Medical image encryption system and method based on double adversarial neural network

Abstract

The invention relates to a medical image encryption system and method based on a double adversarial neural network, and belongs to the technical field of image security protection. According to the method, an image encryption system is designed by using nonlinearity and instability of a deep neural network, firstly, two pairs of mutually confronting neural networks are adopted as a main framework of the encryption system, then random noise-like images with ideal characteristics are constructed to supervise encryption network training, and original medical image data are used to guide decryption network training. The encryption system provided by the invention dynamically generates a key with strong randomness and high sensitivity through adversarial training, and finally encrypts input medical image data in combination with an encryption network architecture and key parameters.

Description

Technical Field

[0001] The present invention belongs to the technical field of image security protection, and relates to a medical image encryption system and method based on a double adversarial neural network. Background Art

[0002] With the development of the Internet of Medical Things (IoMT), smart medical devices have been widely deployed and applied in the field of healthcare. These smart medical devices achieve efficient collection, rapid transmission and accurate processing of medical data by connecting to the network, which greatly promotes the popularization of telemedicine services. Medical images play an important role in telemedicine services, such as lesion segmentation and disease diagnosis. However, medical images contain a large amount of sensitive information of patients, such as physical characteristics and details of their condition, which are highly private. When medical images are transmitted over the network, they are vulnerable to illegal access. Once the medical image information is leaked, it will pose a serious threat to the privacy and security of patients. Traditional schemes based on DNA coding methods and chaos theory have played a certain role in protecting the security of medical images. For example, encryption schemes based on DNA coding utilize the unique properties of DNA sequences and attempt to encrypt medical images by encoding and converting them. However, this scheme often relies on DNA coding rules and limited coding combinations, resulting in a relatively limited key space. In the face of increasingly powerful computing power and diverse attack methods, the limited key space makes it easy for attackers to crack encryption systems by exhaustive key enumeration. At the same time, the computational complexity of the DNA encoding process may also lead to low efficiency in encryption and decryption, affecting the real-time transmission and processing of medical images. Chaos theory uses the nonlinear dynamic characteristics of chaotic systems to generate encryption keys and perform image encryption operations. Chaotic systems are sensitive to initial conditions and have strong randomness, which provides a good foundation for encryption. However, due to the limited precision representation of computers, chaotic systems may experience dynamic degradation during digital implementation, resulting in the generated key sequence being periodic, which greatly reduces the randomness and security of the key. Secondly, the parameter setting and key generation process of chaotic systems are relatively complex and difficult to accurately control, and are susceptible to external interference and attacks, which makes the stability and reliability of the encryption system questionable. With the continuous development of deep learning technology, it has shown great potential in feature extraction and nonlinear mapping. Deep neural network models can automatically learn complex features and patterns in data and have powerful representation capabilities. However, existing neural network-based encryption schemes still have obvious shortcomings, and the key used in the encryption process remains unchanged throughout the encryption cycle. This static key makes the encryption system vulnerable to attack. Once the attacker obtains the key, it can easily decrypt the medical images encrypted using the key. In addition, existing encryption rules are often pre-set and fixed, lacking flexibility and adaptability, and cannot be dynamically adjusted according to different security requirements and image features, making them powerless when faced with a variety of attack methods.

[0003] Therefore, a new medical image encryption method is urgently needed to provide a more secure, efficient and flexible solution. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide a medical image encryption system and method based on a dual adversarial neural network, which is used to solve the security problem of medical images during network transmission. The method of the present invention designs a unique dual adversarial network architecture and makes full use of the adversarial training mechanism to generate dynamic keys to ensure the secure transmission of medical images.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] Option 1:

[0007] A medical image encryption method based on a dual adversarial neural network specifically includes the following steps:

[0008] S1: Architecture construction: Build an encryption system of dual adversarial neural networks consisting of an encryption network, a decryption network, a ciphertext discrimination network, and a plaintext discrimination network;

