CN109413068B - Wireless signal encryption method based on dual GAN - Google Patents
Wireless signal encryption method based on dual GAN Download PDFInfo
- Publication number
- CN109413068B CN109413068B CN201811268894.7A CN201811268894A CN109413068B CN 109413068 B CN109413068 B CN 109413068B CN 201811268894 A CN201811268894 A CN 201811268894A CN 109413068 B CN109413068 B CN 109413068B
- Authority
- CN
- China
- Prior art keywords
- wireless signal
- network
- size
- encryption
- signal
- 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.)
- Active
Links
Images
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
The invention discloses a wireless signal encryption method based on dual GAN, which comprises the following steps: encrypting an original wireless signal into an encrypted signal image by using a wireless signal encryption model; decrypting the encrypted signal image into a decrypted wireless signal by using a wireless signal decryption model; the wireless signal encryption model and the wireless signal decryption model are obtained by training through the following model training system, wherein the model training system comprises: the wireless signal encryption model comprises a wireless signal encryption network Ag, a wireless signal encryption judgment network Ad, a wireless signal decryption network Bg, a wireless signal decryption judgment network Bd and a wireless signal classification network C, and the model training system adopts pre-training of Ag and C and combined training of Ag, Ad, Bg, Bd and C to determine a wireless signal encryption model and a wireless signal decryption model. The method can effectively encrypt the wireless signals into pictures, and improves the concealment of information transmission.
Description
Technical Field
The invention belongs to the field of deep learning algorithm in the field of artificial intelligence and data security research, and particularly relates to a wireless signal encryption method based on dual GAN.
Background
At present, deep learning has been widely applied to the fields of object detection, image detection, data generation, and the like. Generation of countermeasure networks (GAN) is one of the more advanced techniques in deep learning today, which passes through (at least) two modules in the framework: the mutual game learning of the Generative Model (Generative Model) and the discriminant Model (discriminant Model) produces better output. According to different network structures, a series of GAN networks are derived: DCGAN, WGAN-GP, LSGAN-BEGAN, DUAL-GAN.
The DUAL-GAN generative countermeasure network (DUAL-GAN) is further extended to be two coupled generative countermeasure networks on the basis of the original generative countermeasure network, wherein two generators and two discriminators are provided, and the translation between characters and pictures is taken as an example, wherein the first generator Ag can translate the characters into the pictures, and the corresponding discriminator Ad can identify the true and false of the pictures. Meanwhile, a generator Bg corresponding to the generator Bg is constructed, the photo can be translated into the character, and the same generator Bg also has a discriminator Bd corresponding to the generator Bg to identify whether the character is accurate or not.
The existing wireless signal encryption methods are mainly divided into symmetric encryption and symmetric encryption technologies. Symmetric encryption uses the same key for both encryption and decryption of symmetric encoding technology files. Asymmetric encryption requires two keys: public key and private key, one of which is used to encrypt and the other to decrypt. The existing encryption method has the problems of being easy to attack, single encryption mode and the like. Generally, the wireless signal encryption technology adopts a digital encryption mode, which has a relatively high calculation cost and is easy to be intercepted and utilized.
Disclosure of Invention
The invention aims to provide a wireless signal encryption method based on dual GAN, which generates a picture corresponding to a wireless signal through a generation countermeasure network of the dual GAN and classifies the picture by using a wireless signal classification model C, thereby realizing the encryption process of wireless signal data. The wireless signal encryption method improves the confidentiality and the safety of the wireless signal transmission process.
In order to achieve the purpose, the invention provides the following technical scheme:
a wireless signal encryption method based on dual GAN comprises the following steps:
encrypting an original wireless signal into an encrypted signal image by using a wireless signal encryption model;
decrypting the encrypted signal image into a decrypted wireless signal by using a wireless signal decryption model;
the wireless signal encryption model and the wireless signal decryption model are obtained by training through the following model training system, wherein the model training system comprises:
the wireless signal encryption network Ag is used for encrypting the original wireless signal, inputting the original wireless signal and the decrypted wireless signal output by the wireless signal decryption network Bg, and outputting the decrypted wireless signal as an encrypted signal image;
the wireless signal encryption distinguishing network Ad inputs an encrypted signal image output by a normal color image and a wireless signal encryption network Ag and outputs a judgment result of the normal color image and the encrypted signal image, wherein when the output feedback of the wireless signal encryption distinguishing network Ad is used for training the wireless signal encryption distinguishing network Ad parameter, the class mark of the normal color image is defined as 1, the class mark of the encrypted signal image is defined as 0, when the output feedback of the wireless signal encryption distinguishing network Ad is used for training the wireless signal encryption distinguishing network Ag parameter, only the encrypted signal image is used, and the class mark of the encrypted signal image is defined as 1;
the wireless signal decryption network Bg is used for decrypting the encrypted signal image, inputting the encrypted signal image which is a normal color image and is output by the wireless signal encryption network Ag, and outputting the encrypted signal image as a decrypted wireless signal;
the wireless signal decryption judgment network Bd inputs the original wireless signal and the decrypted wireless signal output by the wireless signal decryption network Bg, and outputs the result of judging the original wireless signal and the decrypted wireless signal, wherein when the output of the wireless signal decryption judgment network Bd is used for feeding back the parameters of the wireless signal decryption judgment network Bd, the class mark of the original wireless signal is defined as 1, the class mark of the decrypted wireless signal is defined as 0, when the output of the wireless signal decryption judgment network Bd is used for feeding back the parameters of the wireless signal decryption network Bg, only the decrypted wireless signal is used, and the class mark of the decrypted wireless signal is defined as 1;
the wireless signal classification network C inputs the encrypted signal image output by the wireless signal encryption network Ag and the class mark of the wireless signal for generating the encrypted signal image, namely the modulation type of the wireless signal, and outputs the classification result of the encrypted signal, wherein each classification result corresponds to the modulation type of the original wireless signal;
the training process is as follows:
stage one: pre-training, namely pre-training a wireless signal encryption network Ag and a wireless signal classification network C by adopting an original wireless signal and a corresponding wireless signal modulation type mark;
and a second stage: retraining, namely firstly fixing parameters of a wireless signal encryption and discrimination network Ad, a wireless signal decryption and discrimination network Bd and a wireless signal classification network C, and training parameters of a wireless signal encryption network Ag and a wireless signal decryption network Bg in a combined manner; then, fixing parameters of a wireless signal encryption network Ag, a wireless signal decryption network Bg and a wireless signal classification network C, and training parameters of a wireless signal encryption judgment network Ad and a wireless signal decryption judgment network Bd in a combined manner; finally, parameters of a fixed wireless signal encryption judging network Ad, a wireless signal decryption network Bg and a wireless signal decryption judging network Bd are combined to train parameters of a wireless signal encryption network Ag and a wireless signal classification network C; through the three combined training, the training is completed until the wireless signal encryption network Ag and the wireless signal encryption judgment network Ad, the wireless signal decryption network Bg and the wireless signal decryption judgment network Bd realize Nash balance, the trained wireless signal encryption network Ag is a wireless signal encryption model, and the trained wireless signal decryption network Bg is a wireless signal decryption model.
