Hash positioning encryption method and device for remote sensing marine target in troposphere waveguide environment and storage medium

Abstract

The invention discloses a remote sensing maritime target Hash positioning encryption method in a troposphere waveguide environment. The method comprises the following steps: step 1, deep Hash network training; step 2, class label encryptor training; and step 3, encrypted image generation antagonism training. According to the encryption method disclosed by the invention, on the basis of a marine remote sensing image deep Hash feature rapid positioning algorithm, a trained original deep Hash network is taken as an original model, Hash encryption which cannot be perceived visually is taken as a target, and encryption feature implantation of an image is realized by training an encrypted image generator with good performance; and the judgment of the deep Hash network on the image is influenced. And the defect of the existing target positioning image encryption method on the ocean remote sensing image encryption effect is overcome.

Description

Technical Field

[0001] The present invention belongs to the field of remote sensing image processing of marine targets, and in particular relates to a visually invisible encryption method, device and storage medium for remote sensing marine target positioning in a tropospheric waveguide environment based on deep hashing in this field. Background Art

[0002] With the rapid development of the marine economy and the increasing importance of marine security, target monitoring and positioning in the marine tropospheric duct environment has become a key technology for marine management and defense. Traditional target positioning methods in the tropospheric duct environment mainly rely on remote sensing technologies such as radar and satellites. However, in complex sea conditions and long-distance monitoring scenarios (such as the tropospheric duct environment), they suffer from slow data processing, insufficient positioning efficiency, and low data processing security. In addition, target positioning data in the tropospheric duct environment often involves sensitive information, such as ship locations and marine resource development areas, which places higher demands on data security.

[0003] Traditional target positioning in the marine tropospheric atmospheric duct environment is unable to achieve effective target positioning due to the high complexity of image processing algorithms and difficulties in model transplantation. Feature hashing, as an algorithm that maps high-dimensional features to low-dimensional binary features, reflects the similarity between images by calculating the Hamming distance of binary feature vectors, which can significantly reduce computing and storage costs. Therefore, it has been effectively applied to marine remote sensing image retrieval and positioning tasks under complex sea conditions (such as tropospheric duct environments). In particular, deep hashing methods achieve superior performance than shallow hashing by learning nonlinear hash functions in an end-to-end manner. At present, positioning methods based on deep hashing have been deeply studied in the field of remote sensing image processing in tropospheric duct environments.

[0004] After years of development, information security technology has expanded from cryptography to the field of hidden technologies. Currently, achieving visually imperceptible image encryption has become a hot topic and a cutting-edge research topic. Entering the 21st century, an increasing amount of high-quality domestic remote sensing data is being widely used. However, the security of information transmission often determines the quality of research results. Therefore, research on the security of information transmission is crucial.

[0005] Although the hash positioning method provides a lot of convenience for the positioning of specific objects in large-scale scenes, sensitive targets need to be protected in the actual marine tropospheric waveguide environment. Under the premise of not affecting image observation, it is not desirable to quickly complete the information extraction and positioning of specific sensitive targets through hash positioning technology. Therefore, it is necessary to encrypt large-scale tropospheric waveguide remote sensing images. This task is called the hash positioning encryption task. Since the hash positioning method mainly converts tropospheric waveguide remote sensing images into binary hash codes for image matching, an encryption algorithm is needed. Under the premise that there is not much difference in the visual appearance of the remote sensing images, the hash code generated by the deep hash network of the tropospheric waveguide remote sensing images is significantly changed, thereby completing the encryption of block images in specific areas and protecting specific targets.

[0006] While traditional tropospheric waveguide image encryption algorithms can effectively reduce the retrieval accuracy of deep hash networks, the encrypted images generated by these algorithms exhibit significant visual changes, which contradicts the original goal of achieving encryption with minimal loss of detail. Currently, many researchers in the image retrieval field have conducted preliminary research on modifying the hash codes generated by deep hash networks, while ensuring minimal loss of image detail. However, these studies have been experimentally tested on general datasets, resulting in limited algorithm specificity and potential for improvement in encryption effectiveness. Furthermore, hash positioning encryption requires processing tropospheric waveguide remote sensing target images. The difficulty in resolving, large scale, and complex semantics of remote sensing images further complicates the task of visually invisible image encryption. Therefore, it is necessary to develop a more targeted and effective algorithm to accomplish visually invisible hash encryption of remote sensing images, thereby providing technical support for the secure transmission and positioning of large-scale remote sensing maritime imagery. Summary of the Invention

