Unmanned aerial vehicle inspection privacy protection method and device, electronic equipment and medium

Abstract

The embodiment of the invention discloses an unmanned aerial vehicle inspection privacy protection method and device, electronic equipment and a medium. The method comprises the following steps: training a semantic encoder and a semantic decoder for semantic communication; performing semantic information extraction on the to-be-transmitted information by adopting a semantic encoder, and adding the generated semantic artificial noise into the semantic information to obtain the to-be-transmitted semantic information; generating a pre-coding vector according to the channel state information, performing pre-coding processing on the semantic information to be transmitted by adopting the pre-coding vector, and sending the pre-coded semantic information to be transmitted; and decoding the pre-coded semantic information to be transmitted by adopting a semantic decoder to obtain target information. By adopting the scheme, the semantic artificial noise is hidden in the semantic communication process, and the specified privacy information is protected in a targeted manner; the information difference transmission is carried out based on the end-to-end semantic communication system, so that the receiving ends of different channels decode different results, and the privacy security in the unmanned aerial vehicle inspection system is ensured.

Description

Technical Field

[0001] Embodiments of the present invention relate to the field of semantic communication technology, and in particular to a method, device, electronic device, and medium for protecting privacy during drone inspections. Background Art

[0002] Grid inspection is a crucial task to ensure the stable operation and efficient management of power systems. Traditional inspection methods often rely on manual labor, especially when inspecting high-altitude power facilities. This is not only labor-intensive and inefficient, but also poses safety risks. To address these issues, the application of drone technology has revolutionized power inspection. Drones can fly unmanned and collect data through a variety of sensors and high-definition cameras. These devices can accurately capture detailed images of cables and other high-altitude facilities, allowing inspectors to remotely assess the condition of equipment. Furthermore, drones can access locations that are difficult or dangerous for humans to reach, such as confined spaces or in extreme weather conditions. This is crucial for promptly identifying potential safety hazards.

[0003] With the continued construction and development of power grids, the workload of drone inspections will gradually increase. Drones primarily collect image data during inspections, which is crucial for monitoring the status of power facilities. However, transmitting large amounts of image data poses significant challenges to drones' communication capabilities and endurance. In past power grid inspections, not all collected data was necessary, especially when it contained sensitive information, such as images of passersby, which could raise privacy concerns.

[0004] Therefore, how to protect privacy information during the inspection process based on secure communication is a technical problem that needs to be solved urgently by those skilled in the art. Summary of the Invention

[0005] The embodiments of the present invention provide a drone inspection privacy protection method, device, electronic device and storage medium to achieve hiding of semantic artificial noise in the drone semantic communication process, so as to provide targeted protection for designated privacy information, thereby ensuring privacy security in the drone inspection system.

[0006] In a first aspect, an embodiment of the present invention provides a method for protecting privacy during drone inspections, comprising:

[0007] Training an end-to-end semantic communication system for semantic communication; the end-to-end semantic communication system includes a semantic encoder and a semantic decoder;

[0008] A semantic encoder is used to extract semantic information from the information to be transmitted, and the generated semantic artificial noise is added to the semantic information to obtain the semantic information to be transmitted;

[0009] Generate a precoding vector based on the channel state information, perform precoding processing on the semantic information to be transmitted using the precoding vector, and send the precoded semantic information to be transmitted;

[0010] A semantic decoder is used to decode the pre-coded semantic information to be transmitted to obtain target information.

[0011] In a second aspect, an embodiment of the present invention further provides a drone inspection privacy protection device, comprising:

[0012] A semantic communication system building module, used to train an end-to-end semantic communication system for semantic communication; the end-to-end semantic communication system includes a semantic encoder and a semantic decoder;

[0013] An artificial noise adding module is used to extract semantic information from the information to be transmitted using a semantic encoder, and add the generated semantic artificial noise to the semantic information to obtain the semantic information to be transmitted;

[0014] A semantic information encoding module, configured to generate a precoding vector based on the channel state information, perform precoding processing on the semantic information to be transmitted using the precoding vector, and send the precoded semantic information to be transmitted;

[0015] The semantic information decoding module is used to use a semantic decoder to decode the pre-coded semantic information to be transmitted to obtain target information.

[0016] In a third aspect, an embodiment of the present invention further provides an electronic device, the electronic device comprising:

[0017] one or more processors;

[0018] a storage device for storing one or more programs;

[0019] When the one or more programs are executed by the one or more processors, the one or more processors implement the drone inspection privacy protection method described in any embodiment of the present invention.

[0020] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the drone inspection privacy protection method described in any embodiment of the present invention is implemented.

[0021] In a fifth aspect, an embodiment of the present invention further provides a computer program product, comprising a computer program, which, when executed by a processor, implements the drone inspection privacy protection method as described in any embodiment of the present invention.

[0022] The embodiment of the present invention provides a privacy protection method, device, electronic device and storage medium for drone inspection. The method comprises training a semantic encoder and a semantic decoder for semantic communication; using the semantic encoder to extract semantic information from the information to be transmitted, and adding the generated semantic artificial noise to the semantic information to obtain the semantic information to be transmitted; generating a precoding vector based on channel state information, and using the precoding vector to precode the semantic information to be transmitted, and sending the precoded semantic information to be transmitted; using the semantic decoder to decode the precoded semantic information to be transmitted to obtain the target information. The technical solution of the embodiment of the present invention is used to hide the semantic artificial noise in the drone semantic communication process to provide targeted protection for the designated privacy information; based on the constructed end-to-end semantic communication system and the defensive anti-attack security technology between the drone and the receiving end, differential transmission of information is performed, which enables the receiving ends of different channels to decode different results, thereby ensuring privacy security in the drone inspection system. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Other features, objects, and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments made with reference to the following drawings. The drawings are for the purpose of illustrating preferred embodiments only and are not to be considered as limiting the present invention. Like reference characters are used throughout the drawings to denote like parts. In the drawings:

[0024] Figure 1 This is a flow chart of a privacy protection method for drone inspections provided by the first embodiment of the present invention;

[0025] Figure 2 This is a flow chart of a privacy protection method for drone inspections provided by the second embodiment of the present invention;

[0026] Figure 3 This is a schematic diagram of the structure of a traditional semantic communication provided by the second embodiment of the present invention;

[0027] Figure 4 This is a schematic diagram of the structure of semantic communication provided by the second embodiment of the present invention;

[0028] Figure 5 This is a schematic diagram of joint optimization of an artificial noise and semantic communication system provided by the second embodiment of the present invention;

[0029] Figure 6This is a schematic diagram of a convolutional layer equivalently representing precoding and wireless channels provided by the second embodiment of the present invention;

[0030] Figure 7 This is a structural diagram of a privacy protection device for drone inspections provided by the third embodiment of the present invention;

[0031] Figure 8 This is a structural diagram of an electronic device provided in Example 4 of the present invention. DETAILED DESCRIPTION

[0032] The present invention will be further described in detail below with reference to the accompanying drawings and examples. It will be understood that the specific embodiments described herein are intended only to illustrate the present invention and are not intended to limit the present invention. It should also be noted that, for ease of description, the accompanying drawings only illustrate portions relevant to the present invention, not all structures.

[0033] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flow charts. Although the flow charts describe the various operations (or steps) as sequential processes, many of the operations (or steps) therein can be implemented in parallel, concurrently, or simultaneously. In addition, the order of the various operations can be rearranged. The process can be terminated when its operation is completed, but can also have additional steps not included in the accompanying drawings. The process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0034] The acquisition, storage, use, and processing of data in the technical solution of this application are in compliance with the relevant provisions of national laws and regulations. It should be noted that in the embodiments of this application, certain software, components, or models, etc., which are already available in the industry, may be mentioned. These should be considered as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. It does not mean that the applicant has already or necessarily used such a solution.

[0035] It should be noted that the terms "first," "second," and the like in the description and claims of the present invention and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate, such that the embodiments of the present invention described herein can be practiced in an order other than that illustrated or described herein.

[0036] Example 1

[0037] Figure 1This is a flow chart of a drone inspection privacy protection method provided in an embodiment of the present invention. This embodiment is applicable to drone inspection privacy protection. The method of this embodiment can be executed by a drone inspection privacy protection device, which can be implemented in hardware and / or software. The device can be configured in a drone inspection privacy protection server. The method specifically includes the following steps:

[0038] S110. Train an end-to-end semantic communication system for semantic communication.

[0039] Among them, drones can cover a large area in a short period of time during power inspections, greatly improving the speed and efficiency of inspections while reducing the need for manpower and the economic losses caused by downtime for maintenance. More importantly, they enhance safety, preventing workers from being directly exposed to high-voltage environments, significantly reducing operational risks, and enabling accurate detection of facilities such as cables through the use of advanced sensors. For example, thermal imaging technology is used to identify potential fault points such as overheating. With technological advancements, drone inspection systems are also constantly developing and improving. In the future, they are expected to further integrate artificial intelligence analysis functions to achieve automatic analysis and processing of collected data, thereby identifying problems more quickly and accurately. This will make power inspections more intelligent and automated, providing strong support for the continued stable operation of power systems.

[0040] The drone acquires the semantic information to be transmitted and transmits it to the line patrolman and base station receivers via the end-to-end semantic communication system. The end-to-end semantic communication system includes a semantic encoder and a semantic decoder. A semantic encoder is a neural network model specifically designed for processing natural language text. Its primary function is to convert text into low-dimensional vector representations that capture the semantic and grammatical structure of the language. The semantic decoder decodes the encoded semantic information to obtain the original information to be transmitted.

[0041] The end-to-end semantic communication system provided by the embodiment of the present invention includes N t UAV with 1 antenna, specifically N r The base station receiving end has N antennas and e The line patrolman receiving end has an antenna; wherein the base station receiving end is used to receive signals and upload data instead of the cloud service.

[0042] In an optional scheme of an embodiment of the present invention, a semantic segmentation dataset is used to train an end-to-end semantic communication system framework. The dataset includes 1,569 urban scene images, which are divided into 9 categories for semantic segmentation as an example. For the semantic communication framework, the training parameters are set to a batch size of 12, the number of training rounds is set to 300, and stochastic gradient descent is used as a deep learning optimizer, adopting a dynamic learning rate strategy. The momentum coefficient is set to 0.9 to enhance optimization stability, and the L2 regularization coefficient (weight decay) is maintained at 0.0005 to reduce overfitting. The loss function used is cross entropy loss.

[0043] In the semantic encoder, unless otherwise noted, all convolutional layers are identical, with a kernel size of 3×3, a stride of 1, and padding of 1, with some max pooling added between layers for subsampling. Each Mini-Inception module consists of two parallel convolutional layers: one with a kernel size of 3×3, a stride of 1, padding of 1, and a dilation rate of 1; and a second dilated convolutional layer with a dilation rate of 2. The output of the Mini-Inception module is the concatenation of the outputs of these two convolutional layers. The number of output channels of 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-decoder is followed by a batch normalization layer, followed by a nonlinear activation function, the Leaky Rectified Linear Unit (Leaky-ReLU). In the last three convolutional layers, the input image is compressed, reshaped, and normalized to meet signal power constraints. Finally, the output of the encoder is the complex-valued wireless channel input symbol.

[0044] In the semantic decoder, since the number of antennas at the receiving end may be different from the number of antennas at each transmitting end, a post-processing layer is used to make the dimension of the received complex-valued signal the same as the transmitted symbols. The input size of the 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 Convolutional kernels are added, followed by batch normalization (Batchnorm) and Leaky-ReLU activation functions. Next, a convolutional layer is used to restore the channels, and an unpooling layer is used to restore the size. The decoder output contains nine categories.

