CN109413068A - A kind of wireless signal encryption method based on antithesis GAN - Google Patents

A kind of wireless signal encryption method based on antithesis GAN Download PDF

Info

Publication number
CN109413068A
CN109413068A CN201811268894.7A CN201811268894A CN109413068A CN 109413068 A CN109413068 A CN 109413068A CN 201811268894 A CN201811268894 A CN 201811268894A CN 109413068 A CN109413068 A CN 109413068A
Authority
CN
China
Prior art keywords
wireless signal
network
size
encryption
decryption
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
CN201811268894.7A
Other languages
Chinese (zh)
Other versions
CN109413068B (en
Inventor
陈晋音
郑海斌
成凯回
熊晖
林翔
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.)
Zhejiang University of Technology ZJUT
Original Assignee
Zhejiang University of Technology ZJUT
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 Zhejiang University of Technology ZJUT filed Critical Zhejiang University of Technology ZJUT
Priority to CN201811268894.7A priority Critical patent/CN109413068B/en
Publication of CN109413068A publication Critical patent/CN109413068A/en
Application granted granted Critical
Publication of CN109413068B publication Critical patent/CN109413068B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L63/00—Network architectures or network communication protocols for network security
    • H04L63/04—Network architectures or network communication protocols for network security for providing a confidential data exchange among entities communicating through data packet networks
    • H04L63/0428—Network architectures or network communication protocols for network security for providing a confidential data exchange among entities communicating through data packet networks wherein the data content is protected, e.g. by encrypting or encapsulating the payload

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Security & Cryptography (AREA)
  • Computer Hardware Design (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Mobile Radio Communication Systems (AREA)
  • Facsimile Transmission Control (AREA)

Abstract

本发明公开了一种基于对偶GAN的无线信号加密方法，包括：利用无线信号加密模型将原始无线信号加密成加密信号图像；利用无线信号解密模型将加密信号图像解密成解密无线信号；所述无线信号加密模型和无线信号解密模型通过以下模型训练体系训练得到，所述模型训练体系包括：无线信号加密网络Ag、无线信号加密判别网络Ad、无线信号解密网络Bg、无线信号解密判别网络Bd、无线信号分类网络C，对上述模型训练体系采用对Ag和C的预训练和对Ag、Ad、Bg、Bd以及C的联合训练，以确定无线信号加密模型和无线信号解密模型。该方法可以有效将无线信号加密为图片，提高信息传输的隐蔽性。

The invention discloses a wireless signal encryption method based on dual GAN, comprising: using a wireless signal encryption model to encrypt an original wireless signal into an encrypted signal image; using a wireless signal decryption model to decrypt the encrypted signal image into a decrypted wireless signal; The signal encryption model and the wireless signal decryption model are obtained through the training of the following model training system, the model training system includes: wireless signal encryption network Ag, wireless signal encryption discrimination network Ad, wireless signal decryption network Bg, wireless signal decryption discrimination network Bd, wireless signal encryption The signal classification network C adopts the pre-training of Ag and C and the joint training of Ag, Ad, Bg, Bd and C for the above model training system to determine the wireless signal encryption model and the wireless signal decryption model. The method can effectively encrypt wireless signals into pictures and improve the concealment of information transmission.