[0009] S2: Network structure and training target design: Based on deep learning technology, the structures of the four networks in step S1 are designed, and their training targets are determined so that the encryption network and the decryption network form an adversarial relationship with the corresponding discriminant network respectively;

[0010] S3: Key generation: Generate dynamic encryption keys and decryption keys through continuous training of the encryption system, medical image datasets, and the generated pseudo-ciphertext image data, where the encryption key and decryption key are the weight parameters of the encryption network and decryption network respectively;

[0011] S4: Encryption: read the medical image to be encrypted and the encryption key, assign the encryption key as a weight parameter to the encryption network, input the medical image into the encryption network, and obtain the ciphertext image after calculation;

[0012] S5: Decryption: Read the ciphertext image to be decrypted and the decryption key, assign the decryption key as a weight parameter to the decryption network, input the ciphertext image into the decryption network, and obtain the original medical image after calculation.

[0013] Furthermore, in step S2, the network structure and training target design specifically include:

[0014] (1) Design the encryption process of the encryption network, including three sub-processes: feature encoding, feature transformation, and feature decoding;

[0015] (2) The ciphertext discrimination network and the plaintext discrimination network adopt the same structure, both of which are designed with the PatchGAN architecture;

[0016] (3) Calculate the adversarial loss of the encryption network and the difference between the decrypted image and the plaintext image, and weight them with the network adversarial loss as one of the training objectives of the encryption system;

[0017] (4) The decryption network is structurally consistent with the encryption network.

[0018] Further, in step S3, the key generation specifically includes the following steps:

[0019] S31: generating pseudo-ciphertext image data, including generating a random sequence using a random sequence generator and performing a randomness test, combining sequence data satisfying the randomness test into a ciphertext image, and performing an entropy test and a pixel correlation test, etc.;

[0020] S32: Use medical image data sets and pseudo-ciphertext image data to train the network in the encryption system, so that the encryption network and the decryption network form an adversarial relationship with the ciphertext discrimination network and the plaintext discrimination network respectively;

[0021] S33: After the network forms a stable adversarial relationship, the weight parameters of the model are read and saved as encryption keys and decryption keys.

[0022] Further, in step S4, the encryption specifically includes the following steps:

[0023] S41: reading the medical image to be encrypted and performing preprocessing;

[0024] S42: The preprocessed medical image is input into the encryption network, and the encryption network calculates layer by layer according to the encryption key to obtain a ciphertext image.

[0025] Further, in step S5, decryption specifically includes the following steps:

[0026] S51: Read the ciphertext image to be decrypted and perform preprocessing;

[0027] S52: The preprocessed ciphertext image is input into the decryption network, and the decryption network calculates layer by layer according to the decryption key to obtain the original medical image.

[0028] Option 2:

[0029] A medical image encryption system based on dual adversarial neural network, comprising:

[0030] An encryption network, used to convert the input plaintext medical image into a ciphertext image;

[0031] A decryption network, used to decrypt the input ciphertext image into the original medical image;

[0032] The ciphertext discrimination network is used to distinguish whether the input image is a pseudo-ciphertext or a real ciphertext image;

[0033] Plaintext discrimination network, used to distinguish whether the input image is a decrypted image or an original plaintext image;

[0034] Among them, the encryption network, decryption network, ciphertext discrimination network and plaintext discrimination network are all constructed based on deep learning technology, and form adversarial relationships through training to generate dynamic encryption keys and decryption keys.

[0035] The encryption network and decryption network are suitable for encryption and decryption of medical images, ensuring the security of medical images during network transmission.

[0036] Preferably, the encryption network includes three sub-processes: feature encoding, feature transformation and feature decoding, and both the ciphertext discrimination network and the plaintext discrimination network are designed using the PatchGAN architecture.

[0037] Preferably, the system also includes a pseudo-ciphertext image data generation module for generating pseudo-ciphertext image data to assist in the training of the encryption network and the decryption network.

