.
Step 3.2.1: firstly, determining the total training times N, and initializing N to be 1; reading the model parameters of the pre-trained self-precoding machine in the step two, including the initialization network parameter phi of the transmitter AliceAAnd receiver initialization network parameters of Bob and Eve
Figure BDA0003492490380000137
And
Figure BDA0003492490380000138
step 3.2.2: initializing channel parameters of training samples and corresponding one-hot coded labels, and reading a training data set;
step 3.2.3: determining a training hyperparameter: the method comprises the steps of optimizing the learning rate of an optimizer, training turns, the length of a sample of each batch, the division ratio of a training data set and a verification data set and the like;
step 3.2.4: starting training, updating all network model parameters phi by using an Adam optimizer based on a loss function (24)A′，
Figure BDA0003492490380000139
And
Figure BDA00034924903800001310
step 3.2.5: n is n + 1; and ending the training until N is equal to N.
Step four: an antagonistic learning mechanism is introduced, and a target loss function aiming at the eavesdropping user is designed by combining the safety loss function of the step 3.1
Figure BDA00034924903800001311
And (2) dividing the whole self-precoding machine into two parts of links, namely a legal link Main Chain and an eavesdropping link Eve Chain, wherein the Main Chain comprises an Alice transmitter and a receiver network of Bob, the Eve Chain comprises the receiver network of Eve, and designing two parts of iterative confrontation training algorithms based on the pre-training model in the second step to obtain a confrontation safety self-encoder model ASAP.
Step 4.1: an antagonistic learning mechanism is introduced, and a target loss function aiming at the eavesdropping user is designed by combining the safety loss function of the step 3.1
Figure BDA0003492490380000141
Figure BDA0003492490380000142
Wherein, Pj,m，
Figure BDA0003492490380000143
Is identical to the expression in formula (24); the purpose of this loss function is to continue to optimize the virtual eavesdropping receiver for the secure transmitter to obtain lower SER values after the secure transmitter training of step 3.2 is completed.
Step 4.2: according to the iterative security training algorithm block diagram shown in fig. 3, the whole self-precoding machine model is divided into two parts, namely a legal link Main Chain and an eavesdropping link Eve Chain, wherein the Main Chain comprises an Alice transmitter and a receiver network of Bob, the Eve Chain comprises an Eve receiver network, and two parts of iterative countertraining algorithms are designed based on the pre-training model in the second step to obtain the countersecurity self-precoding machine ASAP.
Step 4.2.1: firstly, determining the total iteration number N; determining a training hyperparameter: the method comprises the steps of optimizing the learning rate of an optimizer, training turns, the length of a sample of each batch, the division ratio of a training data set and a verification data set and the like;
step 4.2.2: is iteration turn n?(ii) a If n is 1, reading the model parameters of the pre-trained self-precoding machine in the step two, including the initialized network parameter phi of the transmitter AliceAAnd receiver initialization network parameters of Bob and Eve
Figure BDA0003492490380000144
And
Figure BDA0003492490380000145
if n ≠ 1, the model parameters updated in n-1 rounds are read, phiA＝Φ′A,
Figure BDA0003492490380000146
Step 4.2.3: initializing channel parameters of training samples
Figure BDA0003492490380000147
Reading a training data set with the corresponding one-hot coded label P;
step 4.2.4: setting the training times epoch _1, freezing the network parameters of Eve Chain, training the network model of Main Chain according to the loss function (24), and updating the parameter phiA＝Φ′A,
Figure BDA0003492490380000148
Step 4.2.5: setting the training times epoch _2, freezing the network parameters of the Main Chain, training the network model of the Eve Chain according to the loss function (26), and updating the parameters
Figure BDA0003492490380000149
Step 4.2.6: n is n + 1; returning to step 4.2.2, 4.2.3, 4.2.4 and 4.2.5 are continuously executed until N is equal to N, and the training is finished.
