Feature confusion processing method and device for AI generation content, equipment and medium

The AI-generated content features are extracted and obfuscated through convolutional neural networks and deep learning algorithms, which solves the problem of digital watermark affecting visual quality, realizes the recognition of AI-generated content, and improves security.

CN119068520APending Publication Date: 2024-12-03SHANGHAI TONGSHI INFORMATION TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202310653515.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2023-06-02
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

Existing digital watermarking technology will affect the visual quality of AI-generated content, making it difficult to distinguish between AI-generated content from real content, leading to challenges in areas such as identity verification, copyright protection and network security.

Method used

Features of AI generated content are extracted through convolutional neural networks and deep learning algorithms, and obfuscated algorithms are used for obfuscating processing. The obfuscated features are embedded in the original content, and obfuscated rich media content is generated, and frequency domain embedding or graph layering is used to superimpose the obfuscated features. way to maintain visual fidelity and facilitate computer recognition.

Benefits of technology

It significantly reduces the visual quality impact on the original content, makes the confused content difficult to distinguish from the original content, improves the recognition ability of AI-generated content, and enhances the security of identity verification, copyright protection and network security.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119068520A_ABST
    Figure CN119068520A_ABST
Patent Text Reader

Abstract

The invention provides a feature confusion processing method and device for AI generation content, equipment and a medium. The method comprises the following steps: performing feature extraction on original rich media content generated by AI through a convolutional neural network and a deep learning image processing algorithm; performing confusion processing on the extracted features by using a confusion algorithm; and embedding the confused features into the original rich media content to generate confused rich media content, wherein the features are used for identifying the confused rich media content generated by the AI. According to the method, unperceived confusion processing is only carried out on part of the features of the original rich media content, so that the influence on the visual quality of the original rich media content is greatly reduced, and the confused rich media content and the original rich media content are difficult to distinguish by human eyes.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method, device, equipment and medium for processing feature obfuscation of AI-generated content. Background Art

[0002] With the rapid development of artificial intelligence technology, rich media content such as images, videos and texts generated by artificial intelligence (AI) are becoming more and more realistic. This has positive application value in many fields, but it also brings a series of potential problems. For example, it is becoming increasingly difficult to distinguish between AI-generated human images and real human photos. This creates challenges in identity authentication, copyright protection and network security.

[0003] Digital watermarking is a technology that embeds specific information into digital media content. Although digital watermarking can be used to identify AI-generated content, it will affect the visual quality of the media content.

[0004] Therefore, it is necessary to propose a feature obfuscation processing method, device, equipment and medium for AI-generated content to solve the above problems. Summary of the invention

[0005] The purpose of the present invention is to provide a method, device, equipment and medium for feature obfuscation processing of AI-generated content, so as to solve the problem that the existing digital watermarking technology will affect the visual quality of the content.

[0006] In a first aspect, the present application provides a method for feature obfuscation processing of AI-generated content, the method comprising: extracting features of original rich media content generated by AI through a convolutional neural network and a deep learning image processing algorithm; performing obfuscation processing on the extracted features using an obfuscation algorithm; embedding the obfuscated features into the original rich media content to generate obfuscated rich media content, wherein the features are used to identify the obfuscated rich media content generated by AI.

[0007] The beneficial effect of the method provided by the present invention is that compared with digital watermarking that directly embeds information of non-original rich media content into original rich media content, the present invention extracts features of the original rich media content, performs obfuscation processing on the features, and then embeds the obfuscated features into the original rich media content to generate obfuscated rich media content. Since only part of the features of the original rich media content are obfuscated in an imperceptible manner, the impact on the visual quality of the original rich media content is greatly reduced, making it difficult for the human eye to distinguish between the obfuscated rich media content and the original rich media content.

[0008] In a possible embodiment, after the extracted features are obfuscated using an obfuscation algorithm, the method further includes:

[0009] Encode the obfuscated features;

[0010] The obfuscated features are embedded into the original rich media content to generate obfuscated rich media content, including:

[0011] The encoded features are embedded into the original rich media content to generate obfuscated rich media content.