[0038] Preferably, the encryption key and decryption key are weight parameters of the encryption network and decryption network, respectively, and are generated by continuous optimization and adjustment during the training process.

[0039] The beneficial effects of the present invention are as follows: in view of the security challenges faced by medical image transmission in open networks, the present invention uses the nonlinearity and instability of neural networks to design an encryption model, and designs two paired adversarial networks as the core framework of the encryption system. The model generates dynamic keys during adversarial training, overcomes the limitations of traditional encryption schemes that use fixed strategies, and maximizes the confidentiality of medical image transmission.

[0040] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0042] Figure 1 This is a schematic diagram of the architecture of the medical image encryption system based on a dual adversarial neural network of the present invention;

[0043] Figure 2 A schematic diagram for generating encryption keys;

[0044] Figure 3A schematic diagram of an encryption scenario. DETAILED DESCRIPTION

[0045] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0046] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0047] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0048] See also Figure 1 to Figure 3 The present invention proposes a medical image encryption method based on a dual adversarial neural network, which specifically includes the following steps:

[0049] Step 1: Architecture construction: Build an encryption system with double adversarial neural network.

[0050] Step 1 specifically includes the following steps:

[0051] Step 11: Build four deep neural network models to form an encryption system;

[0052] Reference Figure 1The encryption system proposed in the present invention is composed of four networks, namely, an encryption network, a decryption network, a ciphertext discrimination network and a plaintext discrimination network. These four networks form two pairs of adversarial networks, and the two adversarial networks cooperate and restrict each other to form a medical image encryption system.

[0053] Step 12: Design the functions and interaction processes of the four networks;

[0054] The encryption network undertakes the important task of converting the input plaintext image into ciphertext. During the encryption process, the ciphertext will be input into the ciphertext discrimination network together with the constructed pseudo-ciphertext image. The encryption network continuously optimizes and improves its own encryption process by conducting adversarial learning with the ciphertext discrimination network. Its specific working principle is that the encryption network strives to make the output ciphertext as similar to the pseudo-ciphertext as possible in terms of features and forms of expression, thereby confusing the ciphertext discrimination network and making it difficult for the ciphertext discrimination network to accurately distinguish whether the input is real ciphertext or pseudo-ciphertext.

[0055] The decryption network is mainly responsible for decrypting the ciphertext image. After a series of complex calculations and processing, the ciphertext is decrypted into the original image. Subsequently, the decrypted image and the plaintext image are input into the plaintext discrimination network. During the training process, the decryption network continuously learns and adjusts its own decryption algorithm through adversarial training with the plaintext discrimination network, striving to accurately decrypt the image, so that the final decrypted image is indistinguishable from the original image in all aspects (such as pixel value, image structure, texture, etc.).

[0056] Step 2: Network structure and training objective design: Design four deep learning network models and determine their objectives.

[0057] Step 2 specifically includes the following steps:

[0058] Step 21: Construct the structure of the designed neural network based on deep learning technology;

[0059] The encryption process of the encryption network is a multi-step complex process, which mainly includes three key sub-processes: feature encoding, feature transformation, and feature decoding. In the initial stage of encryption, the input plaintext image will first undergo a convolution operation to extract coarse-grained features from the original plaintext image. These coarse-grained features can initially capture some overall information of the image, such as contours, approximate pixel distribution, etc. After completing the convolution operation, the image will be further processed more carefully through a series of convolutions. These convolutions can deeply explore the detailed features of the image and perform more accurate analysis and extraction of local information of the image. Then, multiple residual blocks are used to transform the extracted features. These residual blocks enable the network to effectively learn complex feature relationships and significantly enhance the diffusion effect of the encryption system, so that the features of the image are more fully transformed. After a series of feature transformations, upsampling and feature decoding operations are finally performed through convolution to convert the deeply transformed features into ciphertext images to complete the entire encryption process. The decryption network is structurally consistent with the encryption network. This symmetrical structural design helps to better restore the various transformations performed in the encryption process during the decryption process, thereby improving the accuracy of decryption.