[0007] This paper addresses the significant visual loss and poor encryption effectiveness of existing image encryption methods for target hashing of ocean remote sensing images in tropospheric waveguide environments, hindering the secure development of ocean remote sensing image positioning. This paper provides a method, device, and storage medium for ocean remote sensing image positioning encryption based on deep hashing in tropospheric waveguide environments. Combined with adversarial sample generation technology, this paper provides a visually invisible hash image encryption algorithm, which can provide security and technical support for the field of ocean remote sensing data processing.

[0008] The present invention adopts the following technical solutions:

[0009] A method for remote sensing maritime target hash positioning encryption in a tropospheric waveguide environment, the improvement of which is that it includes the following steps:

[0010] Step 1, deep hashing network training:

[0011] The algorithm involves feature extraction and hash code generation. In the feature extraction phase, a convolutional neural network is used to extract continuous feature values ​​from the image input to the model. The fully connected layers in the convolutional neural network are replaced with hash layers to generate image hash codes. The weight parameters in the convolutional neural network are trained using given image labels. In the hash code generation phase, an activation function converts continuous feature values ​​into binary values, which are then constrained by a loss function to enable the network to generate hash codes.

[0012] Step 2, class label encryptor training:

[0013] Training relies on the calculation of the class encryption loss function, which consists of the quantization loss of the label encryption features, the pairwise similarity loss of the image sample hash codes, and the classification loss of the predicted labels, as well as the iterative implementation of network parameter optimization. The trained class label encryptor will be encapsulated for subsequent extraction of class label semantic information and generation of encrypted images.

[0014] Step 3: Generate adversarial training using encrypted images:

[0015] First, given a new category label, the semantic information in the label is extracted by the class label encryptor to supervise the generation of encrypted examples in the encrypted image generator;

[0016] Afterwards, the encrypted image generator fuses the given original image and the semantic information extracted by the class label encryptor to form an encrypted image;

[0017] After the encrypted image is generated, the encrypted image is input into the hash network to generate the hash code of the encrypted image through feature extraction and hash layer, and the encrypted image is optimized by iteratively minimizing the Hamming distance loss between the encrypted image hash code and the target label hash code;

[0018] Finally, the generated encrypted image is input into the discriminator together with the original image for identification, and the identification result is fed back to the encrypted image generator, encouraging the encrypted image generator to generate more realistic encrypted images while ensuring the semantic information of the new category.

[0019] Furthermore, the training process of the class label encryptor is:

[0020] Given an encrypted label representing the nth class, the class label encryptor takes the encrypted label as input and obtains the encrypted features of the nth class;

[0021] The encrypted features are fed into a fully connected layer and then converted into hash codes through a tanh layer. Quantization loss function L qua for:

[0022]

[0023] In the above formula, is the class center hash code corresponding to the nth class, and N is the total number of data sample labels;

[0024] Input image x (m) The label corresponding to the nth class Class consistency L mn Defined as:

[0025]

[0026] In the above formula, y (m) Represents the true category label of the mth sample;

[0027] Input image x (m) The corresponding real hash code b (m) and Hash code consistency S between mn for:

[0028]

[0029] Through formulas (2) and (3), the pairwise similarity loss function L sim Defined as:

[0030]

[0031] In the above formula, M is the total number of data samples;

[0032] Input the encrypted features into a fully connected layer and output a predicted label Its classification loss function L cla for:

[0033]

[0034] The overall loss function of the class label encryptor is defined as follows through equations (1), (4), and (5):

[0035] L C =α1L qua +α2L sim +α3L cla (6)

[0036] In the above formula, α1, α2 and α3 are all hyperparameters, and the trained class label encryptor is encapsulated and directly used to train the encrypted image generator.

