Semantic camouflage communication method and system, power grid and electronic equipment

Abstract

The invention discloses a semantic camouflage communication method and system, a power grid and electronic equipment, and relates to the technical field of power system safety. According to the method, semantic information extracted from a to-be-sent image by a semantic encoder of the semantic communication system is obtained at a sending end of the semantic communication system, and then a semantic information decoding module used for ensuring that both a legal user and a potential eavesdropper in a deployment environment of the semantic communication system can decode is added to the semantic information, and the legal user decodes to obtain the to-be-sent image, and the potential eavesdropper decodes to obtain the artificial noise of the disguised image, so as to obtain the transmission semantic information, pre-codes the transmission semantic information, and sends the pre-coded transmission semantic information. Therefore, the problems that when semantic communication is adopted to carry out power data transmission, confidentiality is maintained through encryption or a physical layer security system, an eavesdropper is easily aware of protection measures, and data transmission security is not high due to the adoption of more complex attacks are solved.

Description

Technical Field

[0001] The present invention relates to the field of power system safety technology, and more specifically, to a semantic camouflage communication method, system, power grid and electronic equipment. Background Art

[0002] In the process of power inspection of the power grid, in order to ensure the stable operation and efficient management of the power system, a large amount of image data from sensors needs to be collected and transmitted in real time. These data include but are not limited to equipment operation status monitoring, environmental monitoring information, etc. However, with the development of smart grid technology, the power inspection system has become more and more complex, and the amount of image data involved has become increasingly large, which has brought great transmission pressure to the network infrastructure. In order to solve this challenge, an advanced communication method, semantic communication, is currently commonly used to reduce the demand for network bandwidth. Because semantic communication no longer focuses on the transmission of data, but focuses more on understanding the meaning of data and ensuring that the receiving end can accurately obtain the intention of the sender. Therefore, in this way, the amount of data actually transmitted can be significantly reduced, thereby reducing the requirements for network bandwidth.

[0003] Understandably, semantic communication requires all communication participants to have the same knowledge base (KBs) and semantic codec network. However, in widely deployed semantic communication systems, the inherent security of semantic communication can be compromised. This occurs due to the widespread deployment of semantic encoders and decoders on terminal devices, providing eavesdroppers with potential opportunities to access legitimate receivers' decoders and KBs. Therefore, eavesdropping attacks in semantic communication become a critical security issue that needs to be carefully addressed before large-scale deployment of semantic communication systems. Many studies have proposed semantic communication methods that consider security aspects, relying on encryption or physical layer security systems to maintain confidentiality, however, these methods lack stealth, allowing eavesdroppers to be aware of information protection measures. In addition, eavesdroppers can adopt more sophisticated attacks, such as accessing KBs and semantic codec networks, to extract semantic information. Summary of the invention

[0004] The purpose of the present invention is to provide a semantic camouflage communication method, system, power grid and electronic equipment to solve the problem that when semantic communication is used for power data transmission, confidentiality is maintained through encryption or physical layer security system, which makes it easy for eavesdroppers to be aware of protection measures and use more complex attacks, resulting in low data transmission security.

[0005] In a first aspect, the present invention provides a semantic camouflage communication method, which is applied to a sending end of a semantic communication system, and the method comprises:

[0006] Acquire semantic information, where the semantic information is information extracted from the image to be sent by a semantic encoder of the semantic communication system;

[0007] Adding artificial noise to the semantic information to obtain transmission semantic information; the artificial noise is used to ensure that both legitimate users and potential eavesdroppers in the deployment environment of the semantic communication system can decode the transmission semantic information, the legitimate users decode the transmission semantic information to obtain the image to be sent, and the potential eavesdroppers decode the transmission semantic information to obtain the camouflaged image;

[0008] The transmission semantic information is precoded and sent.

[0009] In a preferred solution, the obtaining of semantic information includes:

[0010] Acquire the image to be sent;

[0011] Inputting the image to be sent into a semantic encoder of the semantic communication system to extract semantic information to obtain the semantic information;

[0012] The semantic encoder is jointly trained together with the semantic decoder of the semantic communication system based on channel state information of legitimate users and potential eavesdroppers in the deployment environment of the semantic communication system.

[0013] In a preferred solution, adding artificial noise to the semantic information includes:

[0014] Acquire the camouflage image;

[0015] Inputting the disguised image into an artificial noise generation model to obtain the artificial noise output by the artificial noise generation model;

[0016] applying the artificial noise to the semantic information;

[0017] The artificial noise generation model is a neural network model obtained by collaborative training with the semantic encoder and the semantic decoder.

[0018] In a preferred embodiment, the artificial noise generation model includes two layers of interconnected networks;

[0019] The first layer of the network is an encoder that imitates a semantic communication framework, which is used to ensure the retrieval of equivalent semantic information of the disguised image;

[0020] The second layer network is a U-Net architecture, which is used to receive the disguised image and generate the artificial noise with the same dimension as the disguised image.

[0021] In a preferred embodiment, the artificial noise generation model enables the potential eavesdropper to decode the actual information R e The label M close to the disguised image t ′, the actual information R obtained by the legitimate user decoding b The tag M close to the image to be sent b ′, and R e Not decoded as M b ′ constructs the loss function L for the target:

[0022]

[0023] in, represents the cross entropy loss.