[0012] In a possible embodiment, embedding the obfuscated feature into the original rich media content to generate the obfuscated rich media content includes:

[0013] The obfuscated features are embedded into the original rich media content by frequency domain embedding and layer overlay to generate obfuscated rich media content.

[0014] In a possible embodiment, the extracted features are obfuscated using an obfuscation algorithm, including:

[0015] The extracted features are randomly disturbed or locally deformed using an obfuscation algorithm, wherein the random disturbance includes but is not limited to adding random noise and specific patterns.

[0016] In a possible embodiment, before extracting features from the original rich media content generated by AI using a convolutional neural network or a deep learning image processing algorithm, the process further includes:

[0017] Establishing a correction model, wherein the correction model is used to correct the obfuscated rich media content;

[0018] A modified adversarial network is established, wherein the modified adversarial network is used to perform adversarial generation on the modified model.

[0019] In a possible embodiment, after establishing the correction model, the method further includes:

[0020] The modified model is trained using a training dataset containing obfuscated data.

[0021] In the second aspect, the present application also proposes a feature obfuscation processing device for AI-generated content, which includes a module / unit that executes any possible design method of the first aspect above. These modules / units can be implemented by hardware, or by hardware executing corresponding software implementations.

[0022] In a third aspect, an embodiment of the present application provides an electronic device, including a processor and a memory, wherein the memory is used to store one or more computer programs; when the one or more computer programs stored in the memory are executed by the processor, the electronic device can implement any possible design method of the first aspect.

[0023] In a fourth aspect, a computer-readable storage medium is also provided in an embodiment of the present application. The computer-readable storage medium includes a computer program. When the computer program runs on an electronic device, the electronic device executes any possible design method of the first aspect mentioned above.

[0024] In a fifth aspect, an embodiment of the present application further provides a method comprising a computer program product, which, when executed on an electronic device, enables the electronic device to execute any possible design of any of the above aspects.

[0025] For the beneficial effects of the second to fifth aspects, reference may be made to the description of the first aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a flow chart of a first embodiment of the feature obfuscation processing method for AI-generated content according to the present invention.

[0027] Figure 2 This is a flow chart of a first embodiment of the feature confusion processing method for AI generated content according to the present invention for facial images.

[0028] Figure 3 This is a flow chart of the feature confusion processing method for AI generated content in the second embodiment of the present invention for face images.

[0029] Figure 4 Schematic diagram of a feature obfuscation processing device for AI-generated content according to the present invention.

[0030] Figure 5 It is a schematic diagram of the structure of the electronic device of the present invention. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. Unless otherwise defined, the technical terms or scientific terms used herein should be understood by people with general skills in the field to which the present invention belongs. "Including" and similar words used in this article mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects.

[0032] In view of the problems existing in the prior art, the first embodiment of the present invention provides a feature obfuscation processing method for AI-generated content. Figure 1 This is a flow chart of the feature obfuscation processing method for AI-generated content of the present invention, see Figure 1 , the method comprising:

[0033] S101: extracting features of original rich media content generated by artificial intelligence (AI) through convolutional neural networks and deep learning image processing algorithms, wherein the extracted features include but are not limited to texture, color, and shape.

[0034] S102: Obfuscate the extracted features using an obfuscation algorithm.

[0035] S103: Embed the obfuscated features into the original rich media content to generate obfuscated rich media content, wherein the features are used to identify the obfuscated rich media content generated by AI.

[0036] In a first embodiment, the obfuscated features are embedded into the original rich media content to generate the obfuscated rich media content, including: using a generative adversarial network to fuse the obfuscated features with the original rich media content to generate the obfuscated rich media content.

[0037] In some specific embodiments, the obfuscated features are embedded into the original rich media content to generate the obfuscated rich media content, including: embedding the obfuscated features into the original rich media content by frequency domain embedding and layer overlay to generate the obfuscated rich media content.

[0038] Furthermore, the obfuscated features are embedded into the original rich media content by frequency domain embedding and layer overlay to generate the obfuscated rich media content, including: embedding the obfuscated features into a layer with lower transparency and a frequency domain in the original rich media content to generate the obfuscated rich media content. In this embodiment, the layers with lower transparency and frequency domain embedding are selected to embed in a less noticeable manner to maintain a more realistic visual effect, so as to ensure that the human eye cannot distinguish.