[0037] Furthermore, the reconstruction loss L of the encrypted image is rec Defined as:

[0038]

[0039] In the above formula, To encrypt the image;

[0040] Encrypt the image hash code The Hamming loss L between the hash code of the original image ham Defined as:

[0041]

[0042] In the above formula, K is the length of the hash code;

[0043] The discriminator and the generator’s discrimination loss L disc Defined as:

[0044]

[0045] In the above formula, D represents a discriminator, To encrypt the image The corresponding extended encryption tag,

[0046] The overall loss function L of the encrypted image generator is converted into G Defined as:

[0047] L G =β1L rec +β2L ham +β3L disc (10)

[0048] In the above formula, β1, β2 and β3 are all hyperparameters.

[0049] Furthermore, the loss function L of the encrypted block identifier is D Defined as:

[0050]

[0051] In the above formula, To encrypt the image The corresponding extended encryption tag, For the input image x (m) The corresponding extended original tag,

[0052] A remote sensing maritime target hash positioning encryption device in a tropospheric waveguide environment, the improvement of which is that it includes: a processor; a memory in which executable instructions of the processor are stored; the processor is configured to execute the steps of the above method by executing the executable instructions.

[0053] A computer-readable storage medium is used to store a program, wherein the improvement lies in that the steps of the above method are implemented when the program is executed.

[0054] The beneficial effects of the present invention are:

[0055] The encryption method disclosed in this paper is based on a deep hash feature rapid positioning algorithm for marine remote sensing images. It uses a trained original deep hash network as the original model and aims to achieve visually imperceptible hash encryption. By training a high-performance encrypted image generator, the encryption characteristics of the image are embedded, thereby influencing the deep hash network's judgment of the image. This method overcomes the shortcomings of existing target positioning image encryption methods in encrypting marine remote sensing images, contributes to the development of a secure rapid target positioning algorithm based on marine remote sensing image features, and provides technical support for protecting the positioning security of specific scenes in large-scale marine remote sensing images. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 Locate the flow chart for the hash;

[0057] Figure 2 This is the hash positioning encryption flow chart of this embodiment;

[0058] Figure 3 Schematic diagram of the architecture and training strategy for encrypted image generation;

[0059] Figure 4 Schematic diagram of the training process of the class label encryptor;

[0060] Figure 5 Schematic diagram of the training process of the encrypted image generator;

[0061] Figure 6 This is a comparison chart of remote sensing retrieval and positioning results of the original image and the encrypted image;

[0062] Figure 7-1 This is a comparison chart of target accuracy curves of encrypted images using different hash encryption methods;

[0063] Figure 7-2 This is a comparison chart of PR curves of images encrypted using different hash encryption methods. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0065] Example 1. This embodiment discloses a method for positioning and encryption of marine remote sensing images based on deep hashing in a tropospheric waveguide environment, in order to complete the visually invisible encryption task of specific scenes in large-scale marine remote sensing images. It aims at visually imperceptible hash encryption, and implements the implantation of encryption characteristics of the image by training a good-performance encrypted image generator, thereby affecting the judgment of the deep hash network on the image. This embodiment takes the encryption task of specific scenes in large-scale tropospheric waveguide remote sensing images under complex sea conditions as the main line, and provides a method for hash positioning and encryption of remote sensing marine targets in a tropospheric waveguide environment, which achieves the purpose of protecting the security of remote sensing data by generating an encrypted image that is visually indistinguishable from the original image. In addition, this embodiment designs a flexible framework for hash positioning encryption in combination with the encrypted image generator, so that the hash positioning model has hash loss and hash dislocation effects, thereby providing technical guarantees for protecting the positioning security of specific scenes in large-scale marine remote sensing images. Specifically comprising the following steps:

[0066] Step 1, deep hashing network training:

[0067] The deep hash learning process can be divided into two steps: feature extraction and hash code generation. In the feature extraction stage, a well-trained convolutional neural network extracts continuous feature values ​​from the image input to the model. To maximize the effect of image hash code generation, the deep hash network uses the convolutional neural network as the feature extraction layer and replaces the fully connected layer in the convolutional neural network with a hash layer for image hash code generation. The weight parameters in the convolutional neural network are trained based on the given image labels. Ultimately, the loss of hash code generation between the image and the label is minimized while ensuring efficiency, ensuring that the deep hash network completes the hash code conversion of the image in a short time and has sufficient semantic representativeness. In the hash code generation stage, the activation function converts the continuous feature values ​​into binary values, and then the binary values ​​are constrained by the loss function to enable the network to generate efficient hash codes.