Description

A kind of wireless signal encryption method based on antithesis GAN
Technical field
The invention belongs to the security fields research fields of the deep learning algorithm of artificial intelligence field and data, and in particular to A kind of wireless signal encryption method based on antithesis GAN.
Background technique
Currently, deep learning has been widely used in the fields such as target detection, image detection, data generation.It generates Fighting network (GAN) is one of technology relatively advanced in current deep learning, generates confrontation network and passes through in frame (at least) Two modules: the mutual game of model (Generative Model) and discrimination model (Discriminative Model) are generated Study generates preferable output.According to different network structures, a series of GAN network has been derived: DCGAN, WGAN, WGAN-GP、LSGAN-BEGAN、DUAL-GAN。
Wherein, generation confrontation network (DUAL-GAN) of antithesis GAN be on the basis of original generation fights network into One step is extended to two generation to intercouple confrontation networks, wherein there are two generator and two arbiters, with text and figure For the mutual translation of piece, wherein character translation can be picture by first generator Ag, corresponding also one A arbiter Ad can identify the true and false of picture.Meanwhile a corresponding generator Bg is constructed, photo can be translated as Whether text, same generator Bg also have a corresponding arbiter Bd accurate to identify text.
Current existing wireless signal encryption method is broadly divided into symmetric cryptography and divides symmetric cryptosystem.Symmetric cryptography is adopted It is that the encryption of asymmetric encoding technological document and decryption use identical key.Asymmetric encryption needs two keys: open Key and private cipher key carry out encrypting another key being decrypted using one of key.Existing encryption method exists It is easily attacked, the problems such as encryption mode is single.Under normal circumstances, wireless signal encryption technology is by the way of digital encryption, this It is bigger that kind mode calculates cost, and is easy to be trapped utilization.
Summary of the invention
The object of the present invention is to provide a kind of wireless signal encryption methods based on antithesis GAN, pass through the generation of antithesis GAN It fights network and generates picture corresponding to wireless signal, and classified using wireless signal disaggregated model C, thus to wireless Signal data realizes the process of encryption.The wireless signal encryption method improves the confidentiality and peace of wireless signal transmission process Quan Xing.
For achieving the above object, the present invention the following technical schemes are provided:
A kind of wireless signal encryption method based on antithesis GAN, comprising:
Using wireless signal Encryption Model by original wireless signal encryption at coded signal image；
Using wireless signal decrypted model by coded signal image decryption at decryption wireless signal；
The wireless signal Encryption Model and wireless signal decrypted model are obtained by the training of following model training system, institute Stating model training system includes:
Wireless signal refined net Ag, for encrypting to original wireless signal, input is original wireless signal, nothing The decryption wireless signal of line decrypted signal network B g output, exports as coded signal image；
Wireless signal encryption differentiates network A d, and input is normal color image, wireless signal refined net Ag output Coded signal image exports as the judgement to normal color image and coded signal image as a result, wherein utilizing wireless signal When encryption differentiates that the output feedback training wireless signal encryption of network A d differentiates network A d parameter, the class calibration of normal color image Justice be 1, the category of coded signal image is defined as 0, using wireless signal encryption differentiate network A d output feedback training without When line signal encryption differentiates network A g parameter, coded signal image is used only, and the category of coded signal image is defined as at this time 1；
Wireless signal decryption network Bg, for coded signal image to be decrypted, input is normal color image, nothing The coded signal image of line signal encryption network A g output exports to decrypt wireless signal；
Wireless signal decryption differentiates network B d, and input is original wireless signal, wireless signal decryption network Bg output Wireless signal is decrypted, is exported as the judgement to original wireless signal and decryption wireless signal as a result, wherein utilizing wireless signal When decryption differentiates that the output feedback training wireless signal decryption of network B d differentiates network B d parameter, the class calibration of original wireless signal Justice be 1, the category for decrypting wireless signal is defined as 0, using wireless signal decryption differentiate network B d output feedback training without When line decrypted signal network B g parameter, decryption wireless signal is used only, and the category for decrypting wireless signal at this time is defined as 1；
Wireless signal sorter network C, input are coded signal image and the life of wireless signal refined net Ag output At the category of the wireless signal of the coded signal image, the i.e. modulation type of wireless signal, the classification knot for coded signal is exported Fruit, each classification results correspond to the modulation type of original wireless signal；
Training process are as follows:
Stage one: pre-training, using original wireless signal and corresponding wireless signal modulation type category to wireless signal Refined net Ag and wireless signal sorter network C carries out pre-training；
Stage two: retraining, firstly, fixed wireless signal encryption differentiate network A d, wireless signal decryption differentiate network B d, The parameter of wireless signal sorter network C, the parameter of joint training wireless signal refined net Ag and wireless signal decryption network Bg； Then, the parameter of fixed wireless signal encryption network A g, wireless signal decryption network Bg, wireless signal sorter network C, joint instruction Practice the parameter that wireless signal encryption differentiates network A d and wireless signal decryption differentiates network B d；Finally, fixed wireless signal encryption Differentiate the parameter of network A d, wireless signal decryption network Bg, wireless signal decryption differentiation network B d, joint training wireless signal adds The parameter of close network A g and wireless signal sorter network C；By above three joint training, until wireless signal refined net Ag Network A d is differentiated with wireless signal encryption, and it is assorted equal that wireless signal decryption network Bg and wireless signal decryption differentiate that network B d realization is received Weighing apparatus, training are completed, and trained wireless signal refined net Ag is wireless signal Encryption Model, trained wireless signal decryption Network B g is wireless signal decrypted model.
The graded combination Training strategy provided through the invention, the wireless signal refined net Ag after completing training can make It is signal pattern by wireless signal data encryption for a private cipher key of wireless signal；Wireless signal decryption network Bg can be with Coded signal image decryption is wireless signal data by another key as encrypted signal pattern, and by introducing Wireless signal sorter network C, classifies to the wireless signal for using different modulating type to obtain, and wireless signal can be made to encrypt Network A g learns the characteristic of division attribute of wireless signal modulation type simultaneously in the training process, can be to different modulating type Wireless signal is preferably encrypted.
The effect that the wireless signal encryption method based on antithesis GAN has are as follows: wireless signal can be passed through into wireless communication Number Encryption Model projects image feature space from signal characteristic space, picture signal is encrypted as automatically, to escape wireless communication The identification of number identifying system.The concealment of radio signal propagation, the safety of information transmission can be improved by this encryption method Property.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below There is attached drawing needed in technical description to do simply to introduce, it should be apparent that, the accompanying drawings in the following description is only this Some embodiments of invention for those of ordinary skill in the art, can be with root under the premise of not making the creative labor Other accompanying drawings are obtained according to these attached drawings.
Fig. 1 is the structural schematic diagram for the model training system that embodiment provides；
Fig. 2 is the structural schematic diagram for the wireless signal refined net Ag that embodiment provides；
Fig. 3 is the structural schematic diagram for the wireless signal decryption network Bg that embodiment provides；
Fig. 4 is the structural schematic diagram that the wireless signal encryption that embodiment provides differentiates network A d；
Fig. 5 is the structural schematic diagram that the wireless signal decryption that embodiment provides differentiates network B d；
Fig. 6 is the structural schematic diagram for the wireless signal sorter network C that embodiment provides；
12 class wireless signals are encrypted to picture in the classification method training process that Fig. 7 (a)~7 (l) provides for embodiment Visualization result schematic diagram.
Specific embodiment
To make the objectives, technical solutions, and advantages of the present invention more comprehensible, with reference to the accompanying drawings and embodiments to this Invention is described in further detail.It should be appreciated that the specific embodiments described herein are only used to explain the present invention, And the scope of protection of the present invention is not limited.
For the confidentiality for improving wireless signal transmission, present embodiments provides a kind of wireless signal based on antithesis GAN and add Decryption method, specific steps include: by original wireless signal encryption using wireless signal Encryption Model into coded signal image；It utilizes Wireless signal decrypted model is by coded signal image decryption at decryption wireless signal.
The wireless signal Encryption Model and wireless signal decrypted model are obtained using model training system as shown in Figure 1 , specific model training system includes five modules: being respectively wireless signal refined net Ag, wireless signal encryption differentiation net Network Ad, wireless signal decryption network Bg, wireless signal decryption differentiate network B d and wireless signal sorter network C.