[0024] In a preferred embodiment, the artificial noise complies with L ∞ Norm constraint.

[0025] In a preferred solution, based on a preset formula, the artificial noise is used to ensure that both the legitimate user and the potential eavesdropper can decode the transmission semantic information, and the legitimate user decodes to obtain the image to be sent, while the potential eavesdropper decodes to obtain a disguised image;

[0026] The preset formula is expressed as:

[0027] arg Distance[f(x p ,G),M t ]

[0028]

[0029] d(x p ,x)≤∈,

[0030] Where f(·) represents the semantic decoder and d(·,·) represents the function used to calculate L ∞ norm function, ∈ represents a pre-specified threshold, d(x p ,x)≤∈ indicates L ∞ Norm constraint.

[0031] In a preferred solution, the precoding and sending of the transmission semantic information includes:

[0032] Generate a precoding vector based on the channel state information of the legitimate user;

[0033] Precoding the transmission semantic information based on the precoding vector;

[0034] The precoded transmission semantic information is sent through a wireless channel.

[0035] In a second aspect, the present invention also provides another semantic disguised communication method, which is applied to a target receiving end of a semantic communication system, wherein the target receiving end is a receiving end of a legitimate user in a deployment environment of the semantic communication system, and the method comprises:

[0036] Acquire transmission semantic information, where the transmission semantic information is semantic information with artificial noise added, and the semantic information is information extracted from the image to be sent by a semantic encoder of the semantic communication system;

[0037] Inputting the transmission semantic information into a semantic decoder of the semantic communication system for decoding to obtain the image to be sent;

[0038] The semantic encoder and the semantic decoder are jointly trained based on channel state information of legitimate users and potential eavesdroppers in the deployment environment.

[0039] In a third aspect, the present invention also provides another semantic disguised communication method, which is applied to a disguised receiving end of a semantic communication system, wherein the disguised receiving end is a receiving end of a potential eavesdropper in a deployment environment of the semantic communication system, and the method comprises:

[0040] Acquire transmission semantic information, where the transmission semantic information is semantic information with artificial noise added, and the semantic information is information extracted from the image to be sent by a semantic encoder of the semantic communication system;

[0041] Inputting the transmission semantic information into a semantic decoder of the semantic communication system for decoding to obtain a disguised image output by the semantic decoder;

[0042] The semantic encoder and the semantic decoder are jointly trained based on channel state information of legitimate users and potential eavesdroppers in the deployment environment of the semantic communication system.

[0043] In a preferred solution, the artificial noise is generated by an artificial noise generation model, and the artificial noise generation model is a neural network model obtained by collaborative training with the semantic encoder and the semantic decoder.

[0044] In a fourth aspect, the present invention provides a semantic camouflage communication system, comprising:

[0045] A first acquisition unit, configured to acquire semantic information, wherein the semantic information is information extracted from an image to be sent by a semantic encoder of the semantic communication system;

[0046] A first processing unit is configured to add artificial noise to the semantic information to obtain transmission semantic information; the artificial noise is used to ensure that both legitimate users and potential eavesdroppers in the deployment environment of the semantic communication system can decode the transmission semantic information, the legitimate users decode the transmission semantic information to obtain the image to be sent, and the potential eavesdroppers decode the transmission semantic information to obtain the camouflaged image;

[0047] A sending unit is used to precode and send the transmission semantic information.

[0048] In a fifth aspect, the present invention also provides another semantic camouflage communication system, comprising:

[0049] A second acquisition unit is used to obtain transmission semantic information, where the transmission semantic information is semantic information with artificial noise added, and the semantic information is information extracted from the image to be sent by a semantic encoder of the semantic communication system;

[0050] A second processing unit, configured to input the transmission semantic information into a semantic decoder of the semantic communication system for decoding to obtain the image to be sent;

[0051] The semantic encoder and the semantic decoder are jointly trained based on channel state information of legitimate users and potential eavesdroppers in the deployment environment of the semantic communication system.

[0052] In a sixth aspect, the present invention also provides another semantic disguise communication system, comprising:

[0053] A third acquisition unit is used to obtain transmission semantic information, where the transmission semantic information is semantic information with artificial noise added, and the semantic information is information extracted from the image to be sent by a semantic encoder of the semantic communication system;

[0054] A third processing unit, configured to input the transmission semantic information into a semantic decoder of the semantic communication system for decoding to obtain a disguised image;

[0055] The semantic encoder and the semantic decoder are jointly trained based on channel state information of legitimate users and potential eavesdroppers in the deployment environment of the semantic communication system.

[0056] In a seventh aspect, the present invention also provides a power grid, including a power system body and the semantic camouflage communication system provided in the fourth aspect, the fifth aspect, or the sixth aspect of the present invention.

[0057] In a seventh aspect, the present invention provides an electronic device, the electronic device comprising:

[0058] processor;

[0059] a memory for storing instructions executable by the processor;

[0060] The processor is used to execute the semantic camouflage communication method provided by the first aspect, the second aspect or the third aspect of the present invention.