Step five: according to the countermeasure security self-precoding machine ASAP obtained by training in the step four, under a new security transmission scene, by collecting a small amount of channel samples, the step four is continuously executed to finely adjust the model, and then the security self-precoding machine model with updated parameters is used for carrying out combined optimization of modulation and precoding on confidential information, so that a signal to be transmitted with confidential property for a target eavesdropping user is obtained. And the interception user Eve can only obtain the symbol detection performance of the blind guess level while the legal user Bob has high reliability, so that the safe transmission is realized.
FIG. 4 is a diagram of SER vs SNR performance simulation results of a legal user Bob and an eavesdropping user Eve under the training frames of the security self-precoding machine SAP and the countermeasure security self-precoding machine ASAP in the countermeasure learning-based security self-precoding machine optimization method and embodiment of the invention;
in fig. 4, the abscissa is the SNR, the range is 0 to 30dB, the ordinate is the symbol error rate SER, and the simulation experiment performs comparative analysis on four cases, namely, a single-user self-precoding machine S-AP, two-user self-precoding machines M-AP, a secure self-precoding machine SAP, and a confrontation secure self-precoding machine ASAP. It can be seen that S-AP and M-AP occupy the upper and lower performance limits, respectively. From the angle of beam forming precision, the beam alignment of the S-AP is more accurate, and the received signal power is high; M-AP beam forming is a trade-off between users and therefore the performance of each user is degraded. Compared with the M-AP, the SAP performance curve is characterized in that the directivity of beam forming is definitely biased to a legal user because another user is regarded as an eavesdropper, so that the performance is slightly improved compared with the M-AP, and meanwhile, the SER of Eve is almost equal to the probability of signal blind guess detection in the scene, so that the safety is well guaranteed; the ASAP performance curve, whether Bob or Eve, has some boost after the counterlearning compared to SAP, but Bob' S performance boost is more meaningful because it can be an order of magnitude boost after 10dB compared to SAP and approaches the best SER performance that the self-precoders can provide, i.e., the S-AP curve.
FIG. 5 is a received signal constellation simulation result diagram of a legal user Bob and an eavesdropping user Eve under two safety training frames in the security self-precoding machine optimization method based on countermeasure learning and the embodiment of the invention;
in fig. 5, the abscissa represents the real part of the received constellation, the ordinate represents the imaginary part of the received constellation, and the simulation experiment performs comparative analysis on four received constellations: wherein "Bob w AL" represents the reception constellation of Bob under the ASAP framework; "Eve w AL" represents the receive constellation for Eve under the ASAP framework; "Bob wo AL" represents the reception constellation of Bob under the SAP framework; "Eve wo AL" represents the receive constellation of Eve under SAP framework; it can be seen that the received constellation diagram before counterlearning is similar to the traditional PSK star diagram modulation, only occupies the phase information of the signal, the average distance of the symbol clustering center is smaller, and the error probability of symbol detection is increased; the constellation diagram after the counterstudy is similar to the traditional QAM modulation, and under the condition of the same power constraint, the amplitude and the phase information of the signal can be simultaneously utilized, so that the two-dimensional space is better occupied to control the intersymbol interference. Meanwhile, it can be seen that the received constellation of Eve is relatively chaotic in both SAP and ASAP cases, especially in Eve without counterlearning, the received constellation shrinks to a bunch of noise points in a very small range space, with the worst SER performance. Therefore, the counterlearning has important significance for designing a self-precoding machine with higher safety and a more reasonable safe modulation constellation.
The above detailed description is intended to illustrate the objects, aspects and advantages of the present invention, and it should be understood that the above detailed description is only exemplary of the present invention and is not intended to limit the scope of the present invention, and any modifications, equivalents, improvements and the like within the spirit and principle of the present invention should be included in the scope of the present invention.