[0039] In a preferred embodiment, the extracted features are obfuscated using an obfuscation algorithm, including: randomly perturbing or locally deforming the extracted features using an obfuscation algorithm, wherein the random perturbation includes but is not limited to adding random noise and specific patterns.

[0040] The feature confusion processing method of the first embodiment is described in detail below in conjunction with face images. Figure 2FIG. 1 is a flow chart of a feature obfuscation processing method for AI-generated content in a first embodiment of the present invention for a face image, see Figure 2 , a feature confusion processing method for face images generated by artificial intelligence, including:

[0041] S201: Face feature extraction: Use face recognition algorithms, such as the face detection method based on HOG (Histogram of Oriented Gradients) features provided by the Dlib library, to extract key feature points from AI-generated face images, including eyes, nose, mouth, etc. Among them, the dlib library is an open source library for machine learning, which contains many machine learning algorithms. It is very convenient to use. You can directly include the header file and it does not depend on other libraries (it comes with image encoding and decoding library source code).

[0042] S202: Randomly perturb key feature points: Randomly perturb the detected key feature points, for example, add random noise around the feature points to facilitate recognition and distinction by a computer system while maintaining visual fidelity.

[0043] S203: Generate an obfuscated face image: Use a generative adversarial network (GAN) to fuse the perturbed key feature points with the original face image to generate an obfuscated face image.

[0044] S204: Recognize facial images: Train a deep learning model to enable the deep learning model to recognize the facial image to be detected, thereby determining whether the image is an obfuscated facial image generated by AI, wherein the deep learning model may be a convolutional neural network (CNN).

[0045] The first embodiment of the face feature obfuscation mainly focuses on randomly perturbing key feature points, and uses a generative adversarial network (GAN) for fusion to generate an obfuscated face image. The advantage of this solution is that it can generate an obfuscated image with high fidelity and strong robustness.

[0046] In the second embodiment, after the extracted features are obfuscated using an obfuscation algorithm, the method further includes: encoding the obfuscated features;

[0047] Embedding the obfuscated features into the original rich media content to generate the obfuscated rich media content includes: embedding the encoded features into the original rich media content to generate the obfuscated rich media content.

[0048] In the second embodiment, the obfuscated features are encoded using an unsupervised learning method.

[0049] In some specific embodiments, the encoding method includes but is not limited to a QR code and a data matrix.

[0050] The feature confusion processing method of the second embodiment is described in detail below in conjunction with face images. Figure 3 FIG. 1 is a flow chart of a feature obfuscation processing method for AI-generated content in a second embodiment of the present invention for face images, see Figure 3 , a feature confusion processing method for face images generated by artificial intelligence, including:

[0051] S301: Face feature extraction: Use deep learning methods, such as the OpenFace algorithm based on Convolutional Neural Networks (CNN), to extract recognizable feature vectors from AI-generated face images. OpenFace is an image + machine learning project, and the overall process includes: face discovery, feature extraction, feature neural network training, and face recognition.

[0052] S302: Slight local deformation: Slight local deformation is performed on the original face image, such as slightly changing the shape and position of the eyes, nose, mouth and other areas, so as to achieve the purpose of feature confusion.

[0053] S303: encoding based on feature transformation: encoding the features of the locally deformed facial image, and an autoencoder (AE) network or other unsupervised learning methods may be used to implement feature encoding.

[0054] in,

[0055] S304: embedding coded features: embedding the coded features into the original face image to generate a confused face image, which can be achieved through frequency domain embedding, layer overlay and other methods.

[0056] S305: Recognize facial images: Train a deep learning model to enable the deep learning model to recognize obfuscated facial images, thereby determining whether the image is an obfuscated facial image generated by AI, wherein the deep learning model may be a convolutional neural network (CNN).

[0057] The second embodiment of the human face feature obfuscation focuses on performing a slight local deformation on the original human face image and adopts a coding method based on feature transformation. The advantage of this solution is that the calculation complexity is low and it is easy to implement.