Wireless signal refined net Ag is mainly used for encrypting wireless signal, i.e., is to add by original wireless signal encryption Close signal pattern.Wireless signal refined net Ag is the neural network being made of LSTM unit, convolutional layer and full articulamentum, knot Structure is as shown in Figure 2, comprising: the wireless signal being originally inputted is having a size of [512,2], wherein when the sampling of 512 expression wireless signals Between point, 2 indicate that the characteristic value of wireless signal each time points, training process are minimum using the training of most small quantities of gradient descent method Generally be taken as 64 per a batch of wireless signal data sample number in batch, obtained after LSTM unit having a size of [512, 128] characteristic layer, wherein 512 correspond to original time point, the feature vector that 128 corresponding each time points are calculated makes The characteristic layer having a size of 128 is obtained with full connection, the characteristic layer having a size of 128 is obtained using full connection, is obtained using full articulamentum To the characteristic layer having a size of 64*64*3, this feature layer is deformed to obtain having a size of [64,64,3] using reshape function Characteristic layer, wherein 64 respectively correspond the length and width of characteristic layer, the depth of 3 character pair layers, using having a size of [5,5,64] Convolution module and the maximum pond module for being 2 having a size of [2,2] step-length obtain the characteristic layer having a size of [32,32,64], use [5,5,128] convolution module and the maximum pond module for being 2 having a size of [2,2] step-length are obtained having a size of [16,16,128] Characteristic layer, the maximum pond module for being 2 using the convolution module having a size of [5,5,256] and having a size of [2,2] step-length obtain ruler The very little characteristic layer for [8,8,256], the maximum for being 2 using the convolution module having a size of [5,5,512] and having a size of [2,2] step-length Pond module obtains the characteristic layer having a size of [4,4,512], using having a size of [5,5,512] convolution module and having a size of [2, 2] the maximum pond module that step-length is 2 obtains the characteristic layer having a size of [2,2,512], uses the convolution having a size of [5,5,512] Module and having a size of [2,2] step-length be 2 maximum pond module obtain the characteristic layer having a size of [1,1,512], using having a size of The warp volume module of [5,5,512] obtains the characteristic layer having a size of [2,2,512], the characteristic layer of [2,2,512] before combining, The characteristic layer having a size of [2,2,1024] is obtained, the warp volume module of [5,5,512] is used to obtain the characteristic layer of [4,4,512], The characteristic layer of [4,4,512] before joint obtains the characteristic layer of [4,4,1024], uses the warp volume module of [5,5,256] The characteristic layer of [8,8,256] is obtained, the characteristic layer of [8,8,256] before combining obtains the characteristic layer of [8,8,512], uses The warp volume module of [5,5,128] obtains the characteristic layer of [16,16,128], and the characteristic layer of [16,16,128] before combining obtains To the characteristic layer of [16,16,256], the warp volume module of [5,5,64] is used to obtain the characteristic layer of [32,32,64], before joint [32,32,64] characteristic layer, obtain the characteristic layer of [32,32,128], use the warp volume module of [5,5,3] obtain [64, 64,3] characteristic layer, what is obtained is the picture of corresponding ImageNet64, and picture a length of 64, width 64 includes RGB tri- Channel.
Wireless signal decryption network Bg is mainly used for that coded signal image is decrypted, i.e., by coded signal image decryption To decrypt wireless signal.Wireless signal decryption network Bg is the neural network being made of convolutional layer and full articulamentum, and structure is such as Shown in Fig. 3, comprising: the picture size being originally inputted is [64,64,3], uses convolution module and size having a size of [5,5,64] The maximum pond module for being 2 for [2,2] step-length obtains the characteristic layer having a size of [32,32,64], using having a size of [5,5,128] Convolution module and having a size of [2,2] step-length be 2 maximum pond module obtain making having a size of the characteristic layer of [16,16,128] With having a size of [5,5,256] convolution module and having a size of [2,2] step-length be 2 maximum pond module obtain having a size of [8,8, 256] characteristic layer, the maximum pond module for being 2 using the convolution module having a size of [5,5,1024] and having a size of [2,2] step-length The characteristic layer having a size of [4,4,1024] is obtained, using the convolution module having a size of [5,5,2048] and having a size of [2,2] step-length The characteristic layer having a size of [2,2,2048] is obtained for 2 maximum pond module, uses the convolution module having a size of [5,5,4096] It obtains using full link block having a size of the characteristic layer of [1, Isosorbide-5-Nitrae 096] for 2 maximum pond module with having a size of [2,2] step-length The characteristic layer having a size of 16384 (512*32) is obtained, obtains the characteristic layer having a size of 4096 (512*8) using full link block, The characteristic layer having a size of 1024 (512*2) is obtained using full link block, is become its shape using reshape function [512,2] obtain wireless signal corresponding to picture, wherein 512 corresponding sampling time points, the feature of 2 corresponding each time points Value.
Wireless signal encryption differentiates that network A d is mainly used for carrying out two classification to normal color image and coded signal image, Judge the category of normal color image and coded signal image, the class for defining normal color image is designated as true (real), encryption The class of signal pattern is designated as false (fake), is encrypted by classification results feedback training (feedback training) wireless signal Network A g, the image for generating encryption are more nearly true value.Wireless signal encryption differentiates that network A d is by convolutional layer and Quan Lian The neural network of layer composition being connect, structure is as shown in Figure 4, comprising: the picture size of input is [64,64,3], use [5,5, 64] convolution module and the maximum pond module for being 2 having a size of [2,2] step-length obtain the characteristic layer of [32,32,64], use [5, 5,128] convolution module and the maximum pond module for being 2 having a size of [2,2] step-length obtain the characteristic layer of [16,16,128], make The maximum pond module for being 2 with the convolution module of [5,5,256] and having a size of [2,2] step-length obtains the characteristic layer of [8,8,256], It uses the convolution module of [5,5,512] and obtains the feature of [4,4,512] having a size of [2,2] step-length for 2 maximum pond module Layer uses the convolution module of [5,5,1] and obtains the feature of [1,1,1] having a size of [4,4] step-length for 4 maximum pond module Layer, the matrix of obtained not normalized confidence value identical with shape 0 or 1 do the coded signal that cross entropy calculates Ag output The gap of image and normal color image.Wireless signal encryption differentiates that the input of network A d is the normal color of ImageNet64 The mixing for the coded signal image that image and Ag encryption generate, exports as the judgement to normal color image and coded signal image As a result.
Wireless signal decryption differentiates that network B d is mainly used for carrying out two classification to original wireless signal and decryption wireless signal, Judge original wireless signal and decrypt the category of wireless signal, the class for defining original wireless signal is designated as true (real) and decryption The class of wireless signal is designated as false (fake), and passes through classification results feedback training (feedback training) wireless signal solution Close network B g, the wireless signal for generating decryption are more nearly true value.Wireless signal decryption differentiates that network B d is mono- by LSTM The neural network that member, convolutional layer and full articulamentum are constituted, structure are as shown in Figure 5, comprising: the wireless signal sample being originally inputted Having a size of [512,2], the characteristic layer having a size of [512,64] is obtained after LSTM unit, wherein 512 when corresponding to original Between point, the feature vector that 64 corresponding each time points are calculated obtains the characteristic layer having a size of 64 using full connection, using complete Connection obtains the characteristic layer having a size of 1, and obtained not normalized confidence value and 0 or 1 is done cross entropy calculating wireless signal and added The gap of original wireless signal and decryption wireless signal that close network A g is generated.Wireless signal decryption differentiates the input of network B d It is decrypted into the mixing of decryption wireless signal for normal wireless signal and Ag, exports as to normal wireless signal and decryption wireless signal Judging result.
Wireless signal sorter network C is mainly used for classifying to wireless signal, inputs as wireless signal refined net Ag The coded signal image of output, exports the classification results for coded signal, finely tunes training (fine tune by feedback Training) Ag, it is more accurate to wireless signal refined net Ag to make.Wireless signal sorter network C is by convolutional layer and Quan Lian The neural network of layer composition is connect, specific structure is as shown in Figure 6, comprising: the dimension of picture of input is [64,64,3], is used Convolution kernel having a size of [5,5,64] and the maximum pond module for being 2 having a size of [2,2] step-length are obtained having a size of [32,32,64] Characteristic layer, using having a size of [5,5,128] convolution kernel and having a size of [2,2] step-length be 2 maximum pond module obtain ruler The very little characteristic layer for [16,16,128], the maximum for being 2 using the convolution kernel having a size of [5,5,256] and having a size of [2,2] step-length Pond module obtains the characteristic layer having a size of [8,8,256], using the convolution kernel having a size of [5,5,512] and having a size of [2,2] The maximum pond module that step-length is 2 obtains the characteristic layer having a size of [4,4,512], obtains the spy having a size of 1024 using full connection Layer is levied, full articulamentum is used to obtain (being directed to comprising 12 kinds of wireless signal modulation types having a size of 12 characteristic layer as output 12 classification tasks), the class of obtained not normalized confidence value original wireless signal corresponding with coded signal image Mark does cross entropy and calculates distance.
The design object of above-mentioned model training system are as follows:
(a) original wireless signal is encrypted automatically by network A g, which adds respectively as wireless signal The close input sample for differentiating network A d and wireless signal sorter network C, and Ag is fed back to by the output of Ad and C, adjust Ag's Inner parameter.Wherein Ad is by determining the picture sample of Ag encryption generation and the distribution distance of true normal sample, to adjust Ag Parameter；C is by predicting category at a distance from true category, to adjust Ag parameter.Eventually by the multilateral Game mistake of Ag and Ad and C Journey guarantees the stability for the picture sample that Ag encryption generates, and is trained with the encryption picture generated to C that obtaining can be right The wireless signal Encryption Model that the wireless signal data of different modulating type are more preferably encrypted；
(b) it is automatically signal data to image data decryption by network B g, is decrypted as wireless signal and differentiate network B d Input sample adjust Bg inner parameter by the input feedback of Bd to Bg.The wherein signal that Bd is generated by determining Bg decryption The distribution distance of sample and authentic specimen, to adjust Bg parameter.
In model training system, the network structure of wireless signal sorter network C is related with the complexity of data set；And The network structure of Ag and Ad designs, related with the complexity of C network and data set；Bg and Bd network structure design, with number It is related according to the complexity of collection.In order to realize better cipher round results, and prevent in the training process generation model collapse It bursts, be better achieved the encryption and decryption of wireless model, present invention uses the methods of joint training to have carried out accordingly model Training.
Specifically, the training of above-mentioned model training system is broadly divided into two stages of pre-training and retraining, detailed process Are as follows:
The pre-training stage: wireless signal is encrypted using original wireless signal and corresponding wireless signal modulation type category Network A g and wireless signal sorter network C carries out pre-training.When training, fixed wireless signal encryption differentiates network A d, wireless communication Number decryption differentiate network B d and wireless signal decryption network Bg parameter, trained epochs=N1, i.e. training data are set Collection is used N1 times；The input of wireless signal refined net Ag is original wireless signal, is exported as encrypted coded signal figure Picture, the input of wireless signal sorter network C are the coded signal image of wireless signal refined net Ag output, are exported as to encryption The category of the debug-type of original wireless signal corresponding to signal pattern is predicted；Wireless signal refined net Ag and nothing at this time Disaggregated model Ag-C classifies to the modulation type of wireless signal to line Modulation recognition network C as a whole, wireless signal Refined net Ag is equivalent to the characteristic extracting module of disaggregated model Ag-C, and wireless signal sorter network C is equivalent to disaggregated model Ag- The categorization module of C.
Specifically, original signal data collection x is inputtedsignal, corresponding picture Ag (x is generated by Ag networksignal), By Ag (xsignal) be input in disaggregated model C, 10 epochs of training.
The retraining stage, using joint training strategy, detailed process are as follows:
(1) fixed wireless signal encryption differentiates that network A d, wireless signal decryption differentiate network B d, wireless signal sorter network The parameter of C, by original wireless signal xsignal, wireless signal decryption network Bg output decryption wireless signal Bg (ximage) conduct The input of wireless signal refined net Ag, the parameter of training Ag, the coded signal image for exporting it are more nearly normal color Image, so that the ciphering process of wireless signal is realized, so that the propagation that encrypted information is safer；And by normal color figure As ximage, wireless signal refined net Ag output coded signal image Ag (xsignal) as wireless signal decryption network Bg's Input, the parameter of training Bg, the decryption wireless signal for exporting it are more nearly original wireless signal, realize that encryption is wireless with this The decrypting process of signal；
The optimization aim of the process indicates are as follows:
Wherein, xAgIndicate the coded signal image of wireless signal refined net Ag output, xBgIndicate that wireless signal decrypts net The decryption wireless signal of network Bg output, xAg~pAg indicates xAgSample the output from wireless signal refined net Ag, xBg~pBg table Show xBgThe output from wireless signal decryption network Bg is sampled, pAg and pBg respectively indicate the probability distribution of Ag and Bg output, Ad (xAg) indicate that wireless signal encryption differentiates network A d to xAgDifferentiation probability, Bd (xBg) indicate that wireless signal decryption differentiates network Bd is to xBgDifferentiation probability, E () indicate cross entropy expectation；
(2) parameter of fixed wireless signal encryption network A g, wireless signal decryption network Bg, wireless signal sorter network C, The coded signal image Ag (x that wireless signal refined net Ag is exportedsignal) and normal color imageBlended data, Encrypt the input for differentiating network A d as wireless signal, the parameter of training Ad can distinguish coded signal image and normal Color image；And the decryption wireless signal Bg (x for exporting wireless signal decryption network Bgimage) and original wireless signal Blended data, the input for differentiating network B d is decrypted as wireless signal, it is wireless can to distinguish decryption for the parameter of training Bd Signal and original wireless signal；
The optimization aim of the process indicates are as follows:
Wherein, Pdata indicates raw data set,Indicate normal signal sampleFrom original letter Number Pdata；Indicate normal image dataFrom raw image data Pdata；
(3) fixed wireless signal encryption differentiates that network A d, wireless signal decryption network Bg, wireless signal decryption differentiate network The parameter of Bd, by original wireless signal xsignal, wireless signal decryption network Bg output decryption wireless signal Bg (ximage) conduct The input of wireless signal refined net Ag, and the coded signal image Ag (x that wireless signal refined net Ag is exportedsignal) and Corresponding true category y (namely original wireless signal xsignalCorresponding modulation type category) as wireless signal classification net The input of network C, the parameter of coorinated training wireless signal refined net Ag and wireless signal sorter network C；
The optimization aim of the process are as follows:
Wherein, y indicates the modulation type category of original wireless signal,Indicate normal wireless sample of signalFrom wireless signal data set, C () indicates wireless signal sorter network C to the classification knot of wireless signal modulation type for sampling Fruit；
Wireless signal classification mould network C is introduced, is classified to the wireless signal for using different modulating type to obtain, energy Enough make wireless signal refined net Ag in the training process while learning the characteristic of division attribute of wireless signal modulation type, it can be with It is preferably encrypted for the modulation type of wireless signal.
(4) step (1)~(3) are repeated, until wireless signal refined net Ag and wireless signal encryption differentiate network A d, nothing Line decrypted signal network B g and wireless signal decryption differentiate that network B d realizes Nash Equilibrium, i.e. wireless signal refined net Ag and nothing Line style number encryption differentiates that both network A d game tends to balance, and wireless signal decryption network Bg and wireless signal decryption differentiate network The game of both Bd tends to balance, training cut-off, and trained wireless signal refined net Ag is wireless signal Encryption Model, instruction The wireless signal decryption network Bg perfected is wireless signal decrypted model.
When training, the loss function being related to is respectively as follows:
The loss function Ag_loss of wireless signal refined net Ag are as follows:
Wherein, λBFor introduce parameter, for balance B_loss withCodomain range so that B_loss withCodomain it is close, xAg~pAg indicates xAgIt samples from wireless signal refined net Ag Output, B_loss indicate normal color image and coded signal image mean square error, form are as follows:
B_loss=MSE (image, Ag (xsignal))
Wherein, image indicates the normal color image on ImageNet64, the normal color image and image Ag (xsignal) it is the consistent two images of pattern, MSE () indicates mean square error function；
The loss function Bg_loss of wireless signal decryption network Bg are as follows:
Wherein, λAFor introduce parameter, for balance A_loss withCodomain range so that A_loss withCodomain it is close, A_loss indicates original wireless signal and decrypts wireless signal Mean square error, form are as follows:
A_loss=MSE (signal, Bg (ximage))
Wherein, signal indicates original wireless signal；
Wireless signal encryption differentiates the loss function Ad_loss of network A d are as follows:
Wireless signal decryption differentiates the loss function Bd_loss of network B d are as follows:
The loss function C_loss of wireless signal sorter network C are as follows:
The calculating and optimization of the 5 above loss functions are converted the calculating of following 3 loss functions by the present embodiment:
G_loss=Ag_loss+Bg_loss
D_loss=Ad_loss+Bd_loss
C_loss
By the calculating to above 3 loss functions, each network is updated by backpropagation using Adam optimizer Parameter.
Have when realizing Nash Equilibrium:
Through the above training method based on minimax theorem (Minimax theorem), Ag and Bg, Ad are realized With the competition Game Relationship of Bd, C tripartite, the target of Ag is quickly to encrypt a large amount of wireless signal sample, can be connect as far as possible Nearly authentic specimen distribution, while Ad can be deceived；The target of Ad is to distinguish the picture sample and true figure that Ag is generated as far as possible Piece sample；The target of C be correctly classify as far as possible Ag generation picture, thereby realize wireless signal to the encrypted of picture Journey；The target of Bg is to carry out fast decryption to encrypted signal pattern, and revert to original signal, the target of Bd as far as possible It is the recovery degree distinguished the sample of signal of original normal signal sample and decryption as far as possible, and further increase Bg decryption, Thereby realize the decrypting process of signal.
Training strategy in this way, the wireless signal more adjunction of the wireless signal encryption picture and Bg decryption that generate Ag Nearly corresponding true value, such Ag can serve as a private cipher key of wireless signal, wireless signal are encrypted as picture； Image data is translated as wireless signal data by another key that Bg can serve as wireless signal, and wireless by introducing Modulation recognition MODEL C classifies to the wireless signal for using different modulating type to obtain, and can make wireless signal Encryption Model Ag learns the characteristic of division attribute of wireless signal modulation type simultaneously in the training process, can be to different types of wireless signal Preferably encrypted.
Specific experiment:
Data set basic condition includes: that (a) wireless signal data have 312000 training samples and 156000 test specimens This, each sample-size is the matrix of 512*2, and sample value range is (- 6,6).Verifying collection is to take out from test sample at random Take 10% sample size；(b) data set can be divided into 12 classes according to modulation type, every class equal part, there is every class in training set 26000 samples, every class has 13000 samples in test set；(c) state of signal-to-noise: for the 500*26*12=in test set 156000 samples, every one kind have 26 kinds of signal-to-noise ratio (- 20 to 30 all even numbers), different signal-to-noise ratio data collection in every class sample Distribution is substantially are as follows: the sample that 500 signal-to-noise ratio are -20, the sample that 500 signal-to-noise ratio are -18, the sample that 500 signal-to-noise ratio are -16 This ... ... the sample that 500 signal-to-noise ratio are 28, the sample that 500 signal-to-noise ratio are 30.And for the data distribution feelings of training set Condition is 1000 in the number of samples of one of classification signal-to-noise ratio, then the sample size of training set is 1000*26*12 =312000.(d) the simple reduction process divided by 6 has all been carried out to all wireless signal sample datas, has been input to facilitating In model.
Model training system of the above-mentioned training set to above-mentioned building is trained, trained wireless signal encryption is obtained Model and wireless signal decrypted model.And the sample in test set is input in wireless signal Encryption Model, it exports as such as Fig. 7 (a)~7 the corresponding coded signal image of 12 class wireless signals shown in (l), because the batch (batch) of input is 64, every part Figure contains 64 figures, these pictures are manually difficult its content for being included respectively, illustrates this method adding for wireless signal It is close to produce a desired effect.
Technical solution of the present invention and beneficial effect is described in detail in above-described specific embodiment, Ying Li Solution is not intended to restrict the invention the foregoing is merely presently most preferred embodiment of the invention, all in principle model of the invention Interior done any modification, supplementary, and equivalent replacement etc. are enclosed, should all be included in the protection scope of the present invention.