[0061] In order to achieve the above-mentioned purpose, the semantic camouflage communication method, system, power grid and electronic device provided by the present invention obtain the semantic information extracted from the image to be sent by the semantic encoder of the semantic communication system at the sending end of the semantic communication system, and then add artificial noise to the semantic information to ensure that both legitimate users and potential eavesdroppers in the deployment environment of the semantic communication system can decode it, and the legitimate user decodes to obtain the image to be sent, while the potential eavesdropper decodes to obtain the artificial noise of the camouflage image, obtains the transmission semantic information, and pre-encodes the transmission semantic information and sends it. The purpose of using artificial noise to mislead the decoding results of potential eavesdroppers and retaining accurate decoding for legitimate users is achieved, thereby avoiding the drawbacks of eavesdroppers being aware of the protection measures and adopting more complex attacks, and effectively improving the security of power data transmission through semantic communication technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 A flowchart of a semantic camouflage communication method provided by an embodiment of the present invention.

[0063] Figure 2 The convolutional layer provided for the embodiment of the present invention is a schematic diagram that equivalently represents precoding and wireless channels.

[0064] Figure 3 A structural diagram of a semantic communication network architecture example provided for an embodiment of the present invention.

[0065] Figure 4 A structural diagram of an artificial noise generation model example provided in an embodiment of the present invention.

[0066] Figure 5 A comparison diagram of a common semantic communication method provided in an embodiment of the present invention and a semantic camouflage communication method provided in an embodiment of the present invention.

[0067] Figure 6 A flowchart of another semantic camouflage communication method provided by an embodiment of the present invention.

[0068] Figure 7 A flowchart of another semantic camouflage communication method provided by an embodiment of the present invention.

[0069] Figure 8 A structural diagram of a semantic camouflage communication system provided in an embodiment of the present invention.

[0070] Fig. 9A structural diagram of another semantic camouflage communication system provided by an embodiment of the present invention.

[0071] Fig.10 A structural diagram of another semantic camouflage communication system provided by an embodiment of the present invention.

[0072] Fig.11 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0073] The present invention is further described in detail below through the accompanying drawings and specific embodiments.

[0074] It should be noted that an end-to-end semantic communication system needs to include a semantic encoder and a semantic decoder, so that the system can semantically encode information at the sending end according to the content of the knowledge base, and then semantically decode it at the receiving end.

[0075] Specifically, the process of training an end-to-end semantic communication system generally includes three processes: communication modeling, precoding and wireless channel transmission, and semantic communication framework training.

[0076] Among them, for the communication modeling process, it is assumed that the semantic communication system includes N t The transmitting end of the antenna, N r The legal users with N antennas and e Assuming that the eavesdropping channel obeys Rayleigh fading, the channel from the transmitter to the eavesdropper can be defined as The channel between the sender and the legitimate user can be expressed as Subject to Rician fading. Let x represent the signal of the image message M to be sent through joint semantic coding and channel coding. The precoding matrix is ​​given by Represented by. Then the signals received by the legitimate user and the potential eavesdropper are:

[0077] y b =HFx+n b (1)

[0078] y e =GFx+n e (2)

[0079] Among them, n b and n e Yes Obedience and It is understandable that in this case, due to the widespread deployment of semantic communication systems, both the semantic encoder and the semantic decoder are easily accessible to potential eavesdroppers. Therefore, potential eavesdroppers have the ability to decode the correct information. That is, when the legitimate user and the potential eavesdropper receive the corresponding signal y b and e After that, the same semantic decoder can be used to perform channel decoding and semantic decoding on the signal to obtain respective decoding results M b ′ and M e ′, and there exists M b ′ and M e 'Same possibility.

[0080] As for the precoding and wireless channel transmission process, such as Figure 2 As shown, assuming that convolutional layers are used to simulate precoding and wireless channel effects, they are defined by the precoding layer and the wireless channel layer respectively.

[0081] Specifically, represents the mth component of the nth row of F, where m and n are in the range (1,2,...,N t ).element represents the kth component in the rth row of H, where r∈(1,2,...,N r ). The signal received by the rth antenna in the lth time period is recorded as Then through the wireless channel layer, we can get a size of 1×2×N r The feature map of L is then reshaped into 2N r ×1×L, that is, the signal received by the kth user at the receiver. The precoding layer contains 2N t convolution kernels, and the wireless channel layer consists of 2N r The convolution kernels reflect the inherent complexity of the respective physical processes in semantic-aware communication systems.

[0082] This application aims to solve the above problems. During the power inspection process of the power grid, the security of semantic communication is enhanced by utilizing the channel difference between legitimate users and potential eavesdroppers from the perspective of physical layer security. That is, by applying semantic-level artificial noise to the semantic information, and then utilizing the difference in the channels between legitimate users and potential eavesdroppers, the decoding results of potential eavesdroppers are misled and cannot be detected, while retaining accurate decoding for legitimate users, thereby achieving the purpose of "hiding the truth and revealing the false".

[0083] It should be noted that the semantic camouflage communication method provided by the present invention is executed on an electronic device, which may be an intelligent terminal device such as a laptop, a personal computer, and a tablet computer, or may be a server on the network side.

[0084] like Figure 1 As shown, the semantic disguise communication method provided by the embodiment of the present invention is applied to the sending end of the semantic communication system, and mainly includes the following steps:

[0085] 110. Obtain semantic information.

[0086] The semantic information is information extracted from the image to be sent by a semantic encoder of the semantic communication system.