[0058] Both the first embodiment and the second embodiment can facilitate the computer system to identify and distinguish the obfuscated face images generated by AI while maintaining visual fidelity. This helps to solve the problem of distinguishing AI-generated content from real content in the field of face recognition. The first embodiment and the second embodiment can be selected and adjusted according to actual application scenarios and needs. In practical applications, it may be necessary to optimize the algorithm to adapt to different face generation technologies and recognition systems. In addition, it is possible to consider combining these two solutions to achieve more efficient and accurate facial feature obfuscation and recognition.

[0059] In a preferred embodiment, before extracting features from the original rich media content generated by AI through a convolutional neural network and a deep learning image processing algorithm, it also includes: establishing a correction model, which is used to correct the obfuscated rich media content; establishing a correction adversarial network, which is used to perform adversarial generation on the correction model, and its goal is to make the correction model unable to correct the confusion to ensure the robustness of the confusion. For example, when confusing, ensure a larger feature space and design multiple local minima. Specifically, the correction model is a deep learning model, which can be a convolutional neural network, a generative adversarial network, etc. In this embodiment, based on the adversarial generation method of the generative adversarial network, the obfuscation algorithm is automatically iterated and improved, and the specific steps are as follows:

[0060] When the extracted features are obfuscated, the extracted feature information is stored. Specifically, storing the extracted feature information includes: storing the extracted feature information in a secure database or other storage medium.

[0061] The correction model is used to correct the obfuscated rich media content, specifically including: after identifying the obfuscation features through the deep learning model, the obfuscated rich media content is corrected. This correction model can be trained and adjusted according to actual application scenarios and requirements to achieve effective correction of various obfuscation features.

[0062] The generative adversarial network (GAN) is used to generate adversarial versions of the modified model to achieve automatic iterative obfuscation algorithm and improve its adversarial performance.

[0063] In a specific embodiment, establishing a modified adversarial network includes:

[0064] Definition of generator (obfuscator): The generator is responsible for creating obfuscated rich media content. Its goal is to generate obfuscated rich media content with high fidelity, making it difficult for the correction model to identify and correct these obfuscated rich media contents.

[0065] Define the discriminator (correction model): The discriminator is responsible for identifying and correcting the obfuscated rich media content. Its goal is to identify and correct the obfuscated rich media content generated by the generator as accurately as possible.

[0066] Adversarial training: Through the adversarial training process, the generator and the discriminator compete with each other to continuously improve their respective performance. The generator tries to generate increasingly difficult to identify obfuscated rich media content, while the discriminator strives to improve its ability to identify and correct these obfuscated rich media content.

[0067] Loss function and optimization: Define the loss function to measure the performance of the generator and discriminator. Usually, the loss function includes the misclassification rate of the rich media content generated by the generator in the discriminator and the accuracy of the discriminator for real images. By optimizing the loss function, automatic iteration and improvement of the generator and discriminator can be achieved.

[0068] Iterative update: During the training process, the parameters of the generator and discriminator are gradually updated to gradually adapt to each other's strategies. As the training progresses, the generator will generate increasingly difficult to correct obfuscated rich media content, while the discriminator will improve its correction ability.

[0069] Termination condition: Set the termination condition, such as when the loss function reaches the preset threshold or the number of training rounds reaches the set value, stop training. At this point, the parameters of the generator and discriminator will reach a relatively stable state, which can be used as the final obfuscation algorithm and correction model.

[0070] Through this adversarial generation method based on generative adversarial networks, the obfuscation algorithm can be automatically iterated and improved, thereby improving its adversarial performance. It should be noted that in practical applications, the model structure, loss function, etc. may need to be adjusted according to specific needs and scenarios to achieve more efficient and accurate adversarial generation effects.

[0071] When it is necessary to correct the features after the obfuscation process, the corrected rich media content is verified according to the feature information.

[0072] In a preferred embodiment, after the correction model is established, the method further includes: training the correction model using a training data set containing obfuscated data. Directly including obfuscation in the training process allows the correction model to adapt to and recognize obfuscated data during the training phase, thereby having greater robustness and accuracy in practical applications.