Claims (8)

1.一种基于对偶GAN的无线信号加密方法，包括：1. A wireless signal encryption method based on dual GAN, comprising: 利用无线信号加密模型将原始无线信号加密成加密信号图像；Using the wireless signal encryption model to encrypt the original wireless signal into an encrypted signal image; 利用无线信号解密模型将加密信号图像解密成解密无线信号；Use the wireless signal decryption model to decrypt the encrypted signal image into a decrypted wireless signal; 所述无线信号加密模型和无线信号解密模型通过以下模型训练体系训练得到，所述模型训练体系包括：The wireless signal encryption model and the wireless signal decryption model are obtained through the following model training system training, and the model training system includes: 无线信号加密网络Ag，用于对原始无线信号进行加密，其输入为原始无线信号、无线信号解密网络Bg输出的解密无线信号，输出为加密信号图像；The wireless signal encryption network Ag is used to encrypt the original wireless signal, the input is the original wireless signal, the decrypted wireless signal output by the wireless signal decryption network Bg, and the output is the encrypted signal image; 无线信号加密判别网络Ad，其输入为正常彩色图像、无线信号加密网络Ag输出的加密信号图像，输出为对正常彩色图像和加密信号图像的判定结果，其中在利用无线信号加密判别网络Ad的输出反馈训练无线信号加密判别网络Ad参数时，正常彩色图像的类标定义为1，加密信号图像的类标定义为0，在利用无线信号加密判别网络Ad的输出反馈训练无线信号加密网络Ag参数时，仅使用加密信号图像，且此时加密信号图像的类标定义为1；The wireless signal encryption discrimination network Ad, its input is the normal color image, the encrypted signal image output by the wireless signal encryption network Ag, and the output is the judgment result of the normal color image and the encrypted signal image, wherein the output of the wireless signal encryption discrimination network Ad is used. When training the Ad parameters of the wireless signal encryption discrimination network by feedback, the class label of the normal color image is defined as 1, and the class label of the encrypted signal image is defined as 0. When using the output feedback of the wireless signal encryption discrimination network Ad to train the wireless signal encryption network Ag parameters , only the encrypted signal image is used, and the class label of the encrypted signal image is defined as 1; 无线信号解密网络Bg，用于对加密信号图像进行解密，其输入为正常彩色图像、无线信号加密网络Ag输出的加密信号图像，输出为解密无线信号；The wireless signal decryption network Bg is used to decrypt the encrypted signal image, the input is the normal color image, the encrypted signal image output by the wireless signal encryption network Ag, and the output is the decrypted wireless signal; 无线信号解密判别网络Bd，其输入为原始无线信号、无线信号解密网络Bg输出的解密无线信号，输出为对原始无线信号和解密无线信号的判定结果，其中在利用无线信号解密判别网络Bd的输出反馈训练无线信号解密判别网络Bd参数时，原始无线信号的类标定义为1，解密无线信号的类标定义为0，在利用无线信号解密判别网络Bd的输出反馈训练无线信号解密网络Bg参数时，仅使用解密无线信号，且此时解密无线信号的类标定义为1；The wireless signal decryption and identification network Bd, whose input is the original wireless signal and the decrypted wireless signal output by the wireless signal decryption network Bg, and the output is the judgment result of the original wireless signal and the decrypted wireless signal, wherein the output of the wireless signal decryption and identification network Bd is used. When feedback training the wireless signal decryption to determine the network Bd parameters, the class label of the original wireless signal is defined as 1, and the class label of the decrypted wireless signal is defined as 0. When using the output of the wireless signal decryption and discrimination network Bd to feed back the training wireless signal to decrypt the network Bg parameters , only the decrypted wireless signal is used, and the class label of the decrypted wireless signal is defined as 1; 无线信号分类网络C，其输入为无线信号加密网络Ag输出的加密信号图像以及生成该加密信号图像的无线信号的类标，即无线信号的调制类型，输出为加密信号的分类结果，每个分类结果对应原始无线信号的调制类型；The wireless signal classification network C, whose input is the encrypted signal image output by the wireless signal encryption network Ag and the class label of the wireless signal that generates the encrypted signal image, that is, the modulation type of the wireless signal, and the output is the classification result of the encrypted signal, each classification The result corresponds to the modulation type of the original wireless signal; 训练过程为：The training process is: 阶段一：预训练，采用原始无线信号和对应的无线信号调制类型类标对无线信号加密网络Ag和无线信号分类网络C进行预训练；Stage 1: Pre-training, using the original wireless signal and the corresponding wireless signal modulation type label to pre-train the wireless signal encryption network Ag and the wireless signal classification network C; 阶段二：重训练，首先，固定无线信号加密判别网络Ad、无线信号解密判别网络Bd、无线信号分类网络C的参数，联合训练无线信号加密网络Ag与无线信号解密网络Bg的参数；然后，固定无线信号加密网络Ag、无线信号解密网络Bg、无线信号分类网络C的参数，联合训练无线信号加密判别网络Ad与无线信号解密判别网络Bd的参数；最后，固定无线信号加密判别网络Ad、无线信号解密网络Bg、无线信号解密判别网络Bd的参数，联合训练无线信号加密网络Ag与无线信号分类网络C的参数；通过上述三个联合训练，直到无线信号加密网络Ag和无线信号加密判别网络Ad，无线信号解密网络Bg和无线信号解密判别网络Bd实现纳什均衡，训练完成，训练好的无线信号加密网络Ag为无线信号加密模型，训练好的无线信号解密网络Bg为无线信号解密模型。Stage 2: Retraining. First, the parameters of the wireless signal encryption and discrimination network Ad, the wireless signal decryption and discrimination network Bd, and the wireless signal classification network C are fixed, and the parameters of the wireless signal encryption network Ag and the wireless signal decryption network Bg are jointly trained; The parameters of the wireless signal encryption network Ag, the wireless signal decryption network Bg, and the wireless signal classification network C are jointly trained for the wireless signal encryption and discrimination network Ad and the parameters of the wireless signal decryption and discrimination network Bd; Decrypt the parameters of the network Bg and the wireless signal decryption and discrimination network Bd, and jointly train the parameters of the wireless signal encryption network Ag and the wireless signal classification network C; The wireless signal decryption network Bg and the wireless signal decryption discrimination network Bd realize Nash equilibrium, and the training is completed. The trained wireless signal encryption network Ag is the wireless signal encryption model, and the trained wireless signal decryption network Bg is the wireless signal decryption model. 2.如权利要求1所述的基于对偶GAN的无线信号加密方法，其特征在于，所述无线信号加密网络的结构包括：2. The wireless signal encryption method based on dual GAN as claimed in claim 1, wherein the structure of the wireless signal encryption network comprises: 原始输入的无线信号尺寸为[512,2]，其中512表示无线信号的采样时间点，2表示无线信号每个时间点的特征值，训练过程采用最小批梯度下降方法训练，最小批中每一批次的无线信号数据样本个数一般取为64个，经过LSTM单元后得到尺寸为[512,128]的特征层，其中512对应于原始的时间点，128对应每个时间点计算得到的特征向量，使用全连接得到尺寸为128的特征层，使用全连接得到尺寸为128的特征层，使用全连接层得到尺寸为64*64*3的特征层，使用reshape函数对该特征层进行变形得到尺寸为[64,64,3]的特征层，其中64分别对应特征层的长和宽，3对应特征层的深度，使用尺寸为[5,5,64]的卷积模块和尺寸为[2,2]步长为2的最大池化模块得到尺寸为[32,32,64]的特征层，使用[5,5,128]的卷积模块和尺寸为[2,2]步长为2的最大池化模块得到尺寸为[16,16,128]的特征层，使用尺寸为[5,5,256]的卷积模块和尺寸为[2,2]步长为2的最大池化模块得到尺寸为[8,8,256]的特征层，使用尺寸为[5,5,512]的卷积模块和尺寸为[2,2]步长为2的最大池化模块得到尺寸为[4,4,512]的特征层，使用尺寸为[5,5,512]的卷积模块和尺寸为[2,2]步长为2的最大池化模块得到尺寸为[2,2,512]的特征层，使用尺寸为[5,5,512]的卷积模块和尺寸为[2,2]步长为2的最大池化模块得到尺寸为[1,1,512]的特征层，使用尺寸为[5,5,512]的反卷积模块得到尺寸为[2,2,512]的特征层，联合之前的[2,2,512]的特征层，得到尺寸为[2,2,1024]的特征层，使用[5,5,512]的反卷积模块得到[4,4,512]的特征层，联合之前的[4,4,512]的特征层，得到[4,4,1024]的特征层，使用[5,5,256]的反卷积模块得到[8,8,256]的特征层，联合之前的[8,8,256]的特征层，得到[8,8,512]的特征层，使用[5,5,128]的反卷积模块得到[16,16,128]的特征层，联合之前的[16,16,128]的特征层，得到[16,16,256]的特征层，使用[5,5,64]的反卷积模块得到[32,32,64]的特征层，联合之前的[32,32,64]的特征层，得到[32,32,128]的特征层，使用[5,5,3]的反卷积模块得到[64,64,3]的特征层，得到的即为所对应的ImageNet64的图片，图片长为64，宽为64，包含RGB三通道。