[0087] Specifically, before performing semantic camouflage communication, it is first necessary to train the semantic encoder and semantic decoder of the semantic communication system, and this training process needs to be jointly trained with the knowledge base and channel state information.

[0088] Furthermore, after the semantic encoder and the semantic decoder are trained, the sending end of the semantic communication system uses the semantic encoder to extract semantic information from the image to be sent, thereby obtaining semantic information.

[0089] Furthermore, in an optional embodiment, obtaining semantic information includes:

[0090] Get the image to be sent;

[0091] Inputting the image to be sent into a semantic encoder of a semantic communication system to extract semantic information and obtain semantic information;

[0092] The semantic encoder is jointly trained with the semantic decoder of the semantic communication system based on the channel state information of legitimate users and potential eavesdroppers in the deployment environment of the semantic communication system.

[0093] Specifically, by accurately estimating the wireless communication channel in the deployment environment, the channel state information of legitimate users and potential eavesdroppers can be obtained. By jointly training the semantic encoder and decoder with the channel state information, the semantic encoder and decoder can not only learn the key features in the knowledge base, but also absorb the channel state information, thereby achieving more efficient information transmission.

[0094] That is, through joint training, it is possible to ensure that the semantic encoder automatically adjusts to the channel characteristics when encoding information. In this way, even if the signal is affected by channel noise during transmission, the semantic decoder can accurately restore the original information, thereby minimizing information loss and errors. This optimizes the performance of the semantic communication system and improves the security and reliability of information transmission.

[0095] 120. Artificial noise is added to the semantic information to obtain the transmitted semantic information.

[0096] Among them, artificial noise is used to ensure that both legitimate users and potential eavesdroppers in the deployment environment of the semantic communication system can decode the transmitted semantic information. Legitimate users decode the transmitted semantic information to obtain the image to be sent, and potential eavesdroppers decode the transmitted semantic information to obtain the camouflaged image.

[0097] In an optional embodiment, adding artificial noise to the semantic information includes:

[0098] Obtaining a camouflaged image;

[0099] Inputting the disguised image into the artificial noise generation model to obtain artificial noise output by the artificial noise generation model;

[0100] Imposing artificial noise on semantic information;

[0101] Among them, the artificial noise generation model is a neural network model obtained by collaborative training with the semantic encoder and the semantic decoder.

[0102] In this embodiment, artificial noise is generated by using a disguised picture for misleading potential eavesdroppers, and the artificial noise is applied to the semantic information to obtain the transmitted semantic information, so that the legitimate users in the deployment environment of the semantic communication system can use the receiver to receive the semantic information and use the semantic decoder to decode the original information. However, when the potential eavesdropper uses the receiver to receive the semantic information and uses the semantic decoder to decode it, due to the existence of semantic-level artificial noise, the decoding result of the potential eavesdropper will be disguised information rather than the original information.

[0103] In some possible embodiments, the artificial noise generation model can be a neural network, denoted as Perturbation, which receives a picture M for misleading a potential eavesdropper. t As input, the output result is artificial noise δ, and the formula is summarized as:

[0104] δ=Perturbation(M t ) (3)

[0105] The artificial noise generation network is trained collaboratively with the semantic encoder and semantic decoder.

[0106] Semantic extraction uses the semantic encoder mentioned above to extract semantic information from the image to be sent, that is, the original information:

[0107] x=Encoder(M) (4)

[0108] Among them, Encoder is the semantic encoder, M is the original information, and x is the semantic information extracted by the semantic encoder from the original information.

[0109] The generated artificial noise is then applied to the semantic information to obtain the transmitted semantic information:

[0110] x p =Perturbation(M t )+Encoder(M)=δ+x (5)

[0111] 130. Precode and send the transmission semantic information.

[0112] Specifically, in an optional embodiment, precoding and sending the transmission semantic information includes:

[0113] Generate a precoding vector based on the channel state information of the legitimate user;

[0114] Based on the precoding vector, the transmission semantic information is precoded;

[0115] The precoded transmission semantic information is sent through a wireless channel.

[0116] Specifically, the transmitter generates a precoding vector according to the channel state information of the legitimate user, and uses the vector to precode the transmission semantic information, which is then sent through the wireless channel using a transmitter.

[0117] In some possible embodiments, f and h represent the precoding layer and the wireless channel layer, respectively. Where s represents the data to be sent, x represents the precoded data, and has the same structure as s, and z corresponds to the data received by the receiver, that is, the legitimate user and the potential eavesdropper. In the semantic communication system, in order to avoid the complex multiplication caused by the complex numbers in the training phase, two convolutional layers are used to equivalently represent the precoding and wireless channels for easy training. By dividing the channel s in the transmission phase into L time slots, each time slot sends N t complex symbols, then the signal sent by the user in the Lth time slot of the tth antenna is:

[0118]

[0119] Among them, l∈(1,2,…,L), t∈(1,2…,N t ). The index "1" in the superscript represents the real part, while "2" represents the imaginary part. In this case, the transmitted signal can be considered to be of shape 1×2×N t The feature map of L.