[0073] For technologies such as Low-Rank Adaptation of Large Language Models (LoRA), when training a correction model, a training data set containing obfuscated data is used to train the correction model, including:

[0074] Data preprocessing: Adding obfuscated data to the original data set, specifically by combining the original signal with obfuscation methods such as random noise and distortion. In this process, the degree and type of obfuscation can be controlled to ensure that the training data set has sufficient diversity and difficulty.

[0075] Data Annotation: Annotate the preprocessed dataset to indicate whether each data point contains a confusing signal. This will help the model learn to distinguish and identify confusing data during training.

[0076] Model design: Design deep learning models suitable for large models such as LoRA, including but not limited to convolutional neural networks (CNN), recurrent neural networks (RNN) or long short-term memory networks (LSTM). These deep learning models can effectively process time series data and learn to identify and process confusing signals during training.

[0077] Model training: The deep learning model is trained using a training dataset containing obfuscated data. During the training process, the deep learning model will learn to recognize obfuscated signals and improve its performance in the presence of obfuscated signals.

[0078] Model validation: Use a validation set (including obfuscated data) to validate the deep learning model to evaluate its performance in real-world scenarios. If the deep learning model performs poorly on the validation set, it can be optimized by adjusting the model structure, optimizing the algorithm, etc.

[0079] Model deployment: After the model is trained and validated, it is deployed to actual application scenarios. Since the deep learning model has adapted to and recognized obfuscated data, it should be more robust and accurate in actual use.

[0080] By directly including confusion in the training phase, the modified model can be better adapted to the confused signals in actual application scenarios, thereby improving the performance and reliability of the system. However, it should be noted that in actual applications, the model structure, training data, etc. may need to be adjusted according to specific needs and scenarios to achieve the best performance.

[0081] In a specific embodiment, permission control is set in the application interface of the correction model. By setting a corresponding permission control mechanism, it is ensured that the obfuscated rich media content is only corrected to look like a real person in a specific scenario, for example, only specific users or devices are allowed to access the correction function, or security measures such as double authentication are required to unlock the correction function.

[0082] The feature obfuscation processing method for AI-generated content of the present invention can be used as a standardized process. After AI generates original rich media content, it is obfuscated to generate obfuscated rich media content. When identifying the rich media content to be detected, a recognition algorithm or a deep learning model is used to identify whether a feature identifier is embedded in the rich media content to be detected, so as to determine whether the rich media content to be detected is the obfuscated rich media content generated by AI, wherein the recognition algorithm includes but is not limited to template matching and feature point matching.

[0083] The present invention proposes a technical solution for performing feature obfuscation processing on AI-generated rich media content, by randomly perturbing or partially deforming the features extracted from the original rich media content, and then fusing the perturbed features with the original rich media content through a generative adversarial network, or using an unsupervised learning method to encode the locally deformed features and embed the encoded features into the original rich media content as feature identifiers, thereby generating obfuscated rich media content. The present invention performs tiny, difficult-to-recognize obfuscation processing on the features of the original rich media content, so that the obfuscated rich media content can maintain a high degree of visual fidelity, making it difficult for the human eye to recognize the difference between the rich media content before and after the obfuscation, but the obfuscated rich media content generated by AI can be recognized through a recognition algorithm or a deep learning model, thereby solving the problem of distinguishing the obfuscated rich media content generated by AI from the original rich media content. The present invention can be widely used in fields such as identity authentication, copyright protection, and network security, and improves the security and reliability of these fields.

[0084] In addition, this application also proposes a feature obfuscation processing device for AI-generated content. Figure 4 This is a schematic diagram of a feature obfuscation processing device for AI-generated content according to the present invention, see Figure 4 The device includes: a feature extraction unit 401, which is used to extract features from the original rich media content generated by AI through a convolutional neural network and a deep learning image processing algorithm; an obfuscation processing unit 402, which is used to perform obfuscation processing on the extracted features using an obfuscation algorithm; a feature embedding unit 403, which is used to embed the obfuscated features into the original rich media content to generate obfuscated rich media content, and the features are used to identify the obfuscated rich media content generated by AI. All relevant contents of each step involved in the above method embodiment can be referred to the functional description of the corresponding functional module, and will not be repeated here.