Claims (6)

1. The safety self-precoding machine optimization method based on the counterstudy is characterized in that: comprises the following steps of (a) preparing a solution,
step one, setting system parameters of an MIMO communication system based on a self-encoder frame, wherein the system parameters comprise: the number of antennas M, N of the transmitter Alice, the legal user Bob and the eavesdropping user EveBAnd NEBit information R of each symbol, length J of symbol sequence, channel multipath quantity L and channel parameter alphal，θl，
Figure FDA0003492490370000011
Distribution of (d), transmit power constraint p, signal-to-noise ratio SNR; setting a neural network model structure, training/testing dataset parameters, and training hyper-parameters, the training hyper-parameters comprising: the selected optimizer, the training round Epoch, and the sample length of each round Batch Size;
step two, building an MIMO communication system supporting multi-user, large-scale multi-antenna and multi-stream data transmission based on a training framework of a depth self-coding machine, designing a multi-user average SER as a loss function, and training a transmitter network model formed by cascading a modulation module and a pre-coding module; the trained model can realize the reliable transmission of multiple users in the system under limited transmitting power through space beam forming;
step three, designing a new safety loss function by introducing a fuzzy matrix P aiming at the eavesdropping user according to the multi-user and multi-stream MIMO self-precoding machine model built in the step two
Figure FDA0003492490370000012
Guiding model training by using a new safety loss function, giving a self-precoding machine safety attribute, generating a new safety constellation diagram, and ensuring that a receiving end of a legal user Bob can finish symbol detection, and a receiving end of an eavesdropping user Eve cannot finish symbol detection correctly;
step four: an antagonistic learning mechanism is introduced, and a target loss function aiming at the eavesdropping user is designed by combining the safety loss function of the step 3.1
Figure FDA0003492490370000013
Dividing the whole self-precoding machine into two parts of links, namely a legal link Main Chain and an eavesdropping link Eve Chain, wherein the Main Chain comprises an Alice transmitter and a receiver network of Bob, the Eve Chain comprises an Eve receiver network, and designing two parts of iterative confrontation training algorithms based on the pre-training model in the second step to obtain a confrontation safety self-coder model ASAP;
step five: according to the anti-security self-pre-coder ASAP obtained by training in the step four, in a new security transmission scene, by acquiring a small amount of channel samples, the step four is continuously executed to finely adjust the model, and then the security self-pre-coder model with updated parameters is used for carrying out combined optimization of modulation and pre-coding on confidential information to obtain a signal to be transmitted, which has confidential property for a target eavesdropping user; and the interception user Eve can only obtain the symbol detection performance of the blind guess level while the legal user Bob has high reliability, so that the safe transmission is realized.
2. The method of claim 1 for secure self-precoding machine optimization based on antagonistic learning, characterized in that: the implementation method of the step one is that,
setting system parameters of the MIMO communication system based on the self-encoder framework, wherein the system parameters comprise: the number of antennas M, N of the transmitter Alice, the legal user Bob and the eavesdropping user EveBAnd NEBit information R of each symbol, length J of symbol sequence, channel multipath quantity L and channel parameter alphal，θl，
Figure FDA0003492490370000014
Distribution of (d), transmit power constraint p, signal-to-noise ratio SNR; setting a neural network model structure, training/testing dataset parameters, and training hyper-parameters, the training hyper-parameters comprising: selected optimizer, training round Epoch, sample length per round Batch Size.
3. The method of secure self-precoding machine optimization based on antagonistic learning of claim 1, characterized by: the implementation method of the second step is that,
step 2.1: designing a transmitting terminal neural network, wherein the transmitting terminal neural network comprises a network structure of a signal modulation module and a space pre-coding module;
the transmitting symbol corresponding to the jth secret information after the antenna mapping of the transmitting end is Xj：
Figure FDA0003492490370000021
Wherein the content of the first and second substances,
Figure FDA0003492490370000022
and
Figure FDA0003492490370000023
respectively representing a modulation module and a precoding module; m isjIs secret information to be transmitted from a predetermined limited set of secret information
Figure FDA0003492490370000024
Obtaining; the modulation symbol is output after passing through the modulation neural network module
Figure FDA0003492490370000025
sjEstimating channel parameters with the transmitting end
Figure FDA0003492490370000026
Combining and inputting precoding neural network module together
Figure FDA0003492490370000027
Carrying out precoding operation on the modulation symbols to obtain a precoded signal Xj(ii) a In the same way as above, the first and second,
Figure FDA0003492490370000028
xjThe method can process and send in parallel to realize multi-stream signal transmission of the MIMO system; training and testing all parameters in the data set, including channel parameters and signal parameters, by adopting a real part and imaginary part separated mode to represent, namely, all channels and signals in the system are characterized as a real matrix;