[0120] The original information M is extracted by the sender and artificial noise is added to obtain the semantic information x p , signal x pAfter preprocessing and wireless channel transmission, the signal received by the receiver is recorded as z. The receiver, that is, the legitimate user and the potential eavesdropper, uses their own decoder to decode, namely:

[0121] M d =Decoder(z) (7)

[0122] It is understandable that, as mentioned above, the process of training an end-to-end semantic communication system generally includes three processes: communication modeling, precoding and wireless channel transmission, and semantic communication framework training. The artificial noise generation model is an additional component of the semantic communication framework sender, which is used to protect the image to be sent without affecting the operation of the semantic communication framework.

[0123] In some optional embodiments, a semantic segmentation dataset can be used to train the semantic communication framework. For example, the dataset includes 1569 images of urban scenes, which are divided into 9 categories for semantic segmentation. For the semantic communication framework, the training parameters are set to a batch size of 12, a training round of 300, and a stochastic gradient descent is used as a deep learning optimizer with an initial learning rate of 0.02. The momentum coefficient is set to 0.9 to enhance optimization stability, while the L2 regularization coefficient (weight decay) is maintained at 0.0005 to reduce overfitting. In addition, a learning rate decay of 0.01 is applied, where the learning rate of each training iteration is reduced to 0.99 of its original value. The loss function adopted is the cross entropy loss.

[0124] The model structure is as follows Figure 3 As shown, it should be noted that in the semantic encoder, unless otherwise specified, all convolutional layers are the same, that is, the convolution kernel size is 3×3, the stride is 1, the padding is 1, and some maximum pooling is added between the convolutional layers for subsampling. Each Mini-inception module contains two parallel convolutional layers: one is a convolution kernel size of 3×3, a stride of 1, a padding of 1, and a dilation rate of 1, and the other is an expansion convolution layer with a dilation rate of 2. The output of Mini-inception is the cascade of the outputs of these two convolutional layers. The number of output channels on each convolutional layer in the Mini-inception module is equal to half the number of input channels of the Mini-inception block. In addition, each convolutional layer in the semantic encoder and semantic decoder is followed by a batch normalization layer, followed by a nonlinear activation function leaky rectified linear unit (Leaky-Relu). In the last three convolutional layers, the input image is compressed, reshaped, and normalized to meet the signal power limit. Finally, the output of the semantic encoder is the complex-valued wireless channel input symbol.

[0125] In the semantic decoder, since the number of antennas at the receiver may be different from the number of antennas at each transmitter, a post-process layer can be used to make the received complex-valued signal have the same dimension as the transmitted symbols. The input size of the semantic decoder is N r ×2×1200 / N t , assuming N r Can be N t The post-processing layer is a convolution layer with a convolution kernel size of 2×N. r , the step size is 2×N r , filled with 0, which has 2N t Convolution kernels are used, followed by batch normalization and Leaky-ReLU activation functions. Next, a convolution layer is used to restore the channels and an unpooling layer is used to restore the size. The output of the semantic decoder contains nine categories.

[0126] Furthermore, with the help of the trained semantic communication framework, a deep neural network for generating artificial noise can be trained.

[0127] It is understandable that the network needs to be jointly trained with the knowledge base as well as the eavesdropping channel based on the feedback from the semantic communication framework.

[0128] Specifically, the process of training a deep neural network for generating artificial noise includes two processes: problem definition and model construction and training.

[0129] Among them, for the problem definition, in order to enhance the protection from eavesdropping threats without compromising communication efficiency, the following can be used at the sending end: Figure 4 The artificial noise generation model Perturbation shown in the figure generates semantic-level artificial noise, and then adds this noise to the semantic information to be transmitted. This method can blur any original image and provide defense against eavesdropping attacks without the need for a retraining process.

[0130] That is: given x p =δ+x, x p Substituting into the communication model, the transmission results of the legitimate user are as follows:

[0131]

[0132] Among them, (HFx+n b ) is the result of continuous encoding and channel transmission of the semantic information to be transmitted, and HFδ is regarded as artificial noise controlled by the variable δ, where the channel matrix H and the precoding matrix F are known. Similarly, the signal at the potential eavesdropper is as follows:

[0133]

[0134] By observing the formula representation of legitimate users and potential eavesdroppers, it can be found that by controlling the value of δ to control the value of artificial noise HFδ and GFδ, the purpose of confusing potential eavesdroppers can be achieved, and this process can be achieved through deep neural networks.

[0135] Specifically, by using the loss function to specify the optimization direction of the model, the artificial noise GFδ can be used to generate adversarial perturbations to target semantic information (GFx+n e ) to guide the decoding result of semantic information to the disguised image. Therefore, in the presence of potential eavesdroppers, the problem of defensive adversarial attacks on semantic communication systems can be expressed as the following formula:

[0136]

[0137] where f(·) is the semantic decoder in the semantic communication system, and d(·,·) represents the semantic decoder used to calculate L ∞ norm, and ∈ represents a pre-specified threshold, d(x p ,x)≤∈ indicates L ∞ norm constraint, used to ensure that δ does not affect the correct decoding of x. p =δ+x, so the goal is to find a perturbation δ such that when added to the target input domain When we take any input image of , it will cause the decoder f(·) to misclassify the potential eavesdropper, while the legitimate user can use it for decoding.

[0138] The main difference between a legitimate user and an eavesdropper is the difference in H and G. In order to achieve the goal of distorting the decoding results of a potential eavesdropper, the generated artificial noise needs to be able to exploit the difference between the eavesdropped and legitimate channel state information CSI.