[0045] The end-to-end semantic communication system constructed in the embodiment of the present invention cleverly utilizes the differences between channels, so that the sending end only needs to transmit once, and different receiving ends can decode different information. This feature is used to achieve privacy protection in semantic communication.

[0046] S120 , extracting semantic information from the information to be transmitted using a semantic encoder, and adding generated semantic artificial noise to the semantic information to obtain the semantic information to be transmitted.

[0047] The semantic artificial noise is generated by disguised information used to mislead an eavesdropper, and the disguised information used to mislead an eavesdropper is input into the constructed matrix neural network for iteration to determine the semantic artificial noise.

[0048] In the embodiment of the present invention, a semantic encoder is used to extract semantic information from the information to be transmitted, and semantic artificial noise generated by a matrix neural network is added to the semantic information to obtain the semantic information to be transmitted.

[0049] S130 , generating a precoding vector according to the channel state information, performing precoding processing on the semantic information to be transmitted using the precoding vector, and sending the precoded semantic information to be transmitted.

[0050] The channel state information can refer to the channel state information between the drone and the base station, and the channel state information between the drone and the patrolman. The channel state information is a kind of data describing the channel in wireless communication, which represents the propagation characteristics of the communication link and describes the combined influence of multiple effects such as scattering, fading, and power attenuation in the channel. Since the drone is located at a high position, the patrolman and the drone are often far away and there are obstructions. Therefore, assuming that the wireless channel at the patrolman's receiving end obeys Rayleigh fading, the communication channel from the drone to the patrolman is defined as In addition, the communication channel between the UAV and the base station can be expressed as It obeys Rice fading because base stations are generally deployed at high locations and there is often no obstruction between them and drones.

[0051] In this embodiment of the present invention, a precoding vector is generated based on the channel state information, and the precoding vector is used to precode the semantic information to be transmitted. The precoded semantic information to be transmitted is then transmitted via a wireless channel. The precoding vector is used to adjust the semantic information to be transmitted so that it can better adapt to wireless channel conditions.

[0052] S140: Use a semantic decoder to decode the pre-coded semantic information to be transmitted to obtain target information.

[0053] The target information includes the original transmission information obtained by the base station receiving end and the original transmission information without privacy information obtained by the line patrol receiving end. Let x represent the signal of the image message M through joint semantic coding and channel coding, and the precoding matrix is ​​given by Indicates that H represents the channel matrix between the drone and the base station, and G represents the channel matrix between the drone and the patrolman. The signals received by the base station and the patrolman are:

[0054] y b =HFx+n b

[0055] y e =GFx+n e

[0056] Among them, n b and n e It is obedience and In this case, the base station and the patrolman receive the corresponding signal y b and y e After that, the same decoder will be used to perform channel decoding and semantic decoding on the signal to obtain the respective decoding results M' b and M′ e .

[0057] An embodiment of the present invention provides a privacy protection method for drone inspections, which is achieved by training an end-to-end semantic communication system for semantic communication; the end-to-end semantic communication system includes a semantic encoder and a semantic decoder; the semantic encoder is used to extract semantic information from information to be transmitted, and the generated semantic artificial noise is added to the semantic information to obtain the semantic information to be transmitted; a precoding vector is generated based on channel state information, and the precoding vector is used to precode the semantic information to be transmitted, and the precoded semantic information to be transmitted is sent; the semantic decoder is used to decode the precoded semantic information to be transmitted to obtain target information. Using the technical solution of the embodiment of the present invention, the semantic artificial noise is hidden in the drone semantic communication process to provide targeted protection for designated privacy information; based on the constructed end-to-end semantic communication system and the defensive anti-attack security technology between the drone and the receiving end, differential information transmission is performed, which enables the receiving ends of different channels to decode different results, thereby ensuring privacy security in the drone inspection system.

[0058] Example 2

[0059] Figure 2This is a flow chart of a privacy protection method for drone inspections provided in an embodiment of the present invention. The embodiment of the present invention further optimizes the above embodiment on the basis of the above embodiment, and the embodiment of the present invention can be combined with various optional solutions in one or more of the above embodiments. Figure 2 As shown, the drone inspection privacy protection method provided in the embodiment of the present invention may include the following steps:

[0060] S210. Train an end-to-end semantic communication system for semantic communication; the end-to-end semantic communication system includes a semantic encoder and a semantic decoder.

[0061] Among them, see Figure 3 as well as Figure 4 In traditional semantic communication, transmitting different content to different users requires the use of multiple transmitters or sequential transmission, which consumes double the communication resources and poses the risk of eavesdropping. The end-to-end semantic communication system proposed in this embodiment cleverly exploits the differences between channels, enabling the transmitter to transmit only once, while different receivers can decode different information. This feature enables privacy protection in semantic communication.

[0062] As an optional but non-limiting implementation, the training of the end-to-end semantic communication system for semantic communication includes but is not limited to steps A1-A2:

[0063] Step A1: Estimating the wireless channel between the UAV and the receiving end to determine the channel state information; the receiving end includes a base station receiving end and a line patrol receiving end.

[0064] Step A2: Jointly train the semantic encoder and the semantic decoder based on the channel state information to construct an end-to-end semantic communication system for semantic communication.

[0065] When building an end-to-end semantic communication system for semantic communication, the first step is to accurately estimate the wireless channel between the drone and its communication partner (such as a lineman or base station). This step typically involves sending pilot signals and using them to infer channel characteristics such as multipath effects, fading, and noise levels. The obtained channel state information is then used for subsequent neural network training and precoding vector calculation.

[0066] After obtaining the channel state information, the next task is to jointly train the semantic communication encoder and decoder. This is usually performed on a knowledge base containing a large amount of relevant data. The goal of training is to enable the encoder and decoder to effectively cooperate in a specific wireless environment and accurately transmit and recover semantic information even in the presence of noise and interference.

[0067] Jointly training the semantic encoder and semantic decoder ensures that the semantic encoder can automatically adjust to the channel characteristics when encoding information. This allows the semantic decoder to accurately recover the original information even if the signal is affected by channel noise during transmission, minimizing information loss and errors. This comprehensive training method optimizes the performance of the end-to-end semantic communication system provided by embodiments of the present invention, improving the security and reliability of information transmission.