[0060] The ciphertext discrimination network uses the PatchGAN architecture, which has unique advantages in processing image discrimination tasks. During the working process, the ciphertext discrimination network first randomly selects an N×N image block from the input image, then downsamples the image block, extracts the feature information in the image block through a series of convolution operations, and finally determines whether the input image is pseudo-ciphertext or ciphertext based on this feature information. The plaintext discrimination network and the ciphertext discrimination network use the same structure to ensure that the uniform discrimination standard and processing flow are maintained throughout the system.

[0061] Step 22: Design training objectives for the four networks to ensure that the four networks are trained collaboratively.

[0062] In order to accelerate the convergence of the model, improve the training efficiency, and ensure the stability of network training, instance normalization technology is used after each convolution. Instance normalization can effectively normalize the features output by each convolution layer, so that the network can better capture the local features and texture information of the image. At the same time, the LeakyReLU activation function is applied to the entire network, which can introduce nonlinear characteristics to the model and enhance the model's ability to handle complex data relationships. The core goal of the encryption network is to make the output ciphertext and pseudo-ciphertext difficult to distinguish in various features. First, the adversarial loss of the neural network is calculated. At the same time, in order to ensure the accuracy and stability of the encryption and decryption process, the difference between the decrypted image and the plaintext image is calculated to constrain the encryption system, and it is weighted with the network adversarial loss as one of the training objectives of the encryption system.

[0063] The main goal of the decryption network is to accurately decrypt the plaintext image from the ciphertext and confuse it as much as possible in the confrontation process with the plaintext discriminant network. When the plaintext discriminant network cannot accurately distinguish whether the input is the original plaintext image or the decrypted image, it means that the decryption network has successfully achieved the goal of accurately decrypting the image. To achieve this goal, similar to the encryption network, the difference between the decrypted image and the plaintext image is weighted with the network adversarial loss as one of the training objectives of the encryption system.

[0064] The loss of the ciphertext discrimination network is the error between the output image label and the true label. The design of this loss function enables the ciphertext discrimination network to continuously adjust its discrimination ability during the training process, striving to accurately distinguish pseudo-ciphertext from ciphertext. An adversarial relationship is formed between the ciphertext discrimination network and the encryption network. Through this adversarial relationship, the encryption network can continuously optimize its own encryption algorithm, making the generated ciphertext more difficult to be recognized by the ciphertext discrimination network, thereby improving the security of the encryption system.

[0065] The main function of the plaintext discrimination network is to distinguish the input restored image from the original plaintext image. Its loss function is the error value between the output image label and the true label. During the training process, the plaintext discrimination network forms an adversarial relationship with the decryption network by constantly adjusting its own discrimination ability. As the training progresses, the decryption network gradually learns a more accurate decryption algorithm, making the decrypted image closer and closer to the original plaintext image, making it difficult for the plaintext discrimination network to distinguish between the two, improving the decryption accuracy of the decryption network and the stability of the entire system.

[0066] Step 3: Key generation: Generate a large number of keys through continuous training of the network model.

[0067] Step 3 specifically includes the following steps:

[0068] Step 31: Generate pseudo-ciphertext image data.

[0069] First, a random sequence generator is used to generate a random sequence and a randomness test is performed on the sequence. Then, the sequence data that meet the randomness test are combined into a ciphertext image. Finally, the generated pseudo-ciphertext image data is subjected to an entropy test, a pixel correlation test, etc., and the pseudo-ciphertext images that meet the test are saved as a training data set.

[0070] Step 32: Use the medical image dataset and the pseudo-ciphertext image data to train the network in the encryption system.