Through the hierarchical joint training strategy provided by the invention, the trained wireless signal encryption network Ag can be used as a private key of a wireless signal, and wireless signal data is encrypted into a signal image; the wireless signal decryption network Bg can be used as another key of the encrypted signal image to decrypt the encrypted signal image into wireless signal data, and the wireless signal classification network C is introduced to classify the wireless signals obtained by using different modulation types, so that the wireless signal encryption network Ag can learn the classification characteristic attributes of the modulation types of the wireless signals simultaneously in the training process, and the wireless signals of different modulation types can be better encrypted.
The dual GAN-based wireless signal encryption method has the following effects: the wireless signal can be projected from the signal characteristic space to the image characteristic space through the wireless signal encryption model and automatically encrypted into the image signal, so that the identification of a wireless signal identification system is avoided. The encryption method can improve the concealment of wireless signal propagation and the safety of information transmission.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to these drawings without creative efforts.
FIG. 1 is a schematic structural diagram of a model training architecture provided by an embodiment;
fig. 2 is a schematic structural diagram of a wireless signal encryption network Ag according to an embodiment;
fig. 3 is a schematic structural diagram of a wireless signal decryption network Bg according to an embodiment;
fig. 4 is a schematic structural diagram of an Ad for wireless signal encryption discrimination provided by the embodiment;
fig. 5 is a schematic structural diagram of a wireless signal decryption decision network Bd according to an embodiment;
fig. 6 is a schematic structural diagram of a wireless signal classification network C according to an embodiment;
fig. 7(a) -7 (l) are schematic diagrams of visualization results obtained by encrypting 12 types of wireless signals into pictures in a training process of the classification method provided by the embodiment.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be further described in detail with reference to the accompanying drawings and examples. It should be understood that the detailed description and specific examples, while indicating the scope of the invention, are intended for purposes of illustration only and are not intended to limit the scope of the invention.
In order to improve the security of wireless signal transmission, the embodiment provides a wireless signal encryption method based on dual GAN, which specifically includes the steps of: encrypting an original wireless signal into an encrypted signal image by using a wireless signal encryption model; and decrypting the encrypted signal image into a decrypted wireless signal by using the wireless signal decryption model.
The wireless signal encryption model and the wireless signal decryption model are obtained by utilizing a model training system shown in fig. 1, and the specific model training system comprises five modules: the wireless signal encryption network Ag, the wireless signal encryption judgment network Ad, the wireless signal decryption network Bg, the wireless signal decryption judgment network Bd and the wireless signal classification network C are respectively.
The wireless signal encryption network Ag is mainly used for encrypting wireless signals, that is, encrypting original wireless signals into encrypted signal images. The wireless signal encryption network Ag is a neural network composed of LSTM units, convolution modules, and full-link modules, and its structure is shown in fig. 2, and includes: the original input wireless signal has the size of [512,2], wherein 512 represents the sampling time point of the wireless signal, 2 represents the characteristic value of each time point of the wireless signal, the training process adopts a minimum batch gradient descent method for training, the number of wireless signal data samples of each batch in the minimum batch is generally 64, a characteristic layer with the size of [512,128] is obtained after passing through an LSTM unit, wherein 512 corresponds to the original time point, 128 corresponds to the characteristic vector calculated by each time point, a characteristic layer with the size of 128 is obtained by using full connection, a characteristic layer with the size of 64 × 3 is obtained by using a full connection module, the characteristic layer is deformed by using a reshape function to obtain the characteristic layer with the size of [64,64,3], wherein 64 corresponds to the length and the width of the characteristic layer respectively, 3 corresponds to the depth of the characteristic layer, a characteristic layer with the size of [5,5,64 convolution module and max pooling module of size [2,2] step size 2 get feature layers of size [32,32,64], use convolution module of size [5, 128] and max pooling module of size [2,2] step size 2 get feature layers of size [16, 128], use convolution module of size [5, 256] and max pooling module of size [2,2] step size 2 get feature layers of size [8, 256], use convolution module of size [5, 512] and max pooling module of size [2,2] step size 2 get feature layers of size [4, 512], use convolution module of size [5, 512] and max pooling module of size [2,2] step size 2 get feature layers of size [2, 512], use convolution module of size [5, 2] and max pooling module of size [ 2] step size 512,2 maximum pooling module with step size 2 to get feature layers of size [1, 512], using deconvolution module of size [5,5,512] to get feature layers of size [2, 512], combining previous feature layers of [2, 512] to get feature layers of size [2, 1024], using deconvolution module of [5,5,512] to get feature layers of [4, 512], combining previous feature layers of [4, 512] to get feature layers of [4, 1024], using deconvolution module of [5,5,256] to get feature layers of [8, 256], combining previous feature layers of [8, 256] to get feature layers of [8, 512], using deconvolution module of [5,5,128] to get feature layers of [16, 128], combining previous feature layers of [16, 128] to get feature layers of [16, 256] to get deconvolution module of [5, 64, 32 ],32, and (2) combining the feature layers of the previous [32,32,64] to obtain the feature layers of [32,32,128], and obtaining the feature layers of [64,64,3] by using a deconvolution module of [5,5,3], so that the obtained picture is the corresponding ImageNet64, the picture has the length of 64 and the width of 64, and comprises RGB three channels.