[0068] S220: Use a semantic encoder to extract semantic information from the information to be transmitted.

[0069] Among them, semantic extraction requires using the semantic encoder described in the embodiment of the present invention to extract semantic information from the information to be transmitted, which can be expressed as:

[0070] x=Encoder(M)

[0071] Where Encoder is the semantic encoder, M is the original information, and x is the semantic information extracted by the semantic encoder from the information to be transmitted.

[0072] S230: Create a matrix neural network with the same output dimension as the semantic encoder, and use the matrix neural network to generate semantic artificial noise.

[0073] The matrix neural network acts as a carrier for semantic artificial noise, generated from disguised information designed to mislead eavesdroppers. Artificial noise δ, an additional component on the transmitter side of the end-to-end semantic communication system framework, protects private information within the transmitted information without affecting the operation of the semantic communication framework. This noise must be optimized for the lineman's channel characteristics without affecting base station communication efficiency. This noise must be resistant to adversarial attacks against the semantic decoder after passing through the lineman's wireless channel, protecting private information. Furthermore, the base station channel must be robust to this noise, ensuring it does not affect communication efficiency.

[0074] Given p(x) = x + δ, substitute p(x) into the communication model. The transmission results of the base station channel are as follows:

[0075] y b =HFp(x)+n b

[0076] =HF(x+δ)+n b

[0077] =(HFx+n b )+HFδ

[0078] Among them, (HFx+n b ) is the result of continuous encoding and channel transmission of the semantic information to be transmitted. HFδ is regarded as artificial noise controlled by the variable δ. The channel matrix H and precoding matrix F between the drone and the base station are known. Similarly, the signal at the inspector is as follows:

[0079] y e =GFx p +n e

[0080] =(GFx+n e )+GFδ

[0081] By observing the formula for the signals received by the base station and the inspector, it is found that the value of artificial noise HFδ and GFδ can be controlled by controlling the value of δ. This process can be achieved through deep neural networks. 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 mislead the decoding results of semantic information. Specifically, the algorithm will mislead the semantic decoding results of each pixel of private information to be hidden to the background label, that is, the background of the target is used to replace the hidden target in the decoding result, thereby achieving the purpose of privacy protection. Therefore, the problem can be modeled as:

[0082] arg minDistance[M' e ,C p(x),ba ]

[0083]

[0084] d(p(x),x)≤∈,

[0085] Among them C p(x),ba is the true value label representing the background of p(x) during the eavesdropper’s decoding process, C x represents the true value label of x, It is defined as the adversarial attack space against the knowledge base, where each element is an adversarial attack against the semantic decoder, and d(·,·) represents the adversarial attack used to calculate L ∞ The first constraint in the above formula ensures that the base station can correctly decode without being affected by artificial noise. The second constraint maintains the artificial noise in the adversarial attack space, so that it can maintain the effectiveness of the adversarial attack. The third constraint represents the power constraint.

[0086] The primary difference between the patrolman channel and the base station channel is the difference in H and G. To modify the patrolman's decoding results, the generated artificial noise must be able to exploit the difference between eavesdropped and legitimate CSI. Therefore, in this embodiment of the present invention, a matrix neural network δ with the same dimensions as the semantic encoder output is created as the carrier of the artificial noise. This matrix is ​​added to the semantic information output by the encoder to implement the process of superimposing noise on the signal. This noise is obtained through iterative optimization.

[0087] As an optional but non-limiting implementation, a matrix neural network with the same output dimension as the semantic encoder is created, and the matrix neural network is used to generate semantic artificial noise, including but not limited to steps B1-B2:

[0088] Step B1: Initialize the artificial noise into a matrix with the same dimension as the output of the semantic encoder; the artificial noise is implemented in the form of universal adversarial perturbation.

[0089] Step B2: Iteratively optimize the initialized artificial noise in the matrix neural network to obtain semantic artificial noise.

[0090] First, the artificial noise is initialized to a matrix δ with the same dimension as the encoder output. This artificial noise is implemented in the form of a universal adversarial perturbation. That is, after optimization, it can have an effect on the data of the entire dataset without having to find a specific perturbation for each image. The initialized artificial noise is then iteratively optimized in the matrix neural network. The iterative optimization of the initialized artificial noise in the matrix neural network is expressed as:

[0091] δ (0) =0

[0092]

[0093] Among them, δ (0) =0 means that the artificial noise after initialization is initialized to 0; δ n Represents the artificial noise matrix after the nth iteration; function Clip ∈ {·} is used to ensure that the proposed constraint condition can be satisfied, which is to limit the value of the artificial noise δ to [-∈,∈], where ∈ is the disturbance size; y target The label or output result of the prediction result; represents the perturbation increment after the nth iteration; where α is the step size parameter used to control the perturbation update amplitude during the iteration process; sgn is the sign function used to return the sign of each element; represents the gradient of the loss function J with respect to the information to be transmitted x; f d(p(x)) represents the result predicted by the model; p(x) represents the semantic information to be transmitted.

[0094] The artificial noise δ is limited to the range [-∈,∈] to ensure that the perturbation is small enough and imperceptible to the semantic decoder of the eavesdropper. Unlike the natural image input in traditional adversarial attack tasks, the input in this paper is semantic information, so ∈ is set to 4.5 to meet the needs of the scenario. α is the step size parameter used to control the perturbation update amplitude during the iteration process, and is set to 1e-1 in this invention. The loss function J is defined as follows:

[0095]

[0096] in Represents the total number of samples in the knowledge base and is used to normalize the loss function. and They represent hidden target pixels and non-target pixels, respectively. cs (·) represents the cross entropy loss. It represents the cross entropy loss between the prediction result of the hidden target pixel and the preset misleading result, which is then multiplied by the weight w. Similarly, the cross entropy loss between non-target pixels and their true labels is calculated and then multiplied by a weight (1-w). The weight w is used to weigh the importance of different targets for the task. In this embodiment of the present invention, it is set to 0.9, indicating that the hidden target is more important.