[0068] Training an encrypted image generator for a deep hashing network. The encrypted image generator takes an encrypted label and an original image as input. It encrypts the image by making subtle changes to the image and generates the corresponding encrypted image. The encrypted image generated by the encrypted image generator is visually indistinguishable from the original image, but contains semantic information of the new class. This causes the deep hashing network to favor the hash code of the new class when performing hash code conversion. As a result, the image cannot be detected by the hashing localization algorithm during large-scale remote sensing image localization, achieving both encryption and protection. To achieve this, the encrypted image generator should be trained according to the following two criteria: First, the encrypted image should be as visually similar as possible to its original block image. Second, the hash code of the encrypted image should be close to the class indicated by the encrypted label, not the original class. To achieve this, not only an encrypted image generator but also a discriminator are required to undergo iterative adversarial training with it. Furthermore, a key component of the encrypted image generator is a class label encryptor, which is trained independently of the encrypted image generator. The following briefly describes the adversarial training process for the class label encryptor, the encrypted image generator, and the discriminator.

[0069] Step 2, class label encryptor training:

[0070] The class label encryptor takes the encrypted label as input and outputs the encrypted features. Its purpose is to ensure that the encrypted image generated by the subsequent generator can have the semantic information of the encrypted label class, thereby inducing the deep hashing network to generate a hash code belonging to the class to which the encrypted label belongs. Its training mainly relies on the calculation of the class encryption loss function composed of three losses (quantization loss of the label encryption feature, pairwise similarity loss of the image sample hash code, and classification loss of the predicted label) and iterative implementation of network parameter optimization. The trained class label encryptor will be encapsulated for subsequent extraction of class label semantic information and generation of encrypted images.

[0071] Step 3: Generate adversarial training using encrypted images:

[0072] First, given a new category label, the semantic information in the label is extracted by the class label encryptor to supervise the generation of encrypted examples in the encrypted image generator;

[0073] Afterwards, the given original image and the semantic information extracted by the class label encryptor are fused together through the encrypted image generator to form an encrypted image; while the encrypted image is visually indistinguishable from the original image, the semantic information is more biased towards the semantic information of the new category, thereby making the hash code of the encrypted image more biased towards the new category and away from the original category.

[0074] After the encrypted image is generated, the encrypted image is input into the hash network to generate the hash code of the encrypted image through feature extraction and hash layer. The encrypted image is optimized by iteratively minimizing the Hamming distance loss between the encrypted image hash code and the target label hash code, making the hash code generated by the encrypted image more biased towards the new category.

[0075] Finally, the generated encrypted image is input into the discriminator together with the original image for identification, and the identification result is fed back to the encrypted image generator, encouraging the encrypted image generator to generate more realistic encrypted images while ensuring the semantic information of the new category.

[0076] To improve model training efficiency, experiments were conducted on an RTX2080ti graphics processing unit (GPU) running the PyTorch framework. The learning rate was set to 0.05, the batch size to 32, and the number of epochs to 100. ResNet50, pre-trained on ImageNet, was used as the backbone for the deep learning feature extraction method.

[0077] Since the work done in this embodiment is based on the application of hashing methods to marine remote sensing images, an indicator widely used to evaluate the image matching performance of hashing methods, namely the target average precision (MAP), is used to evaluate the remote sensing image retrieval and positioning effects of the deep hashing network before and after image encryption. The experimental results show that the retrieval results of the deep hashing network for encrypted images are obviously biased towards other categories, that is, the categories to which the encrypted labels belong, proving that the encrypted images generated by this embodiment do have an effective hashing image encryption effect under the premise that the visual changes are small enough. In addition, further experiments verified that the encryption effect of the encrypted images generated by this embodiment is better than the existing mainstream hashing encryption methods, and can provide effective technical guarantees for protecting the positioning security of specific scenes in large-scale marine remote sensing images.