The wireless signal decryption network Bg is mainly used for decrypting the encrypted signal image, i.e. decrypting the encrypted signal image into a decrypted wireless signal. The wireless signal decryption network Bg is a neural network formed by a convolution module and a full-connection module, and the structure of the wireless signal decryption network Bg is shown in fig. 3, and comprises: the original input image size is [64,64,3], a feature layer of size [32,32,64] is obtained using a convolution module of size [5,5,64] and a maximum pooling module of size [2,2] step size 2, a feature layer of size [16, 128] is obtained using a convolution module of size [5,5,128] and a maximum pooling module of size [2,2] step size 2, a feature layer of size [8, 256] is obtained using a convolution module of size [5,5,256] and a maximum pooling module of size [2,2] step size 2, a feature layer of size [4, 1024] is obtained using a convolution module of size [5,5,1024] and a maximum pooling module of size [2,2] step size 2, a feature layer of size [4, 1024] is obtained using a convolution module of size [5,5,2048] and a maximum pooling module of size [2,2] step size 2 to obtain a feature layer of size [ 2] step size 2,2,2048], obtaining a feature layer with the size of [1, 4096] by using a convolution module with the size of [5, 4096] and a maximum pooling module with the size of [2,2] and the step size of 2, obtaining a feature layer with the size of 16384(512 x 32) by using a full connection module, obtaining a feature layer with the size of 4096(512 x 8) by using a full connection module, obtaining a feature layer with the size of 1024(512 x 2) by using a full connection module, changing the shape of the feature layer into [512,2] by using a reshape function, and obtaining a wireless signal corresponding to a picture, wherein 512 corresponds to a sampling time point, and 2 corresponds to a feature value of each time point.
The wireless signal encryption judgment network Ad is mainly used for carrying out two classifications on a normal color image and an encrypted signal image, namely, judging the class marks of the normal color image and the encrypted signal image, defining the class mark of the normal color image as true (real), defining the class mark of the encrypted signal image as false (fake), and feeding back training (feedback training) wireless signal encryption network Ag through a classification result to enable the image generated by encryption to be closer to a true value. The wireless signal encryption discrimination network Ad is a neural network composed of a convolution module and a full-connection module, and the structure of the network Ad is shown in fig. 4, and includes: the input image size is [64,64,3], the feature layer of [32,32,64] is obtained using a convolution module of [5,5,64] and a maximum pooling module of [2,2] step size 2, the feature layer of [16, 128] is obtained using a convolution module of [5,5,128] and a maximum pooling module of [2,2] step size 2, the feature layer of [8, 256] is obtained using a convolution module of [5,5,256] and a maximum pooling module of [2,2] step size 2, the feature layer of [4, 512] is obtained using a convolution module of [5,5,512] and a maximum pooling module of [2,2] step size 2, the feature layer of [1, 1] is obtained using a convolution module of [5,5,1] and a maximum pooling module of [4,4] step size 4, the unnormalized confidence value of the same as the cross-normalized confidence value of 0, or the color signal output of the Ag-normalized color encryption matrix as the normal color output signal The difference in the images. The input of the wireless signal encryption discrimination network Ad is a mixture of a normal color image of ImageNet64 and an encrypted signal image generated by Ag encryption, and the output is a result of discrimination of the normal color image and the encrypted signal image.
The wireless signal decryption discrimination network Bd is mainly used for carrying out secondary classification on an original wireless signal and a decrypted wireless signal, namely, the class marks of the original wireless signal and the decrypted wireless signal are judged, the class mark of the original wireless signal is defined as true (real) and the class mark of the decrypted wireless signal is defined as false (fake), and the wireless signal decryption network Bg is trained (feedback training) through the classification result feedback, so that the wireless signal generated by decryption is closer to a true value. The wireless signal decryption discrimination network Bd is a neural network composed of an LSTM unit, a convolution module and a full connection module, and the structure of the network is shown in fig. 5, and includes: the original input wireless signal sample size is [512,2], a feature layer with the size of [512,64] is obtained after passing through an LSTM unit, wherein 512 corresponds to the original time point, 64 corresponds to the feature vector obtained by calculation of each time point, the feature layer with the size of 64 is obtained by using full connection, the feature layer with the size of 1 is obtained by using full connection, and the obtained non-normalized confidence value and 0 or 1 are used for cross entropy calculation of the difference between the original wireless signal generated by the wireless signal encryption network Ag and the decrypted wireless signal. The input of the wireless signal decryption discrimination network Bd is a mixture of normal wireless signals and decrypted wireless signals of Ag, and the output is a judgment result of the normal wireless signals and the decrypted wireless signals.
The wireless signal classification network C is mainly used for classifying wireless signals, inputting encrypted signal images output by the wireless signal encryption network Ag, outputting classification results of the encrypted signals, and making the wireless signal encryption network Ag more accurate by feedback fine tuning training (fine tuning) Ag. The wireless signal classification network C is a neural network composed of a convolution module and a full-connection module, and its specific structure is shown in fig. 6, and includes: the input picture size [64,64,3], using a convolution kernel of size [5,5,64] and a maximum pooling module of size [2,2] step size 2 to obtain a feature layer of size [32,32,64], using a convolution kernel of size [5,5,128] and a maximum pooling module of size [2,2] step size 2 to obtain a feature layer of size [16, 128], using a convolution kernel of size [5,5,256] and a maximum pooling module of size [2,2] step size 2 to obtain a feature layer of size [8, 256], using a convolution kernel of size [5,5,512] and a maximum pooling module of size [2,2] step size 2 to obtain a feature layer of size [4, 512], using full connectivity to obtain a feature layer of size 1024, using full connectivity module to obtain a feature layer of size 12 as output (for a task class of 12 wireless signals including 12 modulation types), and performing cross entropy calculation on the obtained non-normalized confidence value and the class mark of the original wireless signal corresponding to the encrypted signal image to obtain a distance.
