Meta-learning based artificial intelligence model hierarchical authorization control method and system

Abstract

The application discloses a meta-learning-based artificial intelligence model hierarchical authorization control method and system, relates to the technical field of artificial intelligence, and comprises the following steps: S1, constructing a multi-level generation task flow, defining three authorization levels, matching corresponding potential vector processing modes and image generation quality requirements for each level, and clearly defining the core target of multi-task training; S2, performing a meta-learning training process through a double-layer training strategy of inner loop-outer loop; S3, adaptive quality control: periodically evaluating image quality based on FID scores and dynamically adjusting transformation parameters; and S4, performing hierarchical authorization reasoning for image generation requests of users and outputting images corresponding to different authorization levels. The meta-learning-based artificial intelligence model hierarchical authorization control method and system provided by the application solve the problem that the prior art cannot realize safe, dynamic and multi-level image distribution through a single generation model.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a hierarchical authorization control method and system for artificial intelligence models based on meta-learning. Background Technology

[0002] In the current field of AIGC (Artificial Intelligence Content Generation), especially in image generation technology, service providers typically need to offer content of different qualities or privacy levels based on users' access levels (e.g., guests, regular users, VIP users). There are currently two main approaches: The first is post-processing at the pixel level, where the model first generates a uniformly high-quality image, and then performs post-processing at the pixel level, such as Gaussian blurring, adding mosaic effects, adding visible watermarks, or cropping the resolution to generate a low-quality version for users with lower access levels. The second approach is independent deployment of multiple models, where models of different precision are trained and deployed separately for different quality requirements (e.g., a large model generates a high-resolution image, while a small model generates a preview image).

[0003] However, existing technologies still have many shortcomings. First, they lack security. Pixel-level post-processing (such as blurring or mosaicking) often destroys high-frequency information in images but preserves structural information. Using existing super-resolution or denoising AI models, attackers can easily restore low-quality images to high-quality images close to the original, causing graded authorization to fail. Second, they have high resource consumption. Deploying multiple models will multiply the storage space and GPU memory usage. While post-processing solutions only require one model, they require additional computational steps to process the image and need to store multiple copies of different qualities, increasing storage costs. Third, they lack intrinsic model control. Existing models are usually optimized only for a single distribution (i.e., high-quality real images) and cannot dynamically output multi-level quality images that conform to statistical laws by simply changing the input conditions under a single model weight. Fourth, quality control is rigid. Existing degradation methods are usually based on fixed parameters, such as a fixed blur radius, and cannot dynamically adjust the degradation intensity according to the actual quality of the generated content, which can easily lead to "over-degradation leading to unusability" or "under-degradation leading to leakage." Summary of the Invention

[0004] The purpose of this invention is to provide a hierarchical authorization control method and system for artificial intelligence models based on meta-learning, which solves the problem that existing technologies cannot achieve secure, dynamic, and multi-level image distribution through a single generative model.

[0005] To achieve the above objectives, this invention provides a hierarchical authorization control method for artificial intelligence models based on meta-learning, comprising the following steps: S1. Construct a multi-level generation task flow, define 3 authorization levels, match the corresponding potential vector processing method and generated image quality requirements for each level, and clarify the core objective of multi-task training. S2. Meta-learning training process is carried out through a two-layer training strategy of inner loop-outer loop; S3. Adaptive Quality Control: Based on FID score, periodically evaluate image quality and dynamically adjust transformation parameters; S4. For the user's image generation request, perform hierarchical authorization reasoning and output the images corresponding to different authorization levels.

[0006] Preferably, in S1: the three authorization levels include the highest level, the intermediate level, and the lowest level; the highest level corresponds to the untransformed initial latent vector, generating a high-quality original image; the intermediate level corresponds to the initial latent vector after linear mapping transformation, generating an image with specific watermark features or retaining structure but with damaged details; the lowest level corresponds to the annotated latent vector after strong noise perturbation, generating a preview image that retains contours and is full of noise.

[0007] Preferably, in S2: S21, obtain image data to get a training set, extract real image samples from the training set, and at the beginning of each iteration of meta-learning, perform parameter synchronization operation to enable the temporary model... Inherit global model All weights; S22. Execute the inner loop operation to generate images of different authorization levels in parallel. S23. Perform loss calculation; S24. Perform the outer loop operation to aggregate parameters.