[0085] In other embodiments of the present application, the present application discloses an electronic device, Figure 5 FIG. 1 is a schematic diagram of the structure of the electronic device of the present invention, see Figure 5, the electronic device may include: one or more processors 501; memory 502; display 503; one or more applications (not shown); and one or more computer programs 504, and the above components may be connected via one or more communication buses 505. The one or more computer programs 504 are stored in the above memory 502 and configured to be executed by the one or more processors 501, and the one or more computer programs 504 include instructions, which can be used to execute the following: Figure 1 and Figure 4 And each step in the corresponding embodiment.

[0086] Through the description of the above implementation methods, technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device and unit described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0087] Each functional unit in each embodiment of the present application can be integrated into a processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The above integrated unit can be implemented in the form of hardware or in the form of software functional units.

[0088] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as flash memory, mobile hard disk, read-only memory, random access memory, disk or optical disk.

[0089] Although the embodiments of the present invention are described in detail above, it is obvious to those skilled in the art that various modifications and variations can be made to these embodiments. However, it should be understood that such modifications and variations are within the scope and spirit of the present invention as described in the claims. Moreover, the present invention described herein may have other embodiments and may be implemented or realized in a variety of ways.

Claims

1. A feature obfuscation processing method for AI-generated content, characterized in that: The method comprises: Feature extraction of original rich media content generated by AI through convolutional neural networks and deep learning image processing algorithms; The extracted features are obfuscated using an obfuscation algorithm; The obfuscated features are embedded into the original rich media content to generate obfuscated rich media content, and the features are used to identify the obfuscated rich media content generated by AI.

2. The method according to claim 1, characterized in that After the extracted features are obfuscated using the obfuscation algorithm, it also includes: Encode the obfuscated features; The obfuscated features are embedded into the original rich media content to generate obfuscated rich media content, including: The encoded features are embedded into the original rich media content to generate obfuscated rich media content.

3. The method according to claim 1 or 2, characterized in that: The extracted features are obfuscated using an obfuscation algorithm, including: The extracted features are randomly disturbed or locally deformed using an obfuscation algorithm, wherein the random disturbance includes but is not limited to adding random noise and specific patterns.

4. The method according to claim 1 or 2, characterized in that: Before extracting features from the original rich media content generated by AI through convolutional neural networks and deep learning image processing algorithms, it also includes: Establishing a correction model, wherein the correction model is used to correct the obfuscated rich media content; A modified adversarial network is established, wherein the modified adversarial network is used to perform adversarial generation on the modified model.

5. A feature obfuscation processing device for AI-generated content, characterized in that: The device comprises: A feature extraction unit is used to extract features from the original rich media content generated by AI through convolutional neural networks and deep learning image processing algorithms; An obfuscation processing unit, used for performing obfuscation processing on the extracted features using an obfuscation algorithm; A feature embedding unit is used to embed the obfuscated features into the original rich media content to generate obfuscated rich media content, wherein the features are used to identify the obfuscated rich media content generated by AI.

6. The device according to claim 5, characterized in that Also includes: An encoding unit, used for performing obfuscation processing on the extracted features using an obfuscation algorithm, and then encoding the obfuscated features; The feature embedding unit embeds the obfuscated features into the original rich media content to generate obfuscated rich media content, and is specifically used to: embed the encoded features into the original rich media content to generate obfuscated rich media content.

7. The device according to claim 5 or 6, characterized in that The obfuscation processing unit performs obfuscation processing on the extracted features using an obfuscation algorithm, and is specifically used to: perform random perturbation or local deformation on the extracted features using an obfuscation algorithm, wherein the random perturbation includes but is not limited to adding random noise and specific patterns.

8. The device according to claim 5 or 6, characterized in that Also includes: A correction model building unit, used to build a correction model, wherein the correction model is used to correct the obfuscated rich media content; The modified adversarial network establishing unit is used to establish a modified adversarial network, and the modified adversarial network is used to perform adversarial generation on the modified model.

9. An electronic device, characterized in that: include: A processor and a memory, wherein the memory is used to store a computer program; The processor is configured to execute the computer program stored in the memory so as to enable the electronic device to perform the method according to any one of claims 1 to 4.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 4 is implemented.