The design target of the model training system is as follows:
(a) the original wireless signals are automatically encrypted through the network Ag, the wireless signals are respectively used as input samples of a wireless signal encryption judgment network Ad and a wireless signal classification network C, and are fed back to the Ag through the outputs of the Ad and the C, and internal parameters of the Ag are adjusted. The Ad adjusts Ag parameters by judging the distribution distance between the picture sample generated by Ag encryption and a real normal sample; and C, adjusting the Ag parameter by predicting the distance between the class mark and the real class mark. Finally, through the multi-party game process of Ag, Ad and C, the stability of the picture sample generated by Ag encryption is ensured, and the generated encrypted picture is used for training C to obtain a wireless signal encryption model which can better encrypt wireless signal data of different modulation types;
(b) the picture data is automatically decrypted into signal data through the network Bg, the signal data is used as an input sample of the wireless signal decryption judgment network Bd, the input sample is fed back to the Bg through the Bd, and internal parameters of the Bg are adjusted. Wherein Bd adjusts Bg parameters by determining the distribution distance between the signal samples generated by Bg decryption and real samples.
In the model training system, the network structure of the wireless signal classification network C is related to the complexity of a data set; the network structure design of Ag and Ad is related to the complexity of the C network and the data set; the Bg and Bd network structure is designed according to the complexity of the data set. In order to realize better encryption effect, prevent model collapse in the training process and better realize encryption and decryption of wireless models, the invention uses a joint training method to correspondingly train the models.
Specifically, the training of the model training system is mainly divided into two stages, namely pre-training and retraining, and the specific process is as follows:
a pre-training stage: and pre-training the wireless signal encryption network Ag and the wireless signal classification network C by adopting the original wireless signal and the corresponding wireless signal modulation type mark. During training, parameters of a fixed wireless signal encryption judgment network Ad, a wireless signal decryption judgment network Bd and a wireless signal decryption network Bg are set, and epochs of training is set to be N1, namely a training data set is used for N1 times; the input of the wireless signal encryption network Ag is an original wireless signal, the output of the wireless signal encryption network Ag is an encrypted signal image after encryption, the input of the wireless signal classification network C is an encrypted signal image output by the wireless signal encryption network Ag, and the output of the wireless signal classification network C is a class mark prediction of a debugging type of the original wireless signal corresponding to the encrypted signal image; at the moment, the wireless signal encryption network Ag and the wireless signal classification network C are used as an integral classification model Ag-C to classify the modulation types of the wireless signals, the wireless signal encryption network Ag is equivalent to a feature extraction module of the classification model Ag-C, and the wireless signal classification network C is equivalent to a classification module of the classification model Ag-C.
In particular, a raw signal data set x is inputsignal, generating a corresponding picture Ag (x) through an Ag networksignal) Mixing Ag (x)signal) Input into classification model C, 10 epochs were trained.
In the retraining stage, a joint training strategy is adopted, and the specific process is as follows:
(1) the parameters of a fixed wireless signal encryption judging network Ad, a wireless signal decryption judging network Bd and a wireless signal classification network C are used for encrypting the original wireless signal xsignalA decrypted wireless signal Bg (x) output by the wireless signal decrypting network Bgimage) As the input of a wireless signal encryption network Ag, training parameters of the Ag to enable an encrypted signal image output by the Ag to be closer to a normal color image, thereby realizing the encryption process of a wireless signal and enabling encrypted information to be more safely transmitted; and the normal color image ximageAnd an encrypted signal image Ag (x) output by the wireless signal encryption network Agsignal) As the input of the wireless signal decryption network Bg, training the parameters of the Bg to enable the output decrypted wireless signals to be closer to the original wireless signals, thereby realizing the decryption process of the encrypted wireless signals;
the optimization objective of the process is expressed as:
wherein x isAgRepresenting an image of an encrypted signal output by a wireless signal encryption network Ag, xBgRepresenting decrypted radio signals, x, output by a radio signal decrypting network BgAgpAg represents xAgSampling from the output, x, of a wireless signal encryption network AgBgpBg denotes xBgThe output sampled from the radio signal decryption network Bg, pAg and pBg represent the probability distribution of Ag and Bg outputs, Ad (x), respectivelyAg) Network Ad pair x for representing wireless signal encryption judgmentAgB d (x) is the discrimination probability ofBg) Network Bd for x for representing wireless signal decryption discriminationBgE (-) represents the expectation of cross entropy;
(2) parameters of fixed wireless signal encryption network Ag, wireless signal decryption network Bg and wireless signal classification network C, and encrypted signal image Ag (x) output by wireless signal encryption network Agsignal) And normal color image
The mixed data is used as the input of a wireless signal encryption judgment network Ad, and parameters of the Ad are trained so that the Ad can distinguish an encrypted signal image from a normal color image; and decrypting the decrypted wireless signal Bg (x) output by the wireless signal decrypting network Bgimage) And the original wireless signal
The mixed data is used as the input of a wireless signal decryption discrimination network Bd, and the parameters of the Bd are trained so that the decrypted wireless signals and the original wireless signals can be distinguished;
the optimization objective of the process is expressed as:
wherein Pdata represents an original data set,
representing normal signal samples
From the raw signal data Pdata;
representing normal image data
From the raw image data Pdata;
(3) the parameters of a fixed wireless signal encryption judgment network Ad, a wireless signal decryption network Bg and a wireless signal decryption judgment network Bd are used for encrypting the original wireless signal xsignalA decrypted wireless signal Bg (x) output by the wireless signal decrypting network Bgimage) An encrypted signal image Ag (x) as input of the wireless signal encryption network Ag and output of the wireless signal encryption network Agsignal) And the corresponding true classmark y (i.e., the original wireless signal x)signalCorresponding modulation type class mark) as the input of the wireless signal classification network C, and training the parameters of the wireless signal encryption network Ag and the wireless signal classification network C in a coordinated manner;
the optimization goals of the process are as follows:
where y denotes a modulation type class index of the original wireless signal,
indicating normal wireless signal samples
Sampling from wireless signal dataSet, C (-) represents the classification result of the wireless signal classification network C on the wireless signal modulation type;
a wireless signal classification model network C is introduced to classify wireless signals obtained by using different modulation types, so that a wireless signal encryption network Ag can learn classification characteristic attributes of the wireless signal modulation types in a training process, and the wireless signal encryption network Ag can better encrypt the wireless signal modulation types.