all transmitting end networks adopt a fully-connected neural network FCNN and a modulation module neural network
Figure FDA00034924903700000223
The calculation process for the secret information sequence m is represented as follows:
Figure FDA0003492490370000029
wherein the content of the first and second substances,
Figure FDA00034924903700000210
and
Figure FDA00034924903700000211
separately representing modulation module neural networks
Figure FDA00034924903700000212
The activation function, the weight vector and the offset vector of the g-th network; then all symbol sequences s after modulation and the estimated channel parameters estimated by the transmitting end
Figure FDA00034924903700000213
Combining to obtain new training sample
Figure FDA00034924903700000214
Neural network as spatial precoding module
Figure FDA00034924903700000215
The input of (1); spatial precoding module neural network
Figure FDA00034924903700000216
The calculation process of the matrix U after combining the modulation symbols and the channel is expressed as follows:
Figure FDA00034924903700000217
wherein the content of the first and second substances,
Figure FDA00034924903700000218
and
Figure FDA00034924903700000219
separately representing modulation module neural networks
Figure FDA00034924903700000220
T ofthAn activation function, a weight vector and an offset vector of a layer network; to limit the transmit signal power | X | ≦ p,
Figure FDA00034924903700000221
t of (A)thThe layer is designed as a power constraint layer and adopts a self-defined activation function
Figure FDA00034924903700000222
The following were used:
Figure FDA0003492490370000031
wherein | X | represents the F-norm of matrix X, and p represents the maximum transmit power; therefore, a normalized signal to be transmitted mapped to the antenna port is obtained through the step 2.1;
step 2.2: designing a receiving end neural network, which comprises a received signal detection module and a probability mapping module;
the legal user Bob and the eavesdropping user Eve are regarded as two legal users in the step, and two receiver models with the same network structure are built; j ththSignal X with normalized powerjRespectively reaches a receiving end through respective MIMO channels of Bob and Eve, and the j-th channel received by Bob and EvethA signal YB,jAnd YE,jRespectively, as follows:
YB,j＝HBXj+nB (5)
YE,j＝HEXj+nE (6)
wherein the content of the first and second substances,
Figure FDA0003492490370000032
and
Figure FDA0003492490370000033
representing additive white gaussian noise;
at the receiving end, the receiver networks of Bob and Eve adopt the same network structure, which is respectively expressed as:
Figure FDA0003492490370000034
and
Figure FDA0003492490370000035
j-th of receiver recovery for Bob and EvethThe secret information is represented as:
Figure FDA0003492490370000036
Figure FDA0003492490370000037
wherein alpha isB,θB,
Figure FDA0003492490370000038
Respectively representing the channel parameters estimated by the Bob receiving end; alpha is alphaE,θE,
Figure FDA0003492490370000039
Respectively representing the channel parameters estimated by an Eve receiving end;
the last layer of the receiving end neural network adopts a Softmax activation function to respectively output prediction probability vectors P of Bob and EveBAnd PE(ii) a The probability vector represents
Figure FDA00034924903700000310
And
Figure FDA00034924903700000311
the predicted secret is a set of secrets
Figure FDA00034924903700000312
The probability corresponding to a certain secret information;
step 2.3: designing an average cross-entropy loss function for multiple users according to the classified cross-entropy loss function
Figure FDA00034924903700000313
Updating the parameters of the self-precoding machine model by adopting a reverse gradient descent strategy;
average cross entropy loss function of legal user Bob and eavesdropping user Eve
Figure FDA00034924903700000314
The design is as follows:
Figure FDA0003492490370000041
wherein, Pj,mOne-hot encoding matrix representing transmitted secret information sequence m
Figure FDA0003492490370000042
Row j, column m;
Figure FDA0003492490370000043
and
Figure FDA0003492490370000044
probability prediction matrices for receiver networks representing Bob and Eve, respectively
Figure FDA0003492490370000045
And
Figure FDA0003492490370000046
row j, column m;
Figure FDA0003492490370000047
expressed in discrete parameter sets
Figure FDA0003492490370000048
Next, the average of the loss function calculated under the number of batch size data samples, where the batch size represents the length of each batch of training samples sent into the neural network;
an optimizer based on Tensorflow deep learning framework for minimizing the average cross entropy loss function
Figure FDA0003492490370000049
The confidential information recovered by the receiver sides of Bob and Eve is processed through an end-to-end neural network, an unsupervised training process is achieved, signals recovered by a receiving end are consistent with signals of a sending end, reliable transmission of the confidential information is achieved, both Eve and Bob can obtain an optimal receiver under the channel scene, meanwhile, the trained Eve is considered to be an optimal eavesdropper based on self-encoding training, and subsequent safety design is conducted on the basis of the eavesdropper of the Eve user.
4. The method of claim 1 for secure self-precoding machine optimization based on antagonistic learning, characterized in that: the implementation method of the third step is that,
step 3.1: similar to the second step, in order to realize the safe transmission of the physical layer signal, a new safety loss function is designed aiming at the fuzzy matrix P of the eavesdropping user
Figure FDA00034924903700000410
Is represented as follows:
Figure FDA00034924903700000411
wherein, Pj,m，
Figure FDA00034924903700000412
And
Figure FDA00034924903700000413