[0071] In the present invention, the parameters of the encryption network and the decryption network have dual identities, and they act as keys in the encryption and decryption processes at the same time. When the encryption network is initialized, the parameters of each convolutional layer are randomly assigned initial values. During the entire training process, the network propagates the loss information of the entire network to each convolutional layer through the back propagation mechanism. Specifically, back propagation calculates the gradient of each parameter according to the loss function, and then uses the gradient information to update the parameters of each convolutional layer through the gradient descent method. As the training continues, the key (i.e., the network parameters) will be continuously updated and optimized in this process, so that the key is dynamic and adaptable. The generation process of the decryption key follows a similar principle and mechanism to the generation process of the encryption key. By alternately training the encryption network and the decryption network, when both networks reach a convergence state, the obtained network parameters are the final encryption and decryption keys. This dynamic key generation method makes the key highly random and unpredictable, greatly improves the security of the encryption system, and effectively resists various potential attacks.

[0072] Step 33: After the network forms a stable adversarial relationship, read the model weight parameters and save them as a key.

[0073] When both networks reach convergence, the networks reach an adversarial equilibrium state, and the parameters of the model are read and saved as the most secure encryption and decryption keys.

[0074] Step 4: Encryption: Use the generated network model and encryption key to perform medical image operations to obtain the ciphertext image.

[0075] Step 4 specifically includes the following steps:

[0076] Step 41: Read the medical image and obtain the encryption key;

[0077] The medical image file to be encrypted is read from the device, and the encryption key is obtained from a secure key management system or key storage location.

[0078] Step 42: Assign the encryption key as a weight parameter to the encryption network, input the image into the encryption network, and obtain the ciphertext image after calculation.

[0079] When performing encryption operations, the medical image to be encrypted must first be preprocessed, converted into a tensor form, and then the tensor is input into the encryption network. The encryption network will calculate layer by layer according to the key. Specifically, for each layer of the network (starting from the input layer, passing through multiple hidden layers in sequence, until the output layer), the corresponding convolution kernel will be used to perform a convolution operation on the input tensor elements. For each convolution kernel, each element in the input tensor will be traversed for convolution calculation to obtain a series of intermediate results. Then, these intermediate results are summed to obtain a feature map. Next, for all convolution kernels of the same layer, the above convolution and summation operations are repeated to obtain a series of feature maps. Subsequently, these feature maps are spliced ​​along a specific dimension to form a new tensor as the input of the next layer of the network. This is repeated, and after convolution, summation, splicing and other operations of multiple layers of the network, the result output from the last layer of the encryption network is finally converted into a ciphertext image.

[0080] Step 5: Decryption: Use the generated network model and decryption key to operate on the ciphertext image to obtain the plaintext image. Step 5 specifically includes the following steps:

[0081] Step 51: Read the ciphertext image and obtain the decryption key.

[0082] The ciphertext image file to be decrypted is read from the device, and the decryption key is obtained from a secure key management system or a key storage location.

[0083] Step 52: Assign the decryption key as a weight parameter to the decryption network, input the ciphertext image into the decryption network, and obtain the original medical image after calculation.

[0084] The decryption process is structurally the same as the encryption process, but uses different keys (i.e., different parameters obtained by the encryption network and the decryption network after training convergence). During decryption, the ciphertext is input into the decryption network, and the decryption network processes the ciphertext in the same operation flow as the encryption process. Through a series of convolution, feature transformation, upsampling and other operations, the information contained in the ciphertext is gradually restored, and finally the ciphertext is restored to the original plaintext image, thus completing the entire decryption process.