[0078] Hash localization is a remote sensing image processing technology that can quickly locate specific scenes in large-scale remote sensing images. Its efficiency depends on a well-performing deep hash network that can generate different hash codes for input images of different categories. The process of hash localization is as follows: Figure 1 Given a block-based large-scale remote sensing image and an example image of a specific category (such as an airplane or oil drum), hash localization aims to use a trained deep hashing network to identify localized block images within the large-scale remote sensing image that share the same category as the example image. By calculating and sorting the similarity of hash codes, it is possible to quickly extract images of specific scenes within the large-scale remote sensing image, thereby achieving effective localization.

[0079] The hash positioning encryption task aims to complete visually imperceptible hash image encryption to protect specific remote sensing scenes from being hashed. The process of hash positioning encryption is as follows: Figure 2 As shown. To achieve this goal, this embodiment uses a convolutional generative adversarial network to design an encrypted image generator that can add specific perturbations to images. Through training, the encrypted image generator can add specific noise perturbations to the original sample image without visually changing the image, thereby affecting the hash code generation of the image by the deep hash network, and ultimately causing the hash positioning method to produce invalid or even erroneous positioning results, thereby achieving a visually imperceptible hash positioning encryption task and protecting the security of remote sensing data. The encrypted image generator takes an encrypted label and an original image as input. It encrypts the image by making slight changes to the image and generates a corresponding encrypted image. The encrypted image generated by the encrypted image generator is visually indistinguishable from the original image, but has semantic information of a new category, which makes the deep hash network more inclined to the hash code of the new category when performing hash code conversion. Therefore, the image will not be detected by the hash positioning algorithm during large-scale remote sensing image positioning, achieving the effect of encryption and protection.

[0080] In order to obtain a sufficiently effective encrypted image generator, this embodiment trains it in a generative adversarial manner. To this end, not only an encrypted image generator is required, but also a discriminator. In addition, an important part of the encrypted image generator is a class label encryptor, which is trained independently of the encrypted image generator. The architecture and training strategy of the class label encryptor, encrypted image generator and discriminator are as follows: Figure 3 As shown in Figure 1, the training data consists of a set of images, their class labels, hash codes, and the class center hash code for each class. The class center hash codes are generated during the training of the deep hashing network and are encapsulated with the trained deep hashing network, making them immutable in practice. Furthermore, the training of the encrypted image generator consists of three iterative optimization steps: the class label encryptor, the encrypted image generator, and the discriminator.

[0081] The following gives the specific steps and principle formulas for training the class label encryptor and encrypted image generator. (1) ,x (2) ,...,x (m) ,...,x (M)} to define the input image dataset, using Y={y (1) ,y (2) ,...,y (m) ,...,y (M)} to define the true category label of the corresponding image in the dataset, and use B = {b (1) ,b (2) ,...,b (m) ,...,b (M)} to define the real hash codes of the corresponding images in the dataset, which are generated by the encapsulated deep hash network. In addition, this embodiment uses To define n labels corresponding to n categories of image samples, and use To define the n class center hash codes corresponding to n categories. It should be noted that the subscript (m) with brackets represents the mth sample in the dataset, and the subscript n without brackets represents the nth class.

[0082] The class label encryptor takes the encrypted label as input and outputs the encrypted feature. Its purpose is to make the encrypted image generated by the subsequent generator have the semantic information of the encrypted label class, thereby inducing the deep hash network to generate a hash code belonging to the class to which the encrypted label belongs. To achieve the above purpose, the training process of the class label encryptor designed in this embodiment is as follows: Figure 4 As shown in the figure, the program first initializes the network structure parameters of the class label encryptor and then checks the number of iterations. If the number of iterations is reached, the model is saved and the program exits. If the number of iterations is not reached, the class label encryptor is updated based on the loss function. After the parameters are updated, the program inputs the encrypted label into the updated class label encryptor to generate the corresponding predicted hash code and predicted category. The class label encryptor is then trained by calculating the class encryption loss function and performing backpropagation.