[0139] Furthermore, for model construction and training, the artificial noise generation model Perturbation consists of two interconnected networks, an encoder that mimics the semantic communication framework to ensure the retrieval of equivalent semantic information of the disguised image, and a U-Net architecture that takes input from the encoder and generates artificial noise of the same dimension.

[0140] Furthermore, in a specific implementation, given the small amount of semantic information data, the model can be kept in a four-layer structure compared to the original U-Net model. In some possible embodiments, the convolution kernel dimension is adjusted to 3×3, the step size is 1, and the padding is (1, 1) to ensure that the output of each layer matches the size of the input.

[0141] Except for the top layers in the semantic encoder and semantic decoder, all layers are normalized to reduce overfitting, and the Leaky ReLU activation function is used in all layers except the top layer of the semantic encoder to enhance the nonlinear representation. Artificial noise is limited to comply with L ∞ Norm constraint, where L ∞ The norm refers to making the norm of the vector the absolute value of the element with the largest absolute value among the elements of the vector, thereby minimizing its influence on the semantic information and ensuring that it reduces the decoding effect at the intended receiver.

[0142] At the same time, the loss function contains the following three items. The first item is to ensure that the legitimate user's decoding is correct, the second item is to ensure that the potential eavesdropper is successfully misled, and the third item is to ensure that the eavesdropper can decode correctly even though the result is false. That is, the loss function L is defined as:

[0143]

[0144] Among them, R b and R e Represent the actual information decoded by the legitimate user and the potential eavesdropper, respectively. represents the cross entropy loss, M b ′ represents the label of the image to be sent, and M t ′ represents the label of the disguised image that is intended to mislead potential eavesdroppers, and the superscript “'” represents the input domain The goal of this loss function is to make the eavesdropper R e The actual information decoded is close to M t ′, while ensuring R b Close to M b ′ and R e Cannot be decoded as M b ' b .

[0145] It should be noted that the training setup for the artificial noise generation model is very similar to that of the semantic communication framework. In actual implementation, the training batch size for the artificial noise generation model can be set to 12, there is no predefined maximum number of training epochs, and the training ends dynamically based on the behavior of the loss function. For example, when the model goes through 20 epochs and the loss function value does not decrease, the training stops. The optimization algorithm can be the stochastic gradient descent algorithm SGD, where the initial learning rate is 0.02, the momentum coefficient is 0.9, and the L2 regularization coefficient (weight decay) is 0.0005, which are used to stabilize the training and reduce over-adaptation, respectively. In addition, a learning rate decay of 0.01 can also be applied.

[0146] In summary, the overall architecture of the semantic disguise communication method provided by the above embodiment of the present invention is as follows: Figure 5 As shown by Figure 5 It can be seen that, on the one hand, this method uses the semantic information of the disguised image to generate artificial noise, and uses a low-complexity architecture to blur any original image, thereby providing defense against semantic communication eavesdropping attacks without the need for a retraining process. On the other hand, by training a neural network to capture channel features, artificial noise with different effects is generated when passing through different channels, so that by adding artificial noise to the semantic information, it is possible to resist eavesdropping attacks in semantic communication with the help of defensive adversarial attack technology, and mislead the decoding results of potential eavesdroppers, thereby achieving secure communication of the end-to-end semantic communication system during power inspection.

[0147] To sum up, by applying the semantic camouflage communication method provided by the embodiment of the present invention, confidential information transmission can be performed between devices based on defensive anti-attack security technology, thereby ensuring low-bandwidth and high-accuracy communication between the communicating parties, while being able to resist eavesdropping attacks from potential eavesdroppers and mislead potential eavesdroppers' decoding results and make them unaware, thereby ensuring that the semantic communication system in the power inspection system can resist eavesdropping attacks.

[0148] like Figure 6 As shown, the embodiment of the present invention also provides another semantic disguise communication method, which is applied to a target receiving end of a semantic communication system, where the target receiving end is a receiving end of a legitimate user in a deployment environment of the semantic communication system, and mainly includes the following steps:

[0149] 210. Obtain transmission semantic information.

[0150] The transmitted semantic information is semantic information with artificial noise added, and the semantic information is information extracted from the image to be sent by a semantic encoder of the semantic communication system.

[0151] 220. Input the transmission semantic information into a semantic decoder of the semantic communication system for decoding to obtain an image to be sent.

[0152] The semantic encoder and semantic decoder are jointly trained based on the channel state information of legitimate users and potential eavesdroppers in the deployment environment.

[0153] like Figure 7 As shown, the embodiment of the present invention also provides another semantic disguised communication method, which is applied to a disguised receiving end of a semantic communication system, wherein the disguised receiving end is a receiving end of a potential eavesdropper in a deployment environment of the semantic communication system, and mainly includes the following steps:

[0154] 310. Obtain transmission semantic information.

[0155] The transmitted semantic information is semantic information with artificial noise added, and the semantic information is information extracted from the image to be sent by a semantic encoder of the semantic communication system.

[0156] 320. Input the transmitted semantic information into a semantic decoder of the semantic communication system for decoding to obtain a camouflaged image.

[0157] The semantic encoder and semantic decoder are jointly trained based on the channel state information of legitimate users and potential eavesdroppers in the deployment environment.