The size of the original input wireless signal is [512, 2], where 512 represents the sampling time point of the wireless signal, and 2 represents the eigenvalue of each time point of the wireless signal. The number of batches of wireless signal data samples is generally 64. After passing through the LSTM unit, a feature layer of size [512, 128] is obtained, of which 512 corresponds to the original time point, and 128 corresponds to the feature vector calculated at each time point. Use full connection to get a feature layer of size 128, use full connection to get a feature layer of size 128, use fully connected layer to get a feature layer of size 64*64*3, use the reshape function to deform the feature layer to get the size of The feature layers of [64, 64, 3], where 64 correspond to the length and width of the feature layer, respectively, and 3 correspond to the depth of the feature layer, use a convolution module of size [5, 5, 64] and size [2, 2] ] A max-pooling module with stride 2 results in a feature layer of size [32, 32, 64], using a convolutional module of [5, 5, 128] and a max-pooling module of size [2, 2] with stride 2 The module gets a feature layer of size [16, 16, 128], using a convolution module of size [5, 5, 256] and a max pooling module of size [2, 2] with stride 2 to get a size of [8, 8, 256] The feature layer of size [4, 4, 512] is obtained using a convolution module of size [5, 5, 512] and a max pooling module of size [2, 2] with stride 2 to obtain a feature layer of size [5, 4, 512] ,5,512] convolutional module of size [2,2] and max-pooling module of size [2,2] stride 2 to obtain feature layers of size [2,2,512], using convolutional module of size [5,5,512] and size Get a feature layer of size [1, 1, 512] for a max pooling module with stride 2 for [2, 2], and use a deconvolution module of size [5, 5, 512] to get features of size [2, 2, 512] layer, combine the previous feature layers of [2, 2, 512] to obtain a feature layer of size [2, 2, 1024], and use the deconvolution module of [5, 5, 512] to obtain a feature layer of [4, 4, 512], and jointly The previous feature layer of [4, 4, 512] gets the feature layer of [4, 4, 1024], and the deconvolution module of [5, 5, 256] is used to obtain the feature layer of [8, 8, 256], combined with the previous [8, 8,256] feature layer, get the feature layer of [8,8,512], use the deconvolution module of [5,5,128] to get the feature layer of [16,16,128], combine the previous feature layer of [16,16,128], get The feature layer of [16, 16, 256], the deconvolution module of [5, 5, 64] is used to obtain the feature layer of [32, 32, 64], and the feature layer of [32, 32, 64] before is combined to obtain [ 32, 32, 128] feature layer, using [5, 5 ,3] The deconvolution module obtains the feature layer of [64, 64, 3], and the obtained image is the corresponding ImageNet64 picture, the picture length is 64, the width is 64, and contains RGB three channels. 3.如权利要求1所述的基于对偶GAN的无线信号加密方法，其特征在于，所述无线信号解密网络Bg的结构包括：3. The wireless signal encryption method based on dual GAN as claimed in claim 1, wherein the structure of the wireless signal decryption network Bg comprises: 原始输入的图像尺寸为[64,64,3]，使用尺寸为[5,5,64]的卷积模块和尺寸为[2,2]步长为2的最大池化模块得到尺寸为[32,32,64]的特征层，使用尺寸为[5,5,128]的卷积模块和尺寸为[2,2]步长为2的最大池化模块得到尺寸为[16,16,128]的特征层，使用尺寸为[5,5,256]的卷积模块和尺寸为[2,2]步长为2的最大池化模块得到尺寸为[8,8,256]的特征层，使用尺寸为[5,5,1024]的卷积模块和尺寸为[2,2]步长为2的最大池化模块得到尺寸为[4,4,1024]的特征层，使用尺寸为[5,5,2048]的卷积模块和尺寸为[2,2]步长为2的最大池化模块得到尺寸为[2,2,2048]的特征层，使用尺寸为[5,5,4096]的卷积模块和尺寸为[2,2]步长为2的最大池化模块得到尺寸为[1,1,4096]的特征层，使用全连接模块得到尺寸为16384(512*32)的特征层，使用全连接模块得到尺寸为4096(512*8)的特征层，使用全连接模块得到尺寸为1024(512*2)的特征层，使用reshape函数将其的形状变为[512,2]，得到图片所对应的无线信号，其中512对应采样时间点，2对应每个时间点的特征值。The original input image size is [64, 64, 3], using a convolution module of size [5, 5, 64] and a max pooling module of size [2, 2] with stride 2 yields a size of [32] , 32, 64], using a convolution module of size [5, 5, 128] and a max pooling module of size [2, 2] with stride 2 to obtain a feature layer of size [16, 16, 128], Use a convolution module of size [5, 5, 256] and a max pooling module of size [2, 2] with stride 2 to obtain a feature layer of size [8, 8, 256], using size [5, 5, 1024] ] and a max-pooling module of size [2, 2] with stride 2 to obtain a feature layer of size [4, 4, 1024], using a convolution module of size [5, 5, 2048] and a max pooling module of size [2, 2] with stride 2 to get a feature layer of size [2, 2, 2048], using a convolution module of size [5, 5, 4096] and a size of [2 ,2] The maximum pooling module with step size 2 obtains a feature layer of size [1, 1, 4096], and the fully connected module is used to obtain a feature layer of size 16384 (512*32), and the fully connected module is used to obtain a size of 4096 (512*8) feature layer, use the fully connected module to get a feature layer with a size of 1024 (512*2), use the reshape function to change its shape to [512,2], and get the wireless signal corresponding to the picture, Among them, 512 corresponds to the sampling time point, and 2 corresponds to the feature value of each time point. 4.如权利要求1所述的基于对偶GAN的无线信号加密方法，其特征在于，所述无线信号加密判别网络Ad的结构包括：4. the wireless signal encryption method based on dual GAN as claimed in claim 1 is characterized in that, the structure of described wireless signal encryption discrimination network Ad comprises: 输入的图像尺寸为[64,64,3]，使用[5,5,64]的卷积模块和尺寸为[2,2]步长为2的最大池化模块得到[32,32,64]的特征层，使用[5,5,128]的卷积模块和尺寸为[2,2]步长为2的最大池化模块得到[16,16,128]的特征层，使用[5,5,256]的卷积模块和尺寸为[2,2]步长为2的最大池化模块得到[8,8,256]的特征层，使用[5,5,512]的卷积模块和尺寸为[2,2]步长为2的最大池化模块得到[4,4,512]的特征层，使用[5,5,1]的卷积模块和尺寸为[4,4]步长为4的最大池化模块得到[1,1,1]的特征层，得到的未归一化的置信度值与形状相同的0或1的矩阵做交叉熵计算Ag输出的加密信号图像与正常彩色图像的差距；The input image size is [64, 64, 3], using a convolution module of [5, 5, 64] and a max pooling module of size [2, 2] with stride 2 to get [32, 32, 64] The feature layer of [16, 16, 128] is obtained using the convolution module of [5, 5, 128] and the max pooling module of size [2, 2] and stride 2, using the convolution of [5, 5, 256] modules and a max pooling module of size [2, 2] with stride 2 to obtain feature layers of [8, 8, 256], using a convolution module of size [5, 5, 512] and stride 2 of size [2, 2] The max-pooling module of [4, 4, 512] is used to obtain feature layers of [4, 4, 512], and the convolution module of [5, 5, 1] and the max-pooling module of size [4, 4] and stride 4 are used to obtain [1, 1, 1], the obtained unnormalized confidence value and the matrix of 0 or 1 with the same shape do cross entropy to calculate the gap between the encrypted signal image output by Ag and the normal color image; 所述无线信号解密判别网络Bd的结构包括：The structure of the wireless signal decryption and discrimination network Bd includes: 原始输入的无线信号样本尺寸为[512,2]，经过LSTM单元后得到尺寸为[512,64]的特征层，其中512对应于原始的时间点，64对应每个时间点计算得到的特征向量，使用全连接得到尺寸为64的特征层，使用全连接得到尺寸为1的特征层，得到的未归一化的置信度值与0或1做交叉熵计算无线信号加密网络Ag生成的原始无线信号和解密无线信号的差距。The size of the original input wireless signal sample is [512, 2]. After passing through the LSTM unit, a feature layer of size [512, 64] is obtained, where 512 corresponds to the original time point, and 64 corresponds to the feature vector calculated at each time point. , using the full connection to obtain a feature layer of size 64, using the full connection to obtain a feature layer of size 1, and the obtained unnormalized confidence value and 0 or 1 for cross entropy calculation The original wireless signal generated by the wireless signal encryption network Ag The gap between signaling and decrypting wireless signals. 5.如权利要求1所述的基于对偶GAN的无线信号加密方法，其特征在于，所述无线信号分类网络C的结构包括：5. The wireless signal encryption method based on dual GAN as claimed in claim 1, wherein the structure of the wireless signal classification network C comprises: 输入的图片尺寸为[64,64,3]，使用尺寸为[5,5,64]的卷积核和尺寸为[2,2]步长为2的最大池化模块得到尺寸为[32,32,64]的特征层，使用尺寸为[5,5,128]的卷积核和尺寸为[2,2]步长为2的最大池化模块得到尺寸为[16,16,128]的特征层，使用尺寸为[5,5,256]的卷积核和尺寸为[2,2]步长为2的最大池化模块得到尺寸为[8,8,256]的特征层，使用尺寸为[5,5,512]的卷积核和尺寸为[2,2]步长为2的最大池化模块得到尺寸为[4,4,512]的特征层，使用全连接得到尺寸为1024的特征层，使用全连接层得到尺寸为12的特征层作为输出(针对于包含12种无线信号调制类型的12分类任务)，所得到的未归一化的置信度值与加密信号图像对应的原始无线信号的类标做交叉熵计算距离。The input image size is [64, 64, 3], using a convolution kernel of size [5, 5, 64] and a max pooling module of size [2, 2] with stride 2 to obtain a size of [32, 32, 64], using a convolution kernel of size [5, 5, 128] and a max pooling module of size [2, 2] with stride 2 to obtain a feature layer of size [16, 16, 128], using A convolution kernel of size [5, 5, 256] and a max pooling module of size [2, 2] with stride 2 yield feature layers of size [8, 8, 256], using volumes of size [5, 5, 512] The kernel and the max pooling module of size [2, 2] and stride 2 get a feature layer of size [4, 4, 512], use a fully connected layer to get a feature layer of size 1024, and use a fully connected layer to get a size of 12 As the output (for 12 classification tasks including 12 wireless signal modulation types), the obtained unnormalized confidence value and the class label of the original wireless signal corresponding to the encrypted signal image are used for cross entropy calculation distance. 