[0083] Specifically, given an encrypted label representing the nth class, the class label encryptor takes the encrypted label as input and obtains the encrypted features of the nth class;

[0084] The encrypted features are fed into a fully connected layer and then converted into hash codes through a tanh layer. Quantization loss function L qua for:

[0085]

[0086] In the above formula, is the class center hash code corresponding to the nth class, and N is the total number of data sample labels;

[0087] Input image x (m) The label corresponding to the nth class Class consistency L mn Defined as:

[0088]

[0089] In the above formula, y (m) Represents the true category label of the mth sample;

[0090] Input image x (m) The corresponding real hash code b(m) The hash code of the corresponding label generated Hash code consistency S between mn for:

[0091]

[0092] Through formulas (2) and (3), the pairwise similarity loss function L sim Defined as:

[0093]

[0094] In the above formula, M is the total number of data samples;

[0095] Input the encrypted features into a fully connected layer and output a predicted label Its classification loss function L cla for:

[0096]

[0097] The overall loss function of the class label encryptor is defined as follows through equations (1), (4), and (5):

[0098] L C =α1L qua +α2L sim +α3L cla (6)

[0099] In the above formula, α1, α2 and α3 are all hyperparameters, and the trained class label encryptor is encapsulated and directly used to train the encrypted image generator.

[0100] Given an original block image and an encrypted label that is different from the original block category, this embodiment develops an encrypted image generator that can use the encrypted label in combination with the original image to generate a corresponding encrypted image. The hash code of the encrypted image generated by the deep hash network will be far away from the class center hash code of the original category and close to the class center hash code of the encrypted category. Figure 5 This paper demonstrates the architecture and training process of an encrypted image generator. First, the program initializes the network parameters of the generator and discriminator. It then checks the number of iterations. If the number of iterations has been reached, the model is saved and the program exits. If the number of iterations has not been reached, the generator and discriminator parameters are updated based on the loss function. After the generator parameters are updated, the original image and encrypted label are input into the updated image generator to generate the corresponding encrypted image sample. The performance of the generator and discriminator is trained by calculating the loss functions of the generator and discriminator and performing backpropagation.

[0101] Specifically, the encrypted tag First, it is converted into encrypted features through the class label encryptor, and then input into the input layer of the generator together with the original image. After processing through the convolution layer, deconvolution layer and output layer, the final output is the corresponding encrypted image. To keep the encrypted image The visual quality of the encrypted image is reconstructed with loss L rec Defined as:

[0102]

[0103] In the above formula, To encrypt the image;

[0104] The reconstruction loss mainly ensures that there is basically no visual difference between the encrypted image and the original image. The encrypted image generated later Converted into the hash code corresponding to the encrypted image through the hash network To implement encryption at the hash code level, the encrypted image hash code The Hamming loss L between the hash code of the original image ham Defined as:

[0105]

[0106] In the above formula, K is the length of the hash code;

[0107] Hamming loss forces the hash code of the encrypted image to Not close to the class center hash code corresponding to the original patch image, but close to the encrypted label The corresponding class center hash code. In order to further enhance The correlation between the nth class and the discriminator and the generator’s identification loss L disc Defined as:

[0108]

[0109] In the above formula, D represents a discriminator, which mainly supervises the training of the encrypted image generator through an extended encrypted label, which will be described later. By optimizing the identification loss, the encrypted image is encouraged to It is more representative than the patch images in the nth class. To encrypt the image The corresponding extended encryption tag,

[0110] The overall loss function L of the encrypted image generator is converted into G Defined as:

[0111] L G =β1L rec +β2Lham +β3L disc (10)

[0112] In the above formula, β1, β2 and β3 are all hyperparameters.

[0113] The purpose of the discriminator is to distinguish between the encrypted image and the original image. The input of the discriminator is the original image x (m) Or encrypt the image The output is a prediction vector that indicates whether the input block image is original (i.e., 0) or encrypted (i.e., 1). This embodiment supervises the training process of the encrypted block discriminator through the extended encrypted label and the extended original label. The loss function L of the encrypted block discriminator is D Defined as:

[0114]

[0115] In the above formula, To encrypt the image The corresponding extended encryption tag, For the input image x (m) The corresponding extended original tag,

[0116] All models introduced in this embodiment are trained using the Adam optimization method to minimize the losses in formulas (6), (10), and (11), so that the encrypted image generator can better encrypt the original image. After the above process, the encrypted image generator and the discriminator are iteratively updated, and finally the trained encrypted image generator model can transfer the hash code of the image under the premise of sufficiently small pixel changes, thereby ensuring that it can achieve hash encryption of remote sensing images under the premise of no visual changes, thereby protecting the positioning security of specific remote sensing scenes.