(4) Repeating the steps (1) to (3) until the wireless signal encryption network Ag and the wireless signal encryption judgment network Ad realize Nash balance, namely the game of the wireless signal encryption network Ag and the game of the wireless model encryption judgment network Ad tend to balance, the game of the wireless signal decryption network Bg and the game of the wireless signal decryption judgment network Bd tend to balance, the training is stopped, the trained wireless signal encryption network Ag is a wireless signal encryption model, and the trained wireless signal decryption network Bg is a wireless signal decryption model.
During training, the loss functions involved are respectively:
the loss function Ag _ loss of the wireless signal encryption network Ag is as follows:
wherein λ isBFor introducing parameters, to balance B _ loss with
So that B _ loss and
are close in value range, xAgpAg represents xAgB _ loss represents the mean square error of the normal color image and the encrypted signal image, and is in the form of:
B_loss＝MSE(image,Ag(xsignal))
wherein, the image represents the normality on ImageNet64Color image, the normal color image and the image Ag (x)signal) Two images with consistent patterns, wherein MSE (DEG) represents a mean square error function;
the loss function Bg _ loss of the wireless signal decryption network Bg is as follows:
wherein λ isAFor introducing parameters, balance A _ loss and
so that A _ loss and
a _ loss represents the mean square error of the original wireless signal and the decrypted wireless signal, and is of the form:
A_loss＝MSE(signal,Bg(ximage))
wherein, signal represents original wireless signal;
the loss function Ad _ loss of the wireless signal encryption judgment network Ad is as follows:
the loss function Bd _ loss of the wireless signal decryption discrimination network Bd is as follows:
the loss function C _ loss of the radio signal classification network C is:
in this embodiment, the calculation and optimization of the above 5 loss functions are converted into the calculation of the following 3 loss functions:
G_loss＝Ag_loss+Bg_loss
D_loss＝Ad_loss+Bd_loss
C_loss
through the calculation of the above 3 loss functions, parameters of each network are updated by back propagation using an Adam optimizer.
When nash equalization is implemented there are:
based on the minimum maximum theorem (Minimax theorem), the competition game relation of three parties of Ag and Bg, Ad, Bd and C is realized by the training method, the goal of Ag is to encrypt a large amount of wireless signal samples quickly, the distribution of the wireless signal samples can be as close to the real sample distribution as possible, and meanwhile, Ad can be fooled; the Ad aims to distinguish the picture sample generated by the Ag from the real picture sample as much as possible; c aims at classifying the picture generated by Ag as correctly as possible, so that the encryption process from the wireless signal to the picture is realized; bg aims to decrypt the encrypted signal image quickly and restore the encrypted signal image to the original signal as far as possible, Bd aims to distinguish the original normal signal sample from the decrypted signal sample as far as possible and further improve the restoration degree of Bg decryption, so that the decryption process of the signal is realized.
Through the training strategy, the wireless signal encrypted picture generated by the Ag and the wireless signal decrypted by the Bg are closer to the corresponding true value, so that the Ag can be used as a private key of the wireless signal to encrypt the wireless signal into the picture; bg can be used as another key of the wireless signal to translate the picture data into wireless signal data, and a wireless signal classification model C is introduced to classify the wireless signals obtained by using different modulation types, so that a wireless signal encryption model Ag can learn the classification characteristic attributes of the modulation types of the wireless signals simultaneously in the training process, and the wireless signals of different types can be better encrypted.
Specific experiments are as follows:
the basic cases of data sets include: (a) the wireless signal data has 312000 training samples and 156000 test samples, each sample having a size of 512 x 2 matrix and a sample span of (-6, 6). The validation set is 10% of the number of samples randomly drawn from the test sample; (b) the data set can be divided into twelve classes according to modulation types, each class is divided equally, 26000 samples exist in each class in a training set, and 13000 samples exist in each class in a testing set; (c) signal-to-noise ratio case: for 500 x 26 x 12 x 156000 samples in the test set, there are 26 snrs per class (-all even numbers from 20 to 30), and the distribution of different snr datasets in each class of samples is roughly: 500 samples with a signal-to-noise ratio of-20, 500 samples with a signal-to-noise ratio of-18, 500 samples with a signal-to-noise ratio of-16, … … 500 samples with a signal-to-noise ratio of 28, and 500 samples with a signal-to-noise ratio of 30. For the data distribution of the training set, the number of samples for one snr in a class is 1000, and the number of samples in the training set is 1000 × 26 × 12 — 312000. (d) All wireless signal sample data is subjected to simple reduction processing of dividing by 6 so as to be conveniently input into the model.
And training the constructed model training system by using the training set to obtain a trained wireless signal encryption model and a trained wireless signal decryption model. And inputting the samples in the test set into the wireless signal encryption model, and outputting the samples as encrypted signal images corresponding to 12 types of wireless signals shown in fig. 7(a) -7 (l), because the input batch (batch) is 64, each image comprises 64 images, and the images are difficult to artificially and respectively contain the contents, which shows that the method achieves the expected effect on the encryption of the wireless signals.
The above-mentioned embodiments are intended to illustrate the technical solutions and advantages of the present invention, and it should be understood that the above-mentioned embodiments are only the most preferred embodiments of the present invention, and are not intended to limit the present invention, and any modifications, additions, equivalents, etc. made within the scope of the principles of the present invention should be included in the scope of the present invention.