[0097] S240 , superimposing the generated semantic artificial noise onto the extracted semantic information to obtain the semantic information to be transmitted.

[0098] Among them, the generated semantic artificial noise is applied to the extracted semantic information, and the actual semantic information to be transmitted p(x) is obtained.

[0099] p(x)=x+δ

[0100] Furthermore, after obtaining the semantic information to be transmitted with artificial noise added, a neural network is used to equivalently replace the precoding and channel transmission processes, wherein the precoding, channel transmission and semantic decoding processes include:

[0101] Convolutional neural networks f and h are used to represent the precoding layer and the wireless channel layer, respectively, to avoid complex multiplication caused by complex numbers in the training phase. s represents the data to be sent; x represents the precoded data, and x and s have the same structure; z corresponds to the data received by the receiver, that is, the inspector and the cloud server of the archive department. In actual scenarios, the cloud server needs to transmit through a nearby base station. The signal s in the transmission phase is divided into L time slots, and each time slot sends Nt complex symbols, the signal sent by the user in the Lth time slot of the tth antenna is:

[0102]

[0103] where l∈(1, 2, ..., L), t∈(1, 2, ..., N t ), j represents a complex number. The index "1" in the superscript represents the real part, while "2" represents the imaginary part. In this case, the transmitted signal can be regarded as having a shape of 1×2×N t The feature map of L.

[0104] The original information M is extracted by the sender and artificial noise is added to obtain semantic information p(x). After the signal p(x) is preprocessed and transmitted through the wireless channel, the signal received by the receiver is recorded as z. The receiver, i.e., the inspector and the cloud server, respectively uses their own decoder to decode, i.e.

[0105] M d =Decoder(z)

[0106] Among them, for the cloud server, the decoding result M d The original information M is consistent with the original information M. However, for the inspector, since the transmitted signal contains artificial noise δ, after passing through the wireless channel, the noise will become an adversarial attack against the decoder, misleading the decoding result of the inspector's semantic decoder. The specific result is to convert the inspector's decoding result M d The private information hidden in the image is replaced with its background.

[0107] As an optional but non-limiting implementation, after obtaining the semantic information to be transmitted, the method further includes but is not limited to steps C1-C3:

[0108] Step C1: The semantic encoder and semantic decoder are jointly trained on the UAV’s lineman channel and the base station channel to obtain a pre-trained semantic communication system.

[0109] Step C2: Add semantic artificial noise to the semantic encoder for adversarial training, fine-tune the pre-trained semantic communication system, and obtain the target semantic communication system.

[0110] Step C3: After obtaining the target semantic communication system, the semantic artificial noise is iteratively optimized in an iterative manner to obtain the target semantic artificial noise.

[0111] Among them, the above embodiment describes the iterative process of universal perturbation. However, this alone is not enough to ensure that semantic communication can find a balance between normal transmission at the base station and modification of the inspector's decoding results. During the training process, if only artificial noise perturbation and the inspector channel are jointly trained, the decoding results at the base station will also be affected. The difference between the inspector channel and the base station channel is not enough to turn the universal perturbation into meaningless noise, and it will still have an impact to a certain extent. This impact exists at the semantic level, and once the semantic information is decoded, this impact will be amplified, resulting in a sharp decline in communication performance.

[0112] In order to protect the communication efficiency of the base station from interference from adversarial attacks, the embodiment of the present invention adopts a defense strategy commonly used in the field of deep learning - adversarial training to fine-tune the model. Figure 5 During the training process, adversarial examples caused by artificial noise are introduced into the training to enhance the robustness of the model against adversarial attacks. By combining the fine-tuning of the semantic communication system with the iterative process of universal perturbations, a training process similar to that of generative adversarial networks is obtained.

[0113] First, the semantic encoder and semantic decoder are jointly trained on the inspector channel and the base station channel of the UAV to obtain a pre-trained end-to-end semantic communication system. After that, the end-to-end semantic communication system is fine-tuned and the artificial noise is iteratively optimized. Specifically, in the first step, artificial noise is added to the semantic encoder for subsequent adversarial training, and the semantic communication system is further fine-tuned in this way. For information transmission through the base station channel, the training target is the true label, and the parameters of the semantic communication model are updated at a lower learning rate; for transmission through the inspector channel, the training target is the preset target. In the second step, after the fine-tuning of the end-to-end semantic communication system is completed, the iterative method mentioned in the embodiment of the present invention is adopted to iteratively optimize the artificial noise.

[0114] The training of the end-to-end semantic communication system in this embodiment further improves the model's sensitivity to both legitimate and eavesdropped channels, enhancing its robustness against similar adversarial attacks. This avoids the "compromise" phenomenon that occurs when training with universal perturbations and legitimate channels, thereby maintaining privacy protection efficiency without affecting accurate decoding by the base station.

[0115] S250 , generating a precoding vector according to the channel state information, performing precoding processing on the semantic information to be transmitted using the precoding vector, and sending the precoded semantic information to be transmitted.

[0116] Among them, see Figure 6 ,The semantic communication method proposed in this invention uses convolutional layers to simulate precoding and wireless channel effects, which are defined by the precoding layer and the wireless channel layer respectively. 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 Through the wireless channel layer, we can get a 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 convolution kernels, thus reflecting the inherent complexity of the respective physical processes in semantic-aware communication systems.

[0117] S260: Use a semantic decoder to decode the pre-coded semantic information to be transmitted to obtain target information.

[0118] The receiving end receives the pre-coded semantic information to be transmitted and decodes it using a semantic decoder to obtain the target information.

[0119] As an optional but non-limiting implementation, the use of a semantic decoder to decode the pre-encoded semantic information to be transmitted to obtain target information includes but is not limited to steps D1-D2:

[0120] Step D1: The receiving end of the end-to-end semantic communication system receives the pre-coded semantic information to be transmitted.