6.如权利要求1所述的基于对偶GAN的无线信号加密方法，其特征在于，预训练的具体过程为：6. the wireless signal encryption method based on dual GAN as claimed in claim 1, is characterized in that, the concrete process of pre-training is: 固定无线信号加密判别网络Ad、无线信号解密判别网络Bd以及无线信号解密网络Bg的参数，设置训练的epochs＝N1，即训练数据集被使用N1次；无线信号加密网络Ag的输入为原始无线信号，输出为加密后的加密信号图像，无线信号分类网络C的输入为无线信号加密网络Ag输出的加密信号图像，输出为对加密信号图像所对应的原始无线信号的调试类型的类标预测；此时的无线信号加密网络Ag和无线信号分类网络C作为一个整体分类模型Ag-C对无线信号的调制类型进行分类，无线信号加密网络Ag相当于分类模型Ag-C的特征提取模块，无线信号分类网络C相当于分类模型Ag-C的分类模块。The parameters of the wireless signal encryption and discrimination network Ad, the wireless signal decryption and discrimination network Bd and the wireless signal decryption network Bg are fixed, and the training epochs=N1 is set, that is, the training data set is used N1 times; the input of the wireless signal encryption network Ag is the original wireless signal , the output is the encrypted encrypted signal image, the input of the wireless signal classification network C is the encrypted signal image output by the wireless signal encryption network Ag, and the output is the classmark prediction of the debugging type of the original wireless signal corresponding to the encrypted signal image; this The wireless signal encryption network Ag and the wireless signal classification network C as a whole classification model Ag-C classify the modulation type of the wireless signal, the wireless signal encryption network Ag is equivalent to the feature extraction module of the classification model Ag-C, and the wireless signal classification The network C is equivalent to the classification module of the classification model Ag-C. 7.如权利要求6所述的基于对偶GAN的无线信号加密方法，其特征在于，重训练的具体过程为：7. the wireless signal encryption method based on dual GAN as claimed in claim 6, is characterized in that, the concrete process of retraining is: (1)固定无线信号加密判别网络Ad、无线信号解密判别网络Bd、无线信号分类网络C的参数，将原始无线信号xsignal、无线信号解密网络Bg输出的解密无线信号Bg(ximage)作为无线信号加密网络Ag的输入，训练Ag的参数，使其输出的加密信号图像更加接近正常彩色图像；并将正常彩色图像ximage、无线信号加密网络Ag输出的加密信号图像Ag(xsignal)作为无线信号解密网络Bg的输入，训练Bg的参数，使其输出的解密无线信号更加接近原始无线信号，以此实现加密无线信号的解密过程；(1) The parameters of the wireless signal encryption and discrimination network Ad, the wireless signal decryption and discrimination network Bd, and the wireless signal classification network C are fixed, and the original wireless signal x signal and the decrypted wireless signal Bg (x image ) output by the wireless signal decryption network Bg are used as the wireless signal The input of the signal encryption network Ag, train the parameters of Ag to make the output encrypted signal image closer to the normal color image; use the normal color image x image and the encrypted signal image Ag(x signal ) output by the wireless signal encryption network Ag as the wireless The input of the signal decryption network Bg, the parameters of Bg are trained to make the decrypted wireless signal output closer to the original wireless signal, so as to realize the decryption process of the encrypted wireless signal; (2)固定无线信号加密网络Ag、无线信号解密网络Bg、无线信号分类网络C的参数，将无线信号加密网络Ag输出的加密信号图像Ag(xsignal)和正常彩色图像的混合数据，作为无线信号加密判别网络Ad的输入，训练Ad的参数，使其能够区分加密信号图像和正常彩色图像；并将无线信号解密网络Bg输出的解密无线信号Bg(ximage)和原始无线信号的混合数据，作为无线信号解密判别网络Bd的输入，训练Bd的参数，使其能够区分解密无线信号和原始无线信号；(2) Fix the parameters of the wireless signal encryption network Ag, the wireless signal decryption network Bg, and the wireless signal classification network C, and the encrypted signal image Ag(x signal ) output by the wireless signal encryption network Ag and the normal color image The mixed data is used as the input of the wireless signal encryption and discrimination network Ad, and the parameters of Ad are trained so that it can distinguish the encrypted signal image and the normal color image; and the decrypted wireless signal Bg(x image ) output by the wireless signal decryption network Bg wireless signal The mixed data is used as the input of the wireless signal decryption and discrimination network Bd, and the parameters of Bd are trained so that it can distinguish the decrypted wireless signal and the original wireless signal; (3)固定无线信号加密判别网络Ad、无线信号解密网络Bg、无线信号解密判别网络Bd的参数，将原始无线信号xsignal、无线信号解密网络Bg输出的解密无线信号Bg(ximage)作为无线信号加密网络Ag的输入，并将无线信号加密网络Ag输出的加密信号图像Ag(xsignal)和对应的真实类标y作为无线信号分类网络C的输入，协同训练无线信号加密网络Ag和无线信号分类网络C的参数；(3) The parameters of the wireless signal encryption and discrimination network Ad, the wireless signal decryption network Bg, and the wireless signal decryption and discrimination network Bd are fixed, and the original wireless signal x signal and the decrypted wireless signal Bg (x image ) output by the wireless signal decryption network Bg are used as the wireless signal The input of the signal encryption network Ag, and the encrypted signal image Ag(x signal ) output by the wireless signal encryption network Ag and the corresponding real class label y are used as the input of the wireless signal classification network C, and the wireless signal encryption network Ag and the wireless signal are jointly trained. the parameters of the classification network C; (4)重复步骤(1)～(3)，直到无线信号加密网络Ag和无线信号加密判别网络Ad，无线信号解密网络Bg和无线信号解密判别网络Bd实现纳什均衡，即无线信号加密网络Ag与无线型号加密判别网络Ad两者博弈趋于平衡，无线信号解密网络Bg与无线信号解密判别网络Bd两者的博弈趋于平衡，训练截止，训练好的无线信号加密网络Ag为无线信号加密模型，训练好的无线信号解密网络Bg为无线信号解密模型。(4) Repeat steps (1) to (3) until the wireless signal encryption network Ag and the wireless signal encryption discrimination network Ad, the wireless signal decryption network Bg and the wireless signal decryption discrimination network Bd achieve Nash equilibrium, that is, the wireless signal encryption network Ag and The game between the wireless model encryption and discrimination network Ad tends to be balanced, the game between the wireless signal decryption network Bg and the wireless signal decryption and discrimination network Bd tends to be balanced, and the training ends, and the trained wireless signal encryption network Ag is the wireless signal encryption model, The trained wireless signal decryption network Bg is a wireless signal decryption model. 8.如权利要求7所述的基于对偶GAN的无线信号加密方法，其特征在于，当实现纳什均衡时有：8. the wireless signal encryption method based on dual GAN as claimed in claim 7, is characterized in that, when realizing Nash equilibrium, have: 其中，xAg表示无线信号加密网络Ag输出的加密信号图像，xBg表示无线信号解密网络Bg输出的解密无线信号，xAg～pAg表示xAg采样自无线信号加密网络Ag的输出，xBg～pBg表示xBg采样自无线信号解密网络Bg的输出，pAg和pBg分别表示Ag和Bg输出的概率分布，Ad(xAg)表示无线信号加密判别网络Ad对xAg的判别概率，Bd(xBg)表示无线信号解密判别网络Bd对xBg的判别概率，E(·)表示交叉熵的期望；Pdata表示原始数据集，表示正常信号样本来自于原始信号数据Pdata；表示正常图像数据来自于原始图像数据Pdata；y表示原始无线信号的调制类型类标，表示正常无线信号样本采样自无线信号数据集，C(·)表示无线信号分类网络C对无线信号调制类型的分类结果。Among them, x Ag represents the encrypted signal image output by the wireless signal encryption network Ag, x Bg represents the decrypted wireless signal output by the wireless signal decryption network Bg, x Ag ~ pAg represents the output of x Ag sampled from the wireless signal encryption network Ag, x Bg ~ pBg represents the output of x Bg sampled from the wireless signal decryption network Bg, pAg and pBg represent the probability distribution of Ag and Bg outputs respectively, Ad(x Ag ) represents the discrimination probability of x Ag by the wireless signal encryption discrimination network Ad, Bd(x Bg ) represents the discrimination probability of the wireless signal decryption and discrimination network Bd for x Bg , E( ) represents the expectation of cross entropy; Pdata represents the original data set, represents a normal signal sample From the original signal data Pdata; Represents normal image data It comes from the original image data Pdata; y represents the modulation type label of the original wireless signal, Indicates normal wireless signal samples Sampled from the wireless signal data set, C(·) represents the classification result of the wireless signal modulation type by the wireless signal classification network C.
CN201811268894.7A 2018-10-29 2018-10-29 Wireless signal encryption method based on dual GAN Active CN109413068B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201811268894.7A CN109413068B (en) 2018-10-29 2018-10-29 Wireless signal encryption method based on dual GAN