[0085] The present invention applies the nonlinearity and instability of neural networks to the medical image encryption system and designs two paired adversarial networks as the core framework of the cryptographic system. The features of the ideal ciphertext image are used to construct the pseudo-ciphertext image, and the medical image dataset and pseudo-ciphertext image are used to guide the network training. The encryption system generates dynamic keys during adversarial training, overcoming the limitations of traditional encryption schemes that use fixed strategies.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A medical image encryption method based on dual adversarial neural network, characterized in that: The method specifically comprises the following steps: S1: Architecture construction: Build an encryption system of dual adversarial neural networks consisting of an encryption network, a decryption network, a ciphertext discrimination network, and a plaintext discrimination network; S2: Network structure and training target design: Based on deep learning technology, the structures of the four networks in step S1 are designed, and their training targets are determined so that the encryption network and the decryption network form an adversarial relationship with the corresponding discriminant network respectively; S3: Key generation: Generate dynamic encryption keys and decryption keys through continuous training of the encryption system, medical image datasets, and the generated pseudo-ciphertext image data, where the encryption key and decryption key are the weight parameters of the encryption network and decryption network respectively; S4: Encryption: read the medical image to be encrypted and the encryption key, assign the encryption key as a weight parameter to the encryption network, input the medical image into the encryption network, and obtain the ciphertext image after calculation; S5: Decryption: Read the ciphertext image to be decrypted and the decryption key, assign the decryption key as a weight parameter to the decryption network, input the ciphertext image into the decryption network, and obtain the original medical image after calculation.

2. The medical image encryption method according to claim 1, characterized in that: In step S2, the network structure and training target design specifically include: (1) Design the encryption process of the encryption network, including three sub-processes: feature encoding, feature transformation, and feature decoding; (2) The ciphertext discrimination network and the plaintext discrimination network adopt the same structure, both of which are designed with the PatchGAN architecture; (3) Calculate the adversarial loss of the encryption network and the difference between the decrypted image and the plaintext image, and weight them with the network adversarial loss as one of the training objectives of the encryption system; (4) The decryption network is structurally consistent with the encryption network.

3. The medical image encryption method according to claim 1, characterized in that: In step S3, key generation specifically includes the following steps: S31: generating pseudo-ciphertext image data, including generating a random sequence using a random sequence generator and performing a randomness test, combining sequence data satisfying the randomness test into a ciphertext image, and performing an entropy test and a pixel correlation test; S32: Use medical image data sets and pseudo-ciphertext image data to train the network in the encryption system, so that the encryption network and the decryption network form an adversarial relationship with the ciphertext discrimination network and the plaintext discrimination network respectively; S33: After the network forms a stable adversarial relationship, the weight parameters of the model are read and saved as encryption keys and decryption keys.

4. The medical image encryption method according to claim 1, characterized in that: In step S4, encryption specifically includes the following steps: S41: reading the medical image to be encrypted and performing preprocessing; S42: The preprocessed medical image is input into the encryption network, and the encryption network calculates layer by layer according to the encryption key to obtain a ciphertext image.

5. The medical image encryption method according to claim 1, characterized in that: In step S5, decryption specifically includes the following steps: S51: Read the ciphertext image to be decrypted and perform preprocessing; S52: The preprocessed ciphertext image is input into the decryption network, and the decryption network calculates layer by layer according to the decryption key to obtain the original medical image.

6. A medical image encryption system based on dual adversarial neural network, characterized in that: The system includes: An encryption network, used to convert the input plaintext medical image into a ciphertext image; A decryption network, used to decrypt the input ciphertext image into the original medical image; The ciphertext discrimination network is used to distinguish whether the input image is a pseudo-ciphertext or a real ciphertext image; Plaintext discrimination network, used to distinguish whether the input image is a decrypted image or an original plaintext image; Among them, the encryption network, decryption network, ciphertext discrimination network and plaintext discrimination network are all constructed based on deep learning technology, and form adversarial relationships through training to generate dynamic encryption keys and decryption keys.

7. The medical image encryption system according to claim 6, characterized in that: The encryption network includes three sub-processes: feature encoding, feature transformation and feature decoding. The ciphertext discrimination network and the plaintext discrimination network are both designed using the PatchGAN architecture.

8. The medical image encryption system according to claim 6 or 7, characterized in that: The system also includes a pseudo-ciphertext image data generation module, which is used to generate pseudo-ciphertext image data to assist in the training of encryption networks and decryption networks.

9. The medical image encryption system according to claim 6, characterized in that: The encryption key and decryption key are weight parameters of the encryption network and the decryption network respectively, and are generated through continuous optimization and adjustment during the training process.

Images