the meaning is in accordance with formula (9); introducing an ambiguity matrix P to confuse eavesdropping subscriber receivers, Pj,mThe elements representing the jth row and mth column of the blur matrix P, which is written as follows:
Figure FDA00034924903700000414
according to the principle of a cross entropy loss function, along with the progress of a training process, a prediction probability matrix of an Eve receiver of an eavesdropping user is closer to a fuzzy matrix P, so that the probabilities of Eve judging that received symbols belong to a certain class are consistent, therefore, symbol detection cannot be performed in a striving way, Bob can perform correct detection, and secure transmission of confidential information is realized;
step 3.2: based on the new safety loss function designed in step 3.1
Figure FDA0003492490370000051
Carrying out safety training under the condition that the parameters of a fixed eavesdropping user receiver are not changed, and training to obtain SAP;
step 3.2.1: firstly, determining the total training times N, and initializing N to be 1; reading the parameters of the pre-trained self-encoder model in the step two, including the initialization network parameter phi of the transmitter AliceAAnd receiver initialization network parameters of Bob and Eve
Figure FDA0003492490370000052
And
Figure FDA0003492490370000053
step 3.2.2: initializing channel parameters of training samples and corresponding one-hot coded labels, and reading a training data set;
step 3.2.3: determining a training hyperparameter: the method comprises the steps of optimizing the learning rate of an optimizer, training turns, the length of each batch of samples, and the division ratio of a training data set and a verification data set;
step 3.2.4: starting training, updating all network model parameters phi by using an Adam optimizer based on a loss function (10)A′，
Figure FDA0003492490370000054
And
Figure FDA0003492490370000055
step 3.2.5: n is n + 1; and ending the training until N is equal to N.
5. The method of secure self-precoding machine optimization based on antagonistic learning of claim 1, characterized by: the implementation method of the fourth step is that,
step 4.1: an antagonistic learning mechanism is introduced, and a target loss function aiming at the eavesdropping user is designed by combining the safety loss function of the step 3.1
Figure FDA0003492490370000056
Figure FDA0003492490370000057
Wherein, Pj,m，
Figure FDA0003492490370000058
Is identical to the expression in formula (10); the purpose of designing the loss function is to continue to optimize the virtual eavesdropping receiver for the secure transmitter to obtain a lower SER value after the secure transmitter training of step 3.2 is completed;
step 4.2: the whole self-precoding machine is divided into two parts, namely a legal link Main Chain and an eavesdropping link Eve Chain, wherein the Main Chain comprises an Alice transmitter and a receiver network of Bob, the Eve Chain comprises an Eve receiver network, and an iterative confrontation training algorithm of the two parts is designed based on the pre-training model of the step two to obtain an confrontation safety self-precoding machine ASAP;
step 4.2.1: firstly, determining the total iteration number N; determining a training hyperparameter: the method comprises the steps of optimizing the learning rate of an optimizer, training turns, the length of each batch of samples, and the division ratio of a training data set and a verification data set;
step 4.2.2: is iteration turn n? (ii) a If n is 1, reading the model parameters of the pre-trained self-precoding machine in the step two, including the initialized network parameter phi of the transmitter AliceAAnd receiver initialization network parameters of Bob and Eve
Figure FDA0003492490370000059
And
Figure FDA0003492490370000061
if n ≠ 1, the model parameters of n-1 times of updating are read,
Figure FDA0003492490370000062
step 4.2.3: initializing channel parameters of training samples
Figure FDA0003492490370000063
Reading a training data set with the corresponding one-hot coded label P;
step 4.2.4: setting the training times epoch _1, freezing the network parameters of Eve Chain, training the network model of Main Chain according to the loss function (10), and updating the parameters
Figure FDA0003492490370000064
Step 4.2.5: setting the training times epoch _2, freezing the network parameters of the Main Chain, training the network model of the Eve Chain according to the loss function (12), and updating the parameters
Figure FDA0003492490370000065
Step 4.2.6: n is n + 1; returning to step 4.2.2, 4.2.3, 4.2.4 and 4.2.5 are continuously executed until N is equal to N, and the training is finished.
6. The method of secure self-precoding machine optimization based on antagonistic learning of claim 1, characterized by: the implementation method of the fifth step is that,
according to the anti-security self-pre-coder ASAP obtained by training in the step four, in a new security transmission scene, by acquiring a small amount of channel samples, the step four is continuously executed to finely adjust the model, and then the security self-pre-coder model with updated parameters is used for carrying out combined optimization of modulation and pre-coding on confidential information to obtain a signal to be transmitted, which has confidential property for a target eavesdropping user; and the interception user Eve can only obtain the symbol detection performance of the blind guess level while the legal user Bob has high reliability, so that the safe transmission is realized.
CN202210112026.XA 2022-01-27 2022-01-27 Security self-precoding machine optimization method based on counterstudy Active CN114553274B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210112026.XA CN114553274B (en) 2022-01-27 2022-01-27 Security self-precoding machine optimization method based on counterstudy