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201811268894.7A CN109413068B (en) 2018-10-29 2018-10-29 Wireless signal encryption method based on dual GAN

Publications (2)

Publication Number Publication Date
CN109413068A true CN109413068A (en) 2019-03-01
CN109413068B CN109413068B (en) 2021-01-26

Family

ID=65469565

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201811268894.7A Active CN109413068B (en) 2018-10-29 2018-10-29 Wireless signal encryption method based on dual GAN

Country Status (1)

Country Link
CN (1) CN109413068B (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111047006A (en) * 2019-10-28 2020-04-21 浙江工业大学 An Adversarial Attack Defense Model and Application Based on Dual Generative Network
CN111654368A (en) * 2020-06-03 2020-09-11 电子科技大学 A Key Generation Method Based on Deep Learning Generative Adversarial Networks
CN114710371A (en) * 2022-06-08 2022-07-05 深圳市乐凡信息科技有限公司 Method, device, equipment and storage medium for safely signing electronic data

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106951919A (en) * 2017-03-02 2017-07-14 浙江工业大学 A Realization Method of Flow Velocity Monitoring Based on Adversarial Generative Network
CN107220600A (en) * 2017-05-17 2017-09-29 清华大学深圳研究生院 A kind of Picture Generation Method and generation confrontation network based on deep learning
CN108573479A (en) * 2018-04-16 2018-09-25 西安电子科技大学 Face Image Deblurring and Restoration Method Based on Dual Generative Adversarial Network

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106951919A (en) * 2017-03-02 2017-07-14 浙江工业大学 A Realization Method of Flow Velocity Monitoring Based on Adversarial Generative Network
CN107220600A (en) * 2017-05-17 2017-09-29 清华大学深圳研究生院 A kind of Picture Generation Method and generation confrontation network based on deep learning
CN108573479A (en) * 2018-04-16 2018-09-25 西安电子科技大学 Face Image Deblurring and Restoration Method Based on Dual Generative Adversarial Network

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
SURYAWANSHI S B: "A triple-key Chaotic neural network for", 《 INTERNATIONAL JOURNAL OF ENGINEERING SCIENCES& EMERGING TECHNOLOGIES》 *
王万良: "生成式对抗网络研究进展", 《通信学报》 *

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111047006A (en) * 2019-10-28 2020-04-21 浙江工业大学 An Adversarial Attack Defense Model and Application Based on Dual Generative Network
CN111047006B (en) * 2019-10-28 2023-04-21 浙江工业大学 Dual generation network-based anti-attack defense model and application
CN111654368A (en) * 2020-06-03 2020-09-11 电子科技大学 A Key Generation Method Based on Deep Learning Generative Adversarial Networks
CN111654368B (en) * 2020-06-03 2021-10-08 电子科技大学 A Key Generation Method Based on Deep Learning Generative Adversarial Networks
CN114710371A (en) * 2022-06-08 2022-07-05 深圳市乐凡信息科技有限公司 Method, device, equipment and storage medium for safely signing electronic data

Also Published As

Publication number Publication date
CN109413068B (en) 2021-01-26

Similar Documents

Publication Publication Date Title
Zheng et al. Spectrum interference-based two-level data augmentation method in deep learning for automatic modulation classification
Xie et al. A generalizable model-and-data driven approach for open-set RFF authentication
Maleki et al. Breast cancer diagnosis from histopathology images using deep neural network and XGBoost
Miao et al. A central difference graph convolutional operator for skeleton-based action recognition
CN103838744B (en) A kind of method and device of query word demand analysis
CN109800785B (en) Data classification method and device based on self-expression correlation
CN109413068A (en) A kind of wireless signal encryption method based on antithesis GAN
Guo et al. Self-supervised GANs with similarity loss for remote sensing image scene classification
CN111079594B (en) Video action classification and identification method based on double-flow cooperative network
Zhang et al. Positional context aggregation network for remote sensing scene classification
Luo et al. Beyond universal attack detection for continuous-variable quantum key distribution via deep learning
CN109408649A (en) Slag piece intelligent analysis method and system based on textural characteristics and machine learning
CN109995501A (en) A kind of side Multiple Channel Analysis method and device, terminal and computer readable storage medium
CN113516199A (en) Image data generation method based on differential privacy
Venkateswarulu et al. DeepExplain: enhancing deepfake detection through transparent and explainable AI model
CN111008940A (en) Image enhancement method and device
Zhao et al. Hybrid domain meta-learning network for face forgery detection and localization in deepfakes
Zhang et al. HSiPu2-a new human physical fitness action dataset for recognition and 3D reconstruction evaluation
Chu et al. SecureAS: a vulnerability assessment system for deep neural network based on adversarial examples
Le et al. Generating high-fidelity cybersecurity data with generative adversarial networks
Abbasnejad et al. Deep Lipschitz networks and dudley GANs
CN113486932B (en) Convolutional neural network steganalysis-oriented optimization method
Peng et al. Multi-Scale Convolutional Neural Networks optimized by elite strategy dung beetle optimization algorithm for encrypted traffic classification
Zia et al. Improving synthetic media generation and detection using generative adversarial networks
Zheng et al. Hardening RGB-D object recognition systems against adversarial patch attacks

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
OL01 Intention to license declared
OL01 Intention to license declared