[0158] In an optional embodiment, the artificial noise is generated by an artificial noise generation model.

[0159] Among them, the artificial noise generation model is a neural network model obtained by collaborative training with the semantic encoder and the semantic decoder.

[0160] A semantic camouflage communication system provided in an embodiment of the present invention is introduced below. The semantic camouflage communication system described below can be considered as a module architecture for implementing the semantic camouflage communication method provided in an embodiment of the present invention; the content described below can be cross-referenced with the above.

[0161] Optional, see Figure 8 , Figure 8 : is a structural block diagram of a semantic camouflage communication system provided by an embodiment of the present invention, and the system may include:

[0162] A first acquisition unit 10 is used to acquire semantic information, where the semantic information is information extracted from the image to be sent by a semantic encoder of the semantic communication system;

[0163] The first processing unit 20 adds artificial noise to the semantic information to obtain transmission semantic information; the artificial noise is used to ensure that both legitimate users and potential eavesdroppers in the deployment environment of the semantic communication system can decode the transmission semantic information, the legitimate users decode the transmission semantic information to obtain the image to be sent, and the potential eavesdroppers decode the transmission semantic information to obtain the disguised image;

[0164] The sending unit 30 is used to precode and send the transmission semantic information.

[0165] Optional, see Fig. 9 , Fig. 9 : is a structural block diagram of another semantic camouflage communication system provided by an embodiment of the present invention, and the system may include:

[0166] The second acquisition unit 40 is used to obtain transmission semantic information, where the transmission semantic information is semantic information with artificial noise added, and the semantic information is information extracted from the image to be sent by a semantic encoder of the semantic communication system;

[0167] The second processing unit 50 is used to input the transmission semantic information into a semantic decoder of the semantic communication system, decode it, and obtain an image to be sent;

[0168] The semantic encoder and semantic decoder are jointly trained based on the channel state information of legitimate users and potential eavesdroppers in the deployment environment of the semantic communication system.

[0169] Optional, see Fig.10 , Fig.10 : is a structural block diagram of another semantic camouflage communication system provided by an embodiment of the present invention, and the system may include:

[0170] The third acquisition unit 60 is used to obtain transmission semantic information, where the transmission semantic information is semantic information with artificial noise added, and the semantic information is information extracted from the image to be sent by a semantic encoder of the semantic communication system;

[0171] The third processing unit 70 is used to input the transmission semantic information into a semantic decoder of the semantic communication system, decode it, and obtain a disguised image;

[0172] The semantic encoder and semantic decoder are jointly trained based on the channel state information of legitimate users and potential eavesdroppers in the deployment environment of the semantic communication system.

[0173] Optionally, an embodiment of the present invention further provides a power grid, which includes a power system body and a semantic camouflage communication system as provided in any of the above embodiments.

[0174] Below, reference Fig.11 To describe the electronic device provided by the embodiment of the present application, the electronic device provided by the embodiment may include: at least one processor 100, at least one communication interface 200, at least one memory 300 and at least one communication bus 400;

[0175] In the embodiment of the present invention, the number of the processor 100, the communication interface 200, the memory 300, and the communication bus 400 is at least one, and the processor 100, the communication interface 200, and the memory 300 communicate with each other through the communication bus 400; obviously, Fig.11 The communication connections shown for the processor 100, the communication interface 200, the memory 300, and the communication bus 400 are merely optional;

[0176] Optionally, the communication interface 200 may be an interface of a communication module, such as an interface of a GSM module; the processor 100 may be a central processing unit CPU, or an application specific integrated circuit ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of the present invention.

[0177] The memory 300 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk memory.

[0178] The processor 100 is specifically used to execute the application program in the memory to implement the steps of the above-mentioned crane hoisting wire rope installation control method.

[0179] The above embodiments are merely examples for the purpose of clear explanation, and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the invention.

Claims

1. A semantic camouflage communication method, applied to the sending end of a semantic communication system, characterized in that: The method comprises: Acquire semantic information, where the semantic information is information extracted from the image to be sent by a semantic encoder of the semantic communication system; Adding artificial noise to the semantic information to obtain transmission semantic information; the artificial noise is used to ensure that both legitimate users and potential eavesdroppers in the deployment environment of the semantic communication system can decode the transmission semantic information, the legitimate users decode the transmission semantic information to obtain the image to be sent, and the potential eavesdroppers decode the transmission semantic information to obtain the camouflaged image; The transmission semantic information is precoded and sent.

2. The semantic disguise communication method according to claim 1, characterized in that: The obtaining of semantic information includes: Acquire the image to be sent; Inputting the image to be sent into a semantic encoder of the semantic communication system to extract semantic information to obtain the semantic information; The semantic encoder is jointly trained together with the semantic decoder of the semantic communication system based on channel state information of legitimate users and potential eavesdroppers in the deployment environment of the semantic communication system.

3. The semantic disguise communication method according to claim 2, characterized in that: The adding artificial noise to the semantic information comprises: Acquire the camouflage image; Inputting the disguised image into an artificial noise generation model to obtain the artificial noise output by the artificial noise generation model; applying the artificial noise to the semantic information; The artificial noise generation model is a neural network model obtained by collaborative training with the semantic encoder and the semantic decoder.