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210112026.XA CN114553274B (en) 2022-01-27 2022-01-27 Security self-precoding machine optimization method based on counterstudy

Publications (2)

Publication Number Publication Date
CN114553274A true CN114553274A (en) 2022-05-27
CN114553274B CN114553274B (en) 2023-03-31

Family

ID=81674271

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210112026.XA Active CN114553274B (en) 2022-01-27 2022-01-27 Security self-precoding machine optimization method based on counterstudy

Country Status (1)

Country Link
CN (1) CN114553274B (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117835246A (en) * 2023-12-29 2024-04-05 山东大学 Task-oriented privacy semantic communication method
WO2024077597A1 (en) * 2022-10-14 2024-04-18 华为技术有限公司 Wireless physical layer secure communication method, and communication apparatus

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200160170A1 (en) * 2018-11-20 2020-05-21 Bank Of America Corporation System and method for incremental learning through state-based real-time adaptations in neural networks
US20200175371A1 (en) * 2018-12-04 2020-06-04 Bank Of America Corporation System and method for self constructing deep neural network design through adversarial learning
CN112468258A (en) * 2019-09-06 2021-03-09 河海大学常州校区 Full-duplex end-to-end automatic encoder communication system and anti-eavesdropping method thereof
CN112924749A (en) * 2021-02-04 2021-06-08 西安电子科技大学 Unsupervised counterstudy electromagnetic spectrum abnormal signal detection method
CN113591905A (en) * 2021-06-17 2021-11-02 中山大学 Deep learning time sequence clustering method based on double-layer attention mechanism and counterstudy
CN113972939A (en) * 2021-09-09 2022-01-25 浙江大学 Antenna system precoding method and device based on double time scales and deep learning