[0008] Preferably, in S22: S221, from Initial latent vectors are generated in the 3D latent space through pseudo-random sampling. The expression is: ; ; In the formula, Represents the mean vector; Represents the unit covariance matrix; Indicates the first 1 initial potential vector, Indicates the index of the initial latent vector. ; N Indicates the sign of the normal distribution; Indicates the transpose symbol; Represents the set of real numbers; S222. The initial potential vectors are transformed through transformation paths corresponding to different authorization levels to obtain a set of transformation vectors corresponding to different authorization levels; For the highest level of transformation path, the initial latent vector is not processed in any way, preserving the most original feature distribution, as expressed in: ; In the formula, This represents the transformation vector obtained through the transformation path corresponding to the highest level. Represents the identity transformation function; The transformation path corresponding to the intermediate level maps the initial latent vector to a specific subspace through a linear transformation layer, expressed as: ; In the formula, This represents the transformation vector obtained through the transformation path corresponding to the intermediate level. Represents the transformation weight matrix; Represents the bias vector; For the lowest level of transformation path, random noise or nonlinear perturbation is injected into the initial latent vector, expressed as: ; In the formula, This represents the transformation vector obtained through the transformation path corresponding to the lowest level. This represents the injected random noise vector or nonlinear perturbation term; This indicates the intensity of the control disturbance under injected random noise or nonlinear disturbance conditions; S223. Input the transformation vector obtained in S222 into the generator simultaneously to obtain three images with different license levels in parallel, including the original image, the watermarked image, and the noise image, which correspond to the highest, middle, and lowest license levels, respectively.

[0009] Preferably, in S23: S231. The discriminator scores images of different authorization levels, and the discriminator compresses the scoring results using the Sigmoid activation function. Within the range, if the scoring result is in Within the interval, the discriminator considers the image to conform to the distribution of the high-quality original image; the scoring result is within... If the image quality is within the specified range, the discriminator will consider it to be of poor quality or a mismatch in features. S232. Calculate the comprehensive loss function, and the comprehensive loss... Including discriminator loss and generator loss ; The formula for calculating the discriminator loss is: ; In the formula, Represents a real image; Indicates a label; Represents the generated images at various levels; Indicates the discriminator network function; Indicates batch size; Indicates the index of the sample in the batch. ; The expression for calculating the generator loss is: ; In the formula, Indicates the number of authorization levels; Indicates the level of authority index. ; Represents the weighting coefficients for each level; Indicates the first The loss function for each authorization level satisfies: ; In the formula, Represents the mathematical expectation operator; Indicates the first Potential vector transformation path functions corresponding to each authorization level; S233. Based on the comprehensive loss, the temporary model is processed using the stochastic gradient descent method. Trainable parameters in Updated in rotation.

[0010] Preferably, in S24: the temporary model is completed. After the round parameters are updated, the outer loop aggregation is triggered to calculate the difference between the original parameters of the temporary model and the global model. Update the global model parameters along the direction of difference, as expressed by: ; ; In the formula, Represents the original parameter vector of the global model; This represents the parameter vector of the temporary model after the inner loop completes the preset number of updates; This represents the parameter difference vector, indicating the direction of model evolution on a specific task; This represents the outer loop learning rate; This represents the updated global model parameter vector.

[0011] Preferably, in S3: S31. Calculate the FID score for different authorization levels under the current outer loop iteration count, using the following expression: ; In the formula, A mean vector representing the distribution of features in a real image; This represents the mean vector of the generated image feature distribution; The covariance matrix representing the distribution of features in a real image; The covariance matrix represents the distribution of features in the generated image; Represents the difference between the means vectors The square of the norm; The trace of a matrix is ​​the sum of the elements on the main diagonal of the square matrix. S32, Set safety threshold If in a certain authorization level This indicates that the image quality is too good, posing a risk of exceeding the weight limit and triggering the inversion mechanism. The inversion mechanism forcibly reduces the generated image quality by increasing the perturbation intensity or adding a penalty term to the loss function. If the disturbance intensity is increased, let the transformation function parameters for this authorization level be... Then the expression for increasing the perturbation strength is: ; In the formula, This represents the updated disturbance intensity parameter; Indicates the step adjustment factor; Indicates the safety threshold; Indicates the current FID score; By adding a penalty term to the loss function, the generator loss function for this authorization level is adjusted as follows: ; In the formula, This represents the punitive loss function after reversal; Penalty coefficient; Represents the binary cross-entropy loss function; This indicates the score given by the discriminator to the image generated for this authorization level; This represents the label of the actual sample.