Claims (8)
1. A wireless signal encryption method based on dual GAN comprises the following steps:
encrypting an original wireless signal into an encrypted signal image by using a wireless signal encryption model;
decrypting the encrypted signal image into a decrypted wireless signal by using a wireless signal decryption model;
the wireless signal encryption model and the wireless signal decryption model are obtained by training through the following model training system, wherein the model training system comprises:
the wireless signal encryption network Ag is used for encrypting the original wireless signal, inputting the original wireless signal and the decrypted wireless signal output by the wireless signal decryption network Bg, and outputting the decrypted wireless signal as an encrypted signal image;
the wireless signal encryption distinguishing network Ad inputs an encrypted signal image output by a normal color image and a wireless signal encryption network Ag and outputs a judgment result of the normal color image and the encrypted signal image, wherein when the output feedback of the wireless signal encryption distinguishing network Ad is used for training the wireless signal encryption distinguishing network Ad parameter, the class mark of the normal color image is defined as 1, the class mark of the encrypted signal image is defined as 0, when the output feedback of the wireless signal encryption distinguishing network Ad is used for training the wireless signal encryption network Ag parameter, only the encrypted signal image is used, and the class mark of the encrypted signal image is defined as 1;
the wireless signal decryption network Bg is used for decrypting the encrypted signal image, inputting the encrypted signal image which is a normal color image and is output by the wireless signal encryption network Ag, and outputting the encrypted signal image as a decrypted wireless signal;
the wireless signal decryption judgment network Bd inputs the original wireless signal and the decrypted wireless signal output by the wireless signal decryption network Bg, and outputs the result of judging the original wireless signal and the decrypted wireless signal, wherein when the output of the wireless signal decryption judgment network Bd is used for feeding back the parameters of the wireless signal decryption judgment network Bd, the class mark of the original wireless signal is defined as 1, the class mark of the decrypted wireless signal is defined as 0, when the output of the wireless signal decryption judgment network Bd is used for feeding back the parameters of the wireless signal decryption network Bg, only the decrypted wireless signal is used, and the class mark of the decrypted wireless signal is defined as 1;
the wireless signal classification network C inputs the encrypted signal image output by the wireless signal encryption network Ag and the class mark of the wireless signal for generating the encrypted signal image, namely the modulation type of the wireless signal, and outputs the classification result of the encrypted signal, wherein each classification result corresponds to the modulation type of the original wireless signal;
the training process is as follows:
stage one: pre-training, namely pre-training a wireless signal encryption network Ag and a wireless signal classification network C by adopting an original wireless signal and a corresponding wireless signal modulation type mark;
and a second stage: retraining, namely firstly fixing parameters of a wireless signal encryption and discrimination network Ad, a wireless signal decryption and discrimination network Bd and a wireless signal classification network C, and training parameters of a wireless signal encryption network Ag and a wireless signal decryption network Bg in a combined manner; then, fixing parameters of a wireless signal encryption network Ag, a wireless signal decryption network Bg and a wireless signal classification network C, and training parameters of a wireless signal encryption judgment network Ad and a wireless signal decryption judgment network Bd in a combined manner; finally, parameters of a fixed wireless signal encryption judging network Ad, a wireless signal decryption network Bg and a wireless signal decryption judging network Bd are combined to train parameters of a wireless signal encryption network Ag and a wireless signal classification network C; through the three combined training, the training is completed until the wireless signal encryption network Ag and the wireless signal encryption judgment network Ad, the wireless signal decryption network Bg and the wireless signal decryption judgment network Bd realize Nash balance, the trained wireless signal encryption network Ag is a wireless signal encryption model, and the trained wireless signal decryption network Bg is a wireless signal decryption model.
2. The dual GAN-based wireless signal encryption method of claim 1 wherein the wireless signal encryption network is structured to include:
the original input wireless signal size is [512,2], wherein 512 represents the sampling time point of the wireless signal, 2 represents the characteristic value of each time point of the wireless signal, the training process adopts a minimum batch gradient descent method for training, the number of wireless signal data samples of each batch in the minimum batch is 64, a characteristic layer with the size of [512,128] is obtained after an LSTM unit, wherein 512 corresponds to the original time point, 128 corresponds to the characteristic vector obtained by calculation of each time point, a full-connection module is used for obtaining the characteristic layer with the size of 128, a full-connection module is used for obtaining the characteristic layer with the size of 64 × 3, a reshape function is used for deforming the characteristic layer to obtain the characteristic layer with the size of [64,64,3], wherein 64 corresponds to the length and the width of the characteristic layer respectively, and 3 corresponds to the depth of the characteristic layer, using a convolution module of size [5,5,64] and a maximum pooling module of size [2,2] step size 2 to obtain a feature layer of size [32,32,64], using a convolution module of size [5,5,128] and a maximum pooling module of size [2,2] step size 2 to obtain a feature layer of size [16, 128], using a convolution module of size [5,5,256] and a maximum pooling module of size [2,2] step size 2 to obtain a feature layer of size [8, 256], using a convolution module of size [5,5,512] and a maximum pooling module of size [2,2] step size 2 to obtain a feature layer of size [4, 512], using a convolution module of size [5,5,512] and a maximum pooling module of size [2,2] step size 2 to obtain a feature layer of size [2, 512], using a feature layer of size [2,2] and a maximum pooling module of size [2,2] step size 2 to obtain a feature layer of size [5, 2,512],512, 5,512 with a maximum pooling module of size [2,2] step size 2 to obtain feature layers of size [1, 512], using a deconvolution module of size [5, 512] to obtain feature layers of size [2, 512], combining the preceding feature layers of [2, 512] to obtain feature layers of size [2, 1024], using a deconvolution module of size [5, 512] to obtain feature layers of [4, 512], combining the preceding feature layers of [4, 512] to obtain feature layers of [4, 1024], using a deconvolution module of [5, 256] to obtain feature layers of [8, 256], combining the preceding feature layers of [8, 256] to obtain feature layers of [8, 512], using a deconvolution module of [5, 128] to obtain feature layers of [16, 128, 256], combining the preceding feature layers of [16, 256] to obtain feature layers of [8, 16,256],512 ],16, and (3) obtaining a feature layer of [32,32,64] by using a deconvolution module of [5,5,64], combining the feature layers of the previous [32,32,64] to obtain a feature layer of [32,32,128], and obtaining a feature layer of [64,64,3] by using a deconvolution module of [5,5,3], wherein the obtained picture is the corresponding ImageNet64, and the picture has the length of 64 and the width of 64 and comprises RGB three channels.