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200160170A1 (en) * 2018-11-20 2020-05-21 Bank Of America Corporation System and method for incremental learning through state-based real-time adaptations in neural networks
US20200175371A1 (en) * 2018-12-04 2020-06-04 Bank Of America Corporation System and method for self constructing deep neural network design through adversarial learning
CN112468258A (en) * 2019-09-06 2021-03-09 河海大学常州校区 Full-duplex end-to-end automatic encoder communication system and anti-eavesdropping method thereof
CN112924749A (en) * 2021-02-04 2021-06-08 西安电子科技大学 Unsupervised counterstudy electromagnetic spectrum abnormal signal detection method
CN113591905A (en) * 2021-06-17 2021-11-02 中山大学 Deep learning time sequence clustering method based on double-layer attention mechanism and counterstudy
CN113972939A (en) * 2021-09-09 2022-01-25 浙江大学 Antenna system precoding method and device based on double time scales and deep learning

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
CHAD A. COLE; CHOWDHURY SHAHRIAR; T. CHARLES CLANCY: "An Anti-jam Communications Technique via Spatial Hiding Precoding", 《 2014 IEEE MILITARY COMMUNICATIONS CONFERENCE》 *
刘广灿，曹宇，许家铭，徐波: "基于对抗正则化的自然语言推理", 《自动化学报》 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2024077597A1 (en) * 2022-10-14 2024-04-18 华为技术有限公司 Wireless physical layer secure communication method, and communication apparatus
CN117835246A (en) * 2023-12-29 2024-04-05 山东大学 Task-oriented privacy semantic communication method

Also Published As

Publication number Publication date
CN114553274B (en) 2023-03-31

Similar Documents

Publication Publication Date Title
O'Shea et al. Physical layer deep learning of encodings for the MIMO fading channel
CN106059972B (en) A kind of Modulation Identification method under MIMO correlated channels based on machine learning algorithm
Wang et al. Pilot-assisted channel estimation and signal detection in uplink multi-user MIMO systems with deep learning
CN110557177A (en) DenseNet-based hybrid precoding method in millimeter wave large-scale MIMO system
CN109194378B (en) Physical layer safety wave beam shaping method based on linear neural network
Younis et al. Information-theoretic treatment of space modulation MIMO systems
CN110880950A (en) Safe transmission method for artificial noise auxiliary vector disturbance precoding in MIMO system
Shamasundar et al. A DNN architecture for the detection of generalized spatial modulation signals
Yun et al. Deep artificial noise: Deep learning-based precoding optimization for artificial noise scheme
CN114553274B (en) Security self-precoding machine optimization method based on counterstudy
Wang et al. Deep learning for joint MIMO detection and channel decoding
Hua et al. Signal detection in uplink pilot-assisted multi-user MIMO systems with deep learning
Ye et al. Bilinear convolutional auto-encoder based pilot-free end-to-end communication systems
CN109936399A (en) A kind of insincere junction network antenna selecting method based on deep neural network
Zhu et al. Index modulation for fluid antenna-assisted MIMO communications: System design and performance analysis
Xue et al. End-to-end learning for uplink MU-SIMO joint transmitter and non-coherent receiver design in fading channels
Saikia et al. Signal detection in GSM-based in-band full-duplex communication using DNN
CN109462429B (en) Beam domain modulation device and method of large-scale multiple-input multiple-output millimeter wave system
Tato et al. Neural network aided computation of generalized spatial modulation capacity
CN111431567A (en) Millimeter wave large-scale beam space MIMO system
Huang et al. Deep learning based parallel detector for MIMO systems
Wang et al. Learning to modulate for non-coherent MIMO
Vía et al. Blind decoding of MISO-OSTBC systems based on principal component analysis
Mohamed et al. Supervised learning classifier based transmit antenna selection for SM-MIMO system
Aghdam et al. Low complexity precoding for MIMOME wiretap channels based on cut-off rate

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