[0121] Step D2: using a semantic decoder to decode the pre-encoded semantic information to be transmitted to obtain target information.

[0122] The target information includes the original transmission information received by the base station receiver and the original transmission information without privacy information received by the line inspector receiver. At the receiving end, the inspector uses their receiving device to capture the pre-coded semantic information to be transmitted and restores it to the original information through a semantic decoder. Due to the artificial noise added to the signal, the decoded result of the inspector's decoder does not contain the privacy information that needs to be hidden. At the same time, the drone can send the original data to the archiving department's cloud server through the base station as a backup.

[0123] An embodiment of the present invention provides a privacy protection method for drone inspections. The method comprises training an end-to-end semantic communication system for semantic communication; employing a semantic encoder to extract semantic information from information to be transmitted; creating a matrix neural network with the same output dimension as the semantic encoder, and employing the matrix neural network to generate semantic artificial noise. The matrix neural network acts as a carrier for the semantic artificial noise, and the semantic artificial noise is generated from disguised information designed to mislead eavesdroppers. The generated semantic artificial noise is superimposed on the extracted semantic information to obtain the semantic information to be transmitted; a precoding vector is generated based on channel state information, and the precoding vector is used to precode the semantic information to be transmitted, and the precoded semantic information to be transmitted is transmitted; and a semantic decoder is used to decode the precoded semantic information to obtain the target information. Using the technical solution of the embodiment of the present invention, during drone inspections of power grids, the channel differences between different users are exploited through physical layer security to achieve "one transmission, differentiated communication," allowing different users to decode different information. Defensive adversarial attack techniques are employed at the drone to generate semantic-level artificial noise, which is then incorporated into the extracted semantic information. The semantic information doped with the artificial noise is precoded as the transmitted information and transmitted over a wireless channel. By exploiting channel variations, artificial noise can hide private information from the inspector’s receiver while ensuring the integrity of the information received by the archiving department’s server.

[0124] Example 3

[0125] Figure 7 This is a schematic diagram of the structure of a drone inspection privacy protection device provided in an embodiment of the present invention. The technical solution of this embodiment can be applied to the situation of drone inspection privacy protection. The device can be implemented by software and / or hardware and is generally integrated into any electronic device with network communication function, including but not limited to: servers, computers, personal digital assistants and other devices. Figure 7 As shown, the drone inspection privacy protection device provided in this embodiment may include: a semantic communication system construction module 710, an artificial noise adding module 720, a semantic information encoding module 730 and a semantic information decoding module 740; wherein,

[0126] A semantic communication system building module 710 is used to train an end-to-end semantic communication system for semantic communication; the end-to-end semantic communication system includes a semantic encoder and a semantic decoder;

[0127] An artificial noise adding module 720 is configured to extract semantic information from the information to be transmitted using a semantic encoder, and add the generated semantic artificial noise to the semantic information to obtain the semantic information to be transmitted;

[0128] The semantic information encoding module 730 is configured to generate a precoding vector based on the channel state information, perform precoding processing on the semantic information to be transmitted using the precoding vector, and send the precoded semantic information to be transmitted;

[0129] The semantic information decoding module 740 is configured to decode the pre-encoded semantic information to be transmitted using a semantic decoder to obtain target information.

[0130] Based on the above embodiment, optionally, the semantic communication system construction module is specifically used to:

[0131] Estimating the wireless channel between the drone and the receiving end to determine the channel state information; the receiving end includes a base station receiving end and a line patrol receiving end;

[0132] The semantic encoder and the semantic decoder are jointly trained according to the channel state information to construct an end-to-end semantic communication system for semantic communication.

[0133] Based on the above embodiment, optionally, the artificial noise adding module is specifically configured to:

[0134] Use semantic encoder to extract semantic information from the information to be transmitted;

[0135] Creating a matrix neural network with the same dimension as the output of the semantic encoder, and using the matrix neural network to generate semantic artificial noise; wherein the matrix neural network serves as a carrier of the semantic artificial noise, and the semantic artificial noise is generated by disguised information;

[0136] The generated semantic artificial noise is superimposed on the extracted semantic information to obtain the semantic information to be transmitted.

[0137] Based on the above embodiment, optionally, the artificial noise adding module is further configured to:

[0138] Initializing artificial noise to a matrix with the same dimension as the output of the semantic encoder; the artificial noise is implemented in the form of universal adversarial perturbation;

[0139] The initialized artificial noise is iteratively optimized in a matrix neural network to obtain semantic artificial noise.

[0140] Based on the above embodiment, optionally, the artificial noise adding module is further configured to:

[0141] The initialized artificial noise is iteratively optimized in the matrix neural network, which can be expressed as:

[0142] δ (0) =0

[0143]

[0144] Among them, δ (0) =0 means that the artificial noise after initialization is initialized to 0; δ n Represents the artificial noise matrix after the nth iteration; function Clip ∈ {·} is used to ensure that the proposed constraint condition can be satisfied, which is to limit the value of the artificial noise δ to [-∈,∈], where ∈ is the disturbance size; y target The label or output result of the prediction result; represents the perturbation increment after the nth iteration; where α is the step size parameter used to control the perturbation update amplitude during the iteration process; sgn is the sign function used to return the sign of each element; represents the gradient of the loss function J with respect to the information to be transmitted x; f d (p(x)) represents the result predicted by the model; p(x) represents the semantic information to be transmitted.

[0145] Based on the above embodiment, optionally, after obtaining the semantic information to be transmitted, the apparatus further includes a fine-tuning optimization module, specifically configured to:

[0146] The semantic encoder and semantic decoder are jointly trained on the UAV’s lineman channel and base station channel to obtain a pre-trained semantic communication system.

[0147] Add semantic artificial noise to the semantic encoder for adversarial training, fine-tune the pre-trained semantic communication system, and obtain the target semantic communication system;

[0148] After obtaining the target semantic communication system, the semantic artificial noise is iteratively optimized in an iterative manner to obtain the target semantic artificial noise.