[0117] To verify the effectiveness of the encryption method, this embodiment conducted experiments on the SpaceNet dataset. This dataset includes 10 classes and a total of 6,800 remote sensing images. This embodiment uses 6,300 images to train the deep hash network and uses 500 images to generate interference samples to test the effectiveness of the encryption method proposed in this embodiment. This embodiment uses ResNet50 pre-trained on ImageNet as the basic backbone of the deep learning feature extraction method. In addition, the hash code length is set to 64, the hyperparameters α1, α2, and α3 are set to 1, 10-4, and 1, respectively, and the hyperparameters β1, β2, and β3 are set to 50, 1, and 1, respectively.

[0118] In order to intuitively compare the changes in retrieval results before and after image encryption, Figure 6The results of a deep hashing network retrieval of the original image and the encrypted image are compared. The results in the figure show that while there is no visual difference between the original and encrypted images, the deep hashing network's retrieval results for the encrypted image clearly favor other categories, namely the category to which the encrypted label belongs. This verifies that the encrypted image generated using the encryption method of this embodiment does achieve effective hash image encryption while minimizing visual changes.

[0119] In order to compare the hash code transfer capabilities of different methods for encrypted images, the target accuracy curves and PR curves of encrypted images generated by different methods are plotted. Specifically, this embodiment uses ResNet50 as the network architecture, uses the encryption method proposed in this embodiment and various comparison methods (DHTA, Pros-GAN and Advhash), and generates encrypted images for three deep hash networks (DPH, DPSH, HashNet) trained for 64-bit hash codes. The encrypted image retrieval accuracy and recall rate under different return results are recorded. The results are shown as follows: Figure 7-1 、 7-2 shown.

[0120] Figure 7-1 The target accuracy curves of encrypted images generated by different methods are shown, which mainly describe the target accuracy of the results returned for different numbers of encrypted images. Figure 7-1 It can be concluded that the target accuracy of the encryption method proposed in this embodiment is higher than that of the other four compared methods, regardless of the different deep hash models or the number of returned results. Furthermore, as the number of returned results increases, the target accuracy of the encryption method proposed in this embodiment decreases more slowly than that of the other four compared methods, further demonstrating the superiority of the encryption method proposed in this embodiment.

[0121] Figure 7-2 The precision-recall (PR) curves of encrypted images generated by different methods are shown. They mainly describe the target precision corresponding to the recall rate of different return results for the encrypted image. The area enclosed under the PR curve is the average target precision of the encrypted image, that is, the encrypted image-encrypted label MAP value. Figure 7-2 As can be seen from the figure, the area under the PR curve of the encryption method proposed in this embodiment is larger than that of the other four compared methods. Furthermore, as the recall rate of returned results increases, the target precision value of the encryption method proposed in this embodiment decreases the slowest, further demonstrating the superiority of the encryption method proposed in this embodiment in hash image encryption tasks.

[0122] For three deep hash networks, the average target accuracy of the encrypted images generated by the encryption method proposed in this embodiment is significantly higher than that of several other hash image encryption attack methods, proving that the hash code transfer capability of the encryption method used in this embodiment is superior to that of several other methods. The reason for this is that methods such as DHTA and Pros-GAN select a limited number of hash codes in each target class for training, resulting in hash codes selected from different classes not being very different from each other. In contrast, the encryption method proposed in this embodiment uses a class label encryptor to generate encrypted features for each encrypted label and uses the class-centered hash code of the encrypted label for training. In this way, the encryption method used in this embodiment can generate encrypted hash codes that are more distinctive than DHTA and Pros-GAN, thereby having better hash code transfer and image encryption capabilities.

[0123] This embodiment also discloses a device for remote sensing maritime target hash positioning encryption in a tropospheric waveguide environment, comprising: a processor; a memory storing executable instructions for the processor; the processor configured to execute the steps of the above method by executing the executable instructions; and a computer-readable storage medium for storing a program that, when executed, implements the steps of the above method.