3. The dual GAN-based wireless signal encryption method of claim 1 wherein the structure of the wireless signal decryption network Bg comprises:
the original input image size is [64,64,3], a feature layer of size [32,32,64] is obtained using a convolution module of size [5,5,64] and a maximum pooling module of size [2,2] step size 2, a feature layer of size [16, 128] is obtained using a convolution module of size [5,5,128] and a maximum pooling module of size [2,2] step size 2, a feature layer of size [8, 256] is obtained using a convolution module of size [5,5,256] and a maximum pooling module of size [2,2] step size 2, a feature layer of size [4, 1024] is obtained using a convolution module of size [5,5,1024] and a maximum pooling module of size [2,2] step size 2, a feature layer of size [4, 1024] is obtained using a convolution module of size [5,5,2048] and a maximum pooling module of size [2,2] step size 2 to obtain a feature layer of size [ 2] step size 2,2,2048], obtaining a feature layer with the size of [1, 4096] by using a convolution module with the size of [5, 4096] and a maximum pooling module with the size of [2,2] and the step size of 2, obtaining a feature layer with the size of 16384(512 x 32) by using a full connection module, obtaining a feature layer with the size of 4096(512 x 8) by using a full connection module, obtaining a feature layer with the size of 1024(512 x 2) by using a full connection module, changing the shape of the feature layer into [512,2] by using a reshape function, and obtaining a wireless signal corresponding to a picture, wherein 512 corresponds to a sampling time point, and 2 corresponds to a feature value of each time point.
4. The dual GAN-based wireless signal encryption method according to claim 1, wherein the structure of the wireless signal encryption decision network Ad comprises:
the input image size is [64,64,3], the feature layer of [32,32,64] is obtained using a convolution module of [5,5,64] and a maximum pooling module of [2,2] step size 2, the feature layer of [16, 128] is obtained using a convolution module of [5,5,128] and a maximum pooling module of [2,2] step size 2, the feature layer of [8, 256] is obtained using a convolution module of [5,5,256] and a maximum pooling module of [2,2] step size 2, the feature layer of [4, 512] is obtained using a convolution module of [5,5,512] and a maximum pooling module of [2,2] step size 2, the feature layer of [1, 1] is obtained using a convolution module of [5,5,1] and a maximum pooling module of [4,4] step size 4, the unnormalized confidence value of the same as the cross-normalized confidence value of 0, or the color signal output of the Ag-normalized color encryption matrix as the normal color output signal The difference of the images;
the structure of the wireless signal decryption discrimination network Bd comprises:
the original input wireless signal sample size is [512,2], a feature layer with the size of [512,64] is obtained after passing through an LSTM unit, wherein 512 corresponds to the original time point, 64 corresponds to the feature vector obtained by calculation of each time point, the feature layer with the size of 64 is obtained by using full connection, the feature layer with the size of 1 is obtained by using full connection, and the obtained non-normalized confidence value and 0 or 1 are used for cross entropy calculation of the difference between the original wireless signal generated by the wireless signal encryption network Ag and the decrypted wireless signal.
5. The dual GAN-based wireless signal encryption method of claim 1 wherein the structure of the wireless signal classification network C comprises:
the input picture size is [64,64,3], a feature layer with the size [32,32,64] is obtained by using a convolution kernel with the size [5,5,64] and a maximum pooling module with the size [2,2] step size 2, a feature layer with the size [16, 128] is obtained by using a convolution kernel with the size [5,5,128] and a maximum pooling module with the size [2,2] step size 2, a feature layer with the size [8, 256] is obtained by using a convolution kernel with the size [5,5,256] and a maximum pooling module with the size [2,2] step size 2, a feature layer with the size [4, 512] is obtained by using full connection, a feature layer with the size 1024 is obtained by using full connection module as output, and the obtained feature layer with the size 12 is used as output, and the obtained cross-normalized signal confidence value corresponding to the non-encrypted signal class of the original image is used as the cross-encrypted signal class The distance is calculated by entropy.
6. The dual GAN-based wireless signal encryption method of claim 1, wherein the pre-training is performed by:
parameters of a fixed wireless signal encryption judging network Ad, a wireless signal decryption judging network Bd and a wireless signal decryption network Bg are set, and epochs of training is set to be N1, namely a training data set is used for N1 times; the input of the wireless signal encryption network Ag is an original wireless signal, the output of the wireless signal encryption network Ag is an encrypted signal image after encryption, the input of the wireless signal classification network C is an encrypted signal image output by the wireless signal encryption network Ag, and the output of the wireless signal classification network C is a class mark prediction of a debugging type of the original wireless signal corresponding to the encrypted signal image; at the moment, the wireless signal encryption network Ag and the wireless signal classification network C are used as an integral classification model Ag-C to classify the modulation types of the wireless signals, the wireless signal encryption network Ag is equivalent to a feature extraction module of the classification model Ag-C, and the wireless signal classification network C is equivalent to a classification module of the classification model Ag-C.