[0149] Based on the above embodiment, optionally, the semantic information decoding module is specifically configured to:

[0150] The receiving end of the end-to-end semantic communication system receives the pre-coded semantic information to be transmitted;

[0151] A semantic decoder is used to decode the pre-coded semantic information to be transmitted to obtain target information; wherein the target information includes the original transmission information obtained by the base station receiving end and the original transmission information without privacy information obtained by the line patrol receiving end.

[0152] The drone inspection privacy protection device provided in the embodiment of the present invention can execute the drone inspection privacy protection method provided in any embodiment of the present invention mentioned above, and has the corresponding functions and beneficial effects of executing the drone inspection privacy protection method. For detailed process, please refer to the relevant operations of the drone inspection privacy protection method in the above embodiment.

[0153] Example 4

[0154] Figure 8 1 is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. The electronic device 10 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices (such as helmets, glasses, watches, etc.) and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present invention described and / or required herein.

[0155] like Figure 8 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12, a random access memory (RAM) 13, etc., which is communicatively connected to the at least one processor 11. The memory stores a computer program that can be executed by the at least one processor. The processor 11 can perform various appropriate actions and processes according to the computer program stored in the read-only memory (ROM) 12 or the computer program loaded from the storage unit 18 into the random access memory (RAM) 13. Various programs and data required for the operation of the electronic device 10 can also be stored in the RAM 13. The processor 11, ROM 12, and RAM 13 are connected to each other via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0156] Multiple components in the electronic device 10 are connected to the I / O interface 15, including an input unit 16, such as a keyboard, a mouse, etc.; an output unit 17, such as various types of displays, speakers, etc.; a storage unit 18, such as a magnetic disk, an optical disk, etc.; and a communication unit 19, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 19 allows the electronic device 10 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0157] Processor 11 can be any general-purpose and / or specialized processing component with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various specialized artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any other suitable processor, controller, microcontroller, etc. Processor 11 executes the various methods and processes described above, such as the drone inspection privacy protection method.

[0158] In some embodiments, the drone inspection privacy protection method can be implemented as a computer program, which is tangibly contained in a computer-readable storage medium, such as the storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 10 via the ROM 12 and / or the communication unit 19. When the computer program is loaded into the RAM 13 and executed by the processor 11, one or more steps of the drone inspection privacy protection method described above can be performed. Alternatively, in other embodiments, the processor 11 can be configured to execute the drone inspection privacy protection method in any other appropriate manner (for example, by means of firmware).

[0159] Various embodiments of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), system-on-chip systems (SOCs), programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include being implemented in one or more computer programs that are executable and / or interpreted on a programmable system that includes at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0160] Computer programs for implementing the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when the computer program is executed by the processor, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The computer program may be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0161] In the context of the present invention, computer-readable storage medium can be a tangible medium that can contain or store a computer program for use by an instruction execution system, device or equipment or used in combination with an instruction execution system, device or equipment. Computer-readable storage medium can include but is not limited to electronic, magnetic, optical, electromagnetic, infrared or semiconductor systems, devices or equipment, or any suitable combination of the foregoing. Alternatively, computer-readable storage medium can be a machine-readable signal medium. A more specific example of a machine-readable storage medium can include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0162] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0163] The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), a blockchain network, and the Internet.

[0164] A computing system may include clients and servers. The clients and servers are typically remote from each other and typically interact via a communication network. This client-server relationship arises through computer programs running on the respective computers, creating a client-server relationship. The server may be a cloud server, also known as a cloud computing server or cloud host. This server is a hosting product within a cloud computing service ecosystem that addresses the management difficulties and limited scalability of traditional physical hosting and VPS services.

[0165] Example 5

[0166] An embodiment of the present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements the drone inspection privacy protection method provided in any embodiment of the present application.

[0167] The computer program product may be implemented by writing computer program code for performing the operations of the present invention in one or more programming languages, or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0168] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in the present invention can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solution of the present invention can be achieved. This is not limited herein.

[0169] The above specific embodiments do not limit the scope of protection of the present invention. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.

Claims

1. A privacy protection method for drone inspection, characterized in that: The method comprises: Training an end-to-end semantic communication system for semantic communication; the end-to-end semantic communication system includes a semantic encoder and a semantic decoder; A semantic encoder is used to extract semantic information from the information to be transmitted, and the generated semantic artificial noise is added to the semantic information to obtain the semantic information to be transmitted; Generate a precoding vector according to the channel state information, use the precoding vector to precode the semantic information to be transmitted, and send the precoded semantic information to be transmitted; A semantic decoder is used to decode the pre-coded semantic information to be transmitted to obtain target information.

2. The method according to claim 1, characterized in that The end-to-end semantic communication system trained for semantic communication comprises: Perform channel estimation on the wireless channel between the drone and the receiving end to determine channel state information; the receiving end includes a base station receiving end and a line patrol receiving end; The semantic encoder and the semantic decoder are jointly trained according to the channel state information to construct an end-to-end semantic communication system for semantic communication.

3. The method according to claim 1, characterized in that The method of extracting semantic information from the information to be transmitted by using a semantic encoder and adding the generated semantic artificial noise to the semantic information to obtain the semantic information to be transmitted includes: A semantic encoder is used to extract semantic information from the information to be transmitted; A matrix neural network with the same output dimension as the semantic encoder is created, and the matrix neural network is used to generate semantic artificial noise; wherein the matrix neural network serves as a carrier of the semantic artificial noise, and the semantic artificial noise is generated by camouflage information. The generated semantic artificial noise is added to the extracted semantic information to obtain the semantic information to be transmitted.

4. The method according to claim 3, characterized in that The method of using the matrix neural network to generate semantic artificial noise includes: Initializing the artificial noise to a matrix with the same dimension as the output of the semantic encoder; the artificial noise is implemented in the form of universal adversarial perturbation; The initialized artificial noise is iteratively optimized in the matrix neural network to obtain semantic artificial noise.