4. The semantic disguise communication method according to claim 3, characterized in that: The artificial noise generation model includes two layers of interconnected networks; The first layer of the network is an encoder that imitates a semantic communication framework, which is used to ensure the retrieval of equivalent semantic information of the disguised image; The second layer network is a U-Net architecture, which is used to receive the disguised image and generate the artificial noise with the same dimension as the disguised image.

5. The semantic disguise communication method according to claim 4, characterized in that: The artificial noise generation model enables the potential eavesdropper to decode the actual information R e The label M close to the disguised image t ′ , the actual information R obtained by the legitimate user decoding b The tag M close to the image to be sent b ′ , and R e Not decoded as M b ′ Construct the loss function L for the target: in, represents the cross entropy loss.

6. The semantic disguise communication method according to claim 5, characterized in that: The artificial noise complies with L ∞ Norm constraint.

7. The semantic disguise communication method according to claim 6, characterized in that: Based on a preset formula, the artificial noise is used to ensure that both the legitimate user and the potential eavesdropper can decode the transmission semantic information, and the legitimate user decodes to obtain the image to be sent, while the potential eavesdropper decodes to obtain a disguised image; The preset formula is expressed as: arg Distance[f(x p ,G),M t ] d(x p ,x)≤∈, Where f(·) represents the semantic decoder and d(·,·) represents the function used to calculate L ∞ norm function, ∈ represents a pre-specified threshold, d(x p ,x)≤∈ indicates L ∞ Norm constraint.

8. The semantic disguise communication method according to claim 1, characterized in that: The precoding and sending of the transmission semantic information includes: Generate a precoding vector based on the channel state information of the legitimate user; Precoding the transmission semantic information based on the precoding vector; The precoded transmission semantic information is sent through a wireless channel.

9. A semantic camouflage communication method, applied to a target receiving end of a semantic communication system, wherein the target receiving end is a receiving end of a legitimate user in a deployment environment of the semantic communication system, characterized in that: The method comprises: Acquire transmission semantic information, where the transmission semantic information is semantic information with artificial noise added, and the semantic information is information extracted from the image to be sent by a semantic encoder of the semantic communication system; Inputting the transmission semantic information into a semantic decoder of the semantic communication system for decoding to obtain the image to be sent; The semantic encoder and the semantic decoder are jointly trained based on channel state information of legitimate users and potential eavesdroppers in the deployment environment.

10. A semantic camouflage communication method, applied to a camouflage receiving end of a semantic communication system, wherein the camouflage receiving end is a receiving end of a potential eavesdropper in a deployment environment of the semantic communication system, characterized in that: The method comprises: Acquire transmission semantic information, where the transmission semantic information is semantic information with artificial noise added, and the semantic information is information extracted from the image to be sent by a semantic encoder of the semantic communication system; Inputting the transmission semantic information into a semantic decoder of the semantic communication system for decoding to obtain a disguised image output by the semantic decoder; The semantic encoder and the semantic decoder are jointly trained based on channel state information of legitimate users and potential eavesdroppers in the deployment environment.

11. The semantic disguise communication method according to claim 9 or 10, characterized in that: The artificial noise is generated by an artificial noise generation model, and the artificial noise generation model is a neural network model obtained by collaborative training with the semantic encoder and the semantic decoder.

12. A semantic camouflage communication system, characterized in that: include: A first acquisition unit, configured to acquire semantic information, wherein the semantic information is information extracted from an image to be sent by a semantic encoder of the semantic communication system; A first processing unit is configured to add artificial noise to the semantic information to obtain transmission semantic information; the artificial noise is used to ensure that both legitimate users and potential eavesdroppers in the deployment environment of the semantic communication system can decode the transmission semantic information, the legitimate users decode the transmission semantic information to obtain the image to be sent, and the potential eavesdroppers decode the transmission semantic information to obtain the camouflaged image; A sending unit is used to precode and send the transmission semantic information.

13. A semantic camouflage communication system, characterized in that: include: A second acquisition unit is used to obtain transmission semantic information, where the transmission semantic information is semantic information with artificial noise added, and the semantic information is information extracted from the image to be sent by a semantic encoder of the semantic communication system; A second processing unit, configured to input the transmission semantic information into a semantic decoder of the semantic communication system for decoding to obtain the image to be sent; The semantic encoder and the semantic decoder are jointly trained based on channel state information of legitimate users and potential eavesdroppers in the deployment environment of the semantic communication system.

14. A semantic camouflage communication system, characterized in that: include: A third acquisition unit is used to obtain transmission semantic information, where the transmission semantic information is semantic information with artificial noise added, and the semantic information is information extracted from the image to be sent by a semantic encoder of the semantic communication system; A third processing unit, configured to input the transmission semantic information into a semantic decoder of the semantic communication system for decoding to obtain a disguised image; The semantic encoder and the semantic decoder are jointly trained based on channel state information of legitimate users and potential eavesdroppers in the deployment environment of the semantic communication system.

15. A power grid, characterized in that: It comprises a power system body and a semantic camouflage communication system as described in any one of claims 12 to 14.

16. An electronic device, characterized in that: The electronic device comprises: processor; a memory for storing instructions executable by the processor; The processor is used to execute the semantic camouflage communication method described in any one of claims 1 to 8 or any one of claims 9 to 11.

Images