Claims

1. A remote sensing maritime target hash positioning encryption method in a tropospheric waveguide environment, characterized in that: The steps include: Step 1, deep hashing network training: The algorithm involves feature extraction and hash code generation. In the feature extraction phase, a convolutional neural network is used to extract continuous feature values ​​from the image input to the model. The fully connected layers in the convolutional neural network are replaced with hash layers to generate image hash codes. The weight parameters in the convolutional neural network are trained using given image labels. In the hash code generation phase, an activation function converts continuous feature values ​​into binary values, which are then constrained by a loss function to enable the network to generate hash codes. Step 2, class label encryptor training: The training relies on the calculation of the encryption-like loss function, which is composed of the quantization loss of the label encryption features, the pairwise similarity loss of the image sample hash codes, and the classification loss of the predicted labels, as well as the iterative implementation of network parameter optimization; The trained class label encryptor will be encapsulated for subsequent extraction of class label semantic information and generation of encrypted images; Step 3: Generate adversarial training using encrypted images: First, given a new category label, the semantic information in the label is extracted by the class label encryptor to supervise the generation of encrypted examples in the encrypted image generator; Afterwards, the encrypted image generator fuses the given original image and the semantic information extracted by the class label encryptor to form an encrypted image; After the encrypted image is generated, the encrypted image is input into the hash network to generate the hash code of the encrypted image through feature extraction and hash layer, and the encrypted image is optimized by iteratively minimizing the Hamming distance loss between the encrypted image hash code and the target label hash code; Finally, the generated encrypted image is input into the discriminator together with the original image for identification, and the identification result is fed back to the encrypted image generator, encouraging the encrypted image generator to generate more realistic encrypted images while ensuring the semantic information of the new category.

2. The method for remote sensing maritime target hash positioning encryption in a tropospheric waveguide environment according to claim 1 is characterized in that: The training process of the class label encryptor is: Given an encrypted label representing the nth class, the class label encryptor takes the encrypted label as input and obtains the encrypted features of the nth class; The encrypted features are fed into a fully connected layer and then converted into hash codes through a tanh layer. Quantization loss function L qua for: In the above formula, is the class center hash code corresponding to the nth class, and N is the total number of data sample labels; Input image x (m) The label corresponding to the nth class Class consistency L mn Defined as: In the above formula, y (m) Represents the true category label of the mth sample; Input image x (m) The corresponding real hash code b (m) and Hash code consistency S between mn for: Through formulas (2) and (3), the pairwise similarity loss function L sim Defined as: In the above formula, M is the total number of data samples; Input the encrypted features into a fully connected layer and output a predicted label Its classification loss function L cla for: The overall loss function of the class label encryptor is defined as follows through equations (1), (4), and (5): L C =α1L qua +α2L sim +α3L cla (6) In the above formula, α1, α2 and α3 are all hyperparameters, and the trained class label encryptor is encapsulated and directly used to train the encrypted image generator.

3. The method for remote sensing maritime target hash positioning encryption in a tropospheric waveguide environment according to claim 2 is characterized in that: The reconstruction loss L of the encrypted image rec Defined as: In the above formula, To encrypt the image; Encrypt the image hash code The Hamming loss L between the hash code of the original image ham Defined as: In the above formula, K is the length of the hash code; The discriminator and the generator’s discrimination loss L disc Defined as: In the above formula, D represents a discriminator, To encrypt the image The corresponding extended encryption tag, The overall loss function L of the encrypted image generator is converted into G Defined as: L G =β1L rec +β2L ham +β3L disc (10) In the above formula, β1, β2 and β3 are all hyperparameters.

4. The method for remote sensing maritime target hash positioning encryption in a tropospheric waveguide environment according to claim 3 is characterized by: The loss function L of the encrypted block discriminator D Defined as: In the above formula, To encrypt the image The corresponding extended encryption tag, For the input image x (m) The corresponding extended original tag, 5. A remote sensing maritime target hash positioning encryption device in a tropospheric waveguide environment, characterized in that: include: processor; A memory having executable instructions for the processor stored therein; the processor is configured to perform the steps of the method of claim 1 by executing the executable instructions.

6. A computer-readable storage medium for storing a program, characterized in that: When the program is executed, the steps of the method according to claim 1 are implemented.

Images