5. The method according to claim 4, characterized in that The initialized artificial noise is iteratively optimized in a matrix neural network to obtain semantic artificial noise, including: The initialized artificial noise is iteratively optimized in the matrix neural network, which can be expressed as: d (0) =0 Among them, δ (0) =0 means that the initialized artificial noise is initialized to 0; δ n Represents the artificial noise matrix after the nth iteration; function Clip ∈ {·} is used to ensure that the proposed constraint condition can be satisfied, which is to limit the value of the artificial noise δ to [-∈, ∈], where ∈ is the disturbance size; y target The label or output result representing the prediction result; represents the perturbation increment after the nth iteration; where α is the step size parameter, which is used to control the perturbation update amplitude during the iteration process; sgn is the sign function, which is used to return the sign of each element; represents the gradient of the loss function J with respect to the information to be transmitted x; f d (p(x)) represents the result predicted by the model; p(x) represents the semantic information to be transmitted.

6. The method according to claim 1, characterized in that After obtaining the semantic information to be transmitted, the method further includes: The semantic encoder and the semantic decoder are jointly trained on the lineman channel and the base station channel of the UAV to obtain a pre-trained semantic communication system; Add semantic artificial noise to the semantic encoder for adversarial training, fine-tune the pre-trained semantic communication system, and obtain the target semantic communication system; After obtaining the target semantic communication system, the semantic artificial noise is iteratively optimized in an iterative manner to obtain the target semantic artificial noise.

7. The method according to claim 1, characterized in that The using a semantic decoder to decode the pre-coded semantic information to be transmitted to obtain target information includes: The receiving end of the end-to-end semantic communication system receives the pre-coded semantic information to be transmitted; A semantic decoder is used to decode the pre-coded semantic information to be transmitted to obtain target information; wherein the target information includes the original transmission information obtained by the base station receiving end and the original transmission information without privacy information obtained by the line patrol receiving end.

8. A privacy protection device for drone inspection, characterized in that: The device comprises: A semantic communication system building module, used for training an end-to-end semantic communication system for semantic communication; the end-to-end semantic communication system includes a semantic encoder and a semantic decoder; An artificial noise adding module is used to extract semantic information from the information to be transmitted by using a semantic encoder, and add the generated semantic artificial noise to the semantic information to obtain the semantic information to be transmitted; A semantic information encoding module, used to generate a precoding vector according to the channel state information, and use the precoding vector to precode the semantic information to be transmitted, and send the precoded semantic information to be transmitted; The semantic information decoding module is used to use a semantic decoder to decode the pre-coded semantic information to be transmitted to obtain target information.

9. The device according to claim 8, characterized in that The semantic communication system building module is specifically used for: Perform channel estimation on the wireless channel between the drone and the receiving end to determine channel state information; the receiving end includes a base station receiving end and a line patrol receiving end; The semantic encoder and the semantic decoder are jointly trained according to the channel state information to construct an end-to-end semantic communication system for semantic communication.

10. The device according to claim 8, characterized in that The artificial noise adding module is specifically used for: A semantic encoder is used to extract semantic information from the information to be transmitted; A matrix neural network with the same output dimension as the semantic encoder is created, and the matrix neural network is used to generate semantic artificial noise; wherein the matrix neural network serves as a carrier of the semantic artificial noise, and the semantic artificial noise is generated by camouflage information. The generated semantic artificial noise is added to the extracted semantic information to obtain the semantic information to be transmitted.

11. The device according to claim 10, characterized in that The artificial noise adding module is further specifically used for: Initializing the artificial noise to a matrix with the same dimension as the output of the semantic encoder; the artificial noise is implemented in the form of universal adversarial perturbation; The initialized artificial noise is iteratively optimized in the matrix neural network to obtain semantic artificial noise.

12. The device according to claim 11, characterized in that The artificial noise adding module is further specifically used for: The initialized artificial noise is iteratively optimized in the matrix neural network, which can be expressed as: d (0) =0 Among them, δ (0) =0 means that the initialized artificial noise is initialized to 0; δ n Represents the artificial noise matrix after the nth iteration; function Clip ∈ {·} is used to ensure that the proposed constraint condition can be satisfied, which is to limit the value of the artificial noise δ to [-∈, ∈], where ∈ is the disturbance size; y target The label or output result representing the prediction result; represents the perturbation increment after the nth iteration; where α is the step size parameter, which is used to control the perturbation update amplitude during the iteration process; sgn is the sign function, which is used to return the sign of each element; represents the gradient of the loss function J with respect to the information to be transmitted x; f d (p(x)) represents the result predicted by the model; p(x) represents the semantic information to be transmitted.

13. The device according to claim 8, characterized in that After obtaining the semantic information to be transmitted, the device further includes a fine-tuning optimization module, which is specifically used to: The semantic encoder and the semantic decoder are jointly trained on the lineman channel and the base station channel of the UAV to obtain a pre-trained semantic communication system; Add semantic artificial noise to the semantic encoder for adversarial training, fine-tune the pre-trained semantic communication system, and obtain the target semantic communication system; After obtaining the target semantic communication system, the semantic artificial noise is iteratively optimized in an iterative manner to obtain the target semantic artificial noise.

14. The device according to claim 8, characterized in that The semantic information decoding module is specifically used for: The receiving end of the end-to-end semantic communication system receives the pre-coded semantic information to be transmitted; A semantic decoder is used to decode the pre-coded semantic information to be transmitted to obtain target information; wherein the target information includes the original transmission information obtained by the base station receiving end and the original transmission information without privacy information obtained by the line patrol receiving end.

15. An electronic device, characterized in that: include: one or more processors; A storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the drone inspection privacy protection method described in any one of claims 1-7.

16. A storage medium containing computer executable instructions, characterized in that: The computer executable instructions, when executed by a computer processor, are used to execute the drone inspection privacy protection method as described in any one of claims 1-7.

Images