7. The dual-GAN-based wireless signal encryption method of claim 6, wherein the retraining process is as follows:
(1) parameters of a fixed wireless signal encryption judging network Ad, a wireless signal decryption judging network Bd and a wireless signal classification network C,original wireless signal xsignalA decrypted wireless signal Bg (x) output by the wireless signal decrypting network Bgimage) As the input of a wireless signal encryption network Ag, training parameters of the Ag to enable an encrypted signal image output by the Ag to be closer to a normal color image; and the normal color image ximageAnd an encrypted signal image Ag (x) output by the wireless signal encryption network Agsignal) As the input of the wireless signal decryption network Bg, training the parameters of the Bg to enable the output decrypted wireless signals to be closer to the original wireless signals, thereby realizing the decryption process of the encrypted wireless signals;
(2) parameters of fixed wireless signal encryption network Ag, wireless signal decryption network Bg and wireless signal classification network C, and encrypted signal image Ag (x) output by wireless signal encryption network Agsignal) And normal color image
The mixed data is used as the input of a wireless signal encryption judgment network Ad, and parameters of the Ad are trained so that the Ad can distinguish an encrypted signal image from a normal color image; and decrypting the decrypted wireless signal Bg (x) output by the wireless signal decrypting network Bgimage) And the original wireless signal
The mixed data is used as the input of a wireless signal decryption discrimination network Bd, and the parameters of the Bd are trained so that the decrypted wireless signals and the original wireless signals can be distinguished;
(3) the parameters of a fixed wireless signal encryption judgment network Ad, a wireless signal decryption network Bg and a wireless signal decryption judgment network Bd are used for encrypting the original wireless signal xsignalA decrypted wireless signal Bg (x) output by the wireless signal decrypting network Bgimage) An encrypted signal image Ag (x) as input of the wireless signal encryption network Ag and output of the wireless signal encryption network Agsignal) The corresponding real class mark y is used as the input of the wireless signal classification network C, and the parameters of the wireless signal encryption network Ag and the wireless signal classification network C are trained in a coordinated manner;
(4) repeating the steps (1) to (3) until the wireless signal encryption network Ag and the wireless signal encryption judgment network Ad realize Nash balance, namely the game of the wireless signal encryption network Ag and the game of the wireless model encryption judgment network Ad tend to balance, the game of the wireless signal decryption network Bg and the game of the wireless signal decryption judgment network Bd tend to balance, the training is stopped, the trained wireless signal encryption network Ag is a wireless signal encryption model, and the trained wireless signal decryption network Bg is a wireless signal decryption model.
8. A method for dual GAN-based encryption of wireless signals as claimed in claim 7, wherein when nash equalization is implemented there are:
wherein x isAgRepresenting an image of an encrypted signal output by a wireless signal encryption network Ag, xBgRepresenting decrypted radio signals, x, output by a radio signal decrypting network BgAgpAg represents xAgSampling from the output, x, of a wireless signal encryption network AgBgpBg denotes xBgThe output sampled from the radio signal decryption network Bg, pAg and pBg represent the probability distribution of Ag and Bg outputs, Ad (x), respectivelyAg) Network Ad pair x for representing wireless signal encryption judgmentAgB d (x) is the discrimination probability ofBg) Network Bd for x for representing wireless signal decryption discriminationBgE (-) represents the expectation of cross entropy; pdata represents a set of raw data that is,
representing normal signal samples
From the raw signal data Pdata;
representing normal image data
From the raw image data Pdata; y denotes a modulation type class mark of an original wireless signal,
indicating normal wireless signal samples
Sampled from the radio signal data set, C (-) represents the result of the classification of the radio signal modulation type by the radio signal classification network C.
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 CN109413068A (en) | 2019-03-01 |
| CN109413068B true 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) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111047006B (en) * | 2019-10-28 | 2023-04-21 | 浙江工业大学 | Dual generation network-based anti-attack defense model and application |
| CN111654368B (en) * | 2020-06-03 | 2021-10-08 | 电子科技大学 | Key generation method for generating countermeasure network based on deep learning |
| CN114710371B (en) * | 2022-06-08 | 2022-09-06 | 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)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106951919A (en) * | 2017-03-02 | 2017-07-14 | 浙江工业大学 | A kind of flow monitoring implementation method based on confrontation generation 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 | The facial image deblurring and restoration methods of confrontation type network are generated based on antithesis |
-
2018
- 2018-10-29 CN CN201811268894.7A patent/CN109413068B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106951919A (en) * | 2017-03-02 | 2017-07-14 | 浙江工业大学 | A kind of flow monitoring implementation method based on confrontation generation 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 | The facial image deblurring and restoration methods of confrontation type network are generated based on antithesis |
Non-Patent Citations (2)
| Title |
|---|
| A triple-key Chaotic neural network for;Suryawanshi S B;《 International journal of Engineering sciences& Emerging technologies》;20120430;全文 * |
| 生成式对抗网络研究进展;王万良;《通信学报》;20180228;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109413068A (en) | 2019-03-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109587372B (en) | Invisible image steganography based on generation of countermeasure network | |
| Mo et al. | Fake faces identification via convolutional neural network | |
| CN110084734B (en) | Big data ownership protection method based on object local generation countermeasure network | |
| Xie et al. | A generalizable model-and-data driven approach for open-set RFF authentication | |
| Lerch-Hostalot et al. | Unsupervised steganalysis based on artificial training sets | |
| CN110349136A (en) | A kind of tampered image detection method based on deep learning | |
| CN109413068B (en) | Wireless signal encryption method based on dual GAN | |
| Kong et al. | Detect and locate: Exposing face manipulation by semantic-and noise-level telltales | |
| CN100507943C (en) | Statistic supervision and structure supervision based hidden messages analysis system | |
| CN111860832A (en) | Method for enhancing neural network defense capacity based on federal learning | |
| CN111951149B (en) | Image information steganography method based on neural network | |
| CN109887047B (en) | Signal-image translation method based on generation type countermeasure network | |
| CN111858955B (en) | Knowledge graph representation learning enhancement method and device based on encryption federal learning | |
| CN104636764B (en) | A kind of image latent writing analysis method and its device | |
| CN109388959B (en) | Generation type information hiding method combining difference clustering and minimum error texture synthesis | |
| Lu et al. | Secure halftone image steganography based on feature space and layer embedding | |
| Xiao et al. | Securing the socio-cyber world: Multiorder attribute node association classification for manipulated media | |
| Bi et al. | High-capacity image steganography algorithm based on image style transfer | |
| Jiang et al. | A new steganography without embedding based on adversarial training | |
| Kumar et al. | Secured Cloud Application for Detection of Brain Tumor using Deep Learning Algorithms | |
| CN109447240B (en) | Training method of graphic image replication model, storage medium and computing device | |
| CN113378723B (en) | Automatic safety recognition system for hidden danger of power transmission and transformation line based on depth residual error network | |
| Liu et al. | Adaptive Texture and Spectrum Clue Mining for Generalizable Face Forgery Detection | |
| CN113486932A (en) | Convolutional neural network steganalysis-oriented optimization method | |
| Le et al. | Generating high-fidelity cybersecurity data with generative adversarial networks |
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 |