[0012] Preferably, in S4: S41. Identity Authentication: Verify the user's identity category, including VIP users, regular users, and guests; S42. Select the path based on the user's identity category; If you are a VIP user, select the path corresponding to the highest level to generate a high-definition image; if you are a regular user, select the path corresponding to the middle level to generate an image with noise or a watermark; if you are a guest, select the path corresponding to the lowest level to generate a low-quality preview image. S43. Output the images corresponding to different authorization levels.

[0013] This invention also provides a system for a hierarchical authorization control method for artificial intelligence models based on meta-learning, including a latent vector transformation module, a generator network module, a discriminator network module, a meta-learning control module, and a quality assessment and feedback module. The latent vector transformation module, located at the front end of the generator, receives the original random noise vector and performs data transformations along different paths according to the requested authorization level. The generator network module employs a deep convolutional neural network architecture, receives the latent vectors output from the latent vector transformation module, and maps the latent vectors to RGB images. The discriminator network module uses a classification network with spectral normalization technology to distinguish between real and generated samples and calculates the distribution distance, providing gradient information for generator optimization. The meta-learning control module controls the internal and external loops of the training process, finding the optimal initialization point of parameters by simulating generation tasks at different authorization levels. The quality assessment and feedback module extracts features based on the Inception network, calculates the FID index of the generated image and the real image, and dynamically adjusts the parameters or training strategy of the latent vector transformation module based on the FID index value.

[0014] Therefore, the present invention employs the above-mentioned hierarchical authorization control method and system based on meta-learning artificial intelligence models, which has the following beneficial effects: (1) Source security control and strong anti-reverse restoration capability: abandoning traditional pixel-level post-processing, mathematical transformation is performed on the noise vector in the latent space, so that the low-authority image forms a fundamental difference in the feature layer from the source of generation, which cannot be reversed by super-resolution, denoising and other models. At the same time, implicit watermarks can be introduced in the intermediate level to realize copyright traceability, completely solve the problem of leakage of hierarchical authorization, and greatly improve content security. (2) Lightweight deployment of single model, with dual optimization of resources and efficiency: Only one generator network needs to be trained and stored, reducing the number of parameters by 66.7% compared to the traditional multi-model scheme, reducing hardware overhead such as storage and GPU memory, and eliminating the need to store multiple quality image copies; a single forward propagation can output multiple levels of images in parallel, eliminating complex encryption and decryption and post-processing steps, reducing the inference speed from seconds to milliseconds, improving it by an order of magnitude, and adapting to the needs of real-time and high-concurrency applications; (3) Multi-task compatibility and stability, and controllable generation quality: Based on the training strategy of inner-outer loop of meta-learning, the model parameters converge to the intermediate state that adapts to high-quality and limited-quality generation, avoiding the model collapse and extremely poor quality of single layer in traditional training, and ensuring stable output of a single model under different authorization levels; at the same time, a dynamic feedback closed loop based on the FID index is constructed to monitor the image quality of limited layer in real time, automatically adjust the perturbation intensity or loss function, accurately control the quality difference of each layer, and avoid authorization failure caused by excessive or insufficient degradation; (4) Simplified process and wide scene adaptability: Images of different authorization levels are generated directly from the potential space without post-pixel processing steps, simplifying the generation process and reducing additional computational overhead. The quality of each level of image is designed according to the permission, retaining the practical value of scenarios such as preview, normal viewing, and high-definition use. The three-level parallel authorization implementation logic makes content distribution more flexible and can be adapted to multiple scenarios such as digital collection display, medical data sharing, and paid image distribution.

[0015] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0016] Figure 1 This is an overall flowchart of the hierarchical authorization control method for artificial intelligence models based on meta-learning of the present invention; Figure 2 This is a logical diagram of the hierarchical authorization control method for artificial intelligence models based on meta-learning according to the present invention; Figure 3 The figure shows the experimental results of an embodiment of the present invention; where (a) represents VIP users, (b) represents ordinary users, and (c) represents tourists. Detailed Implementation

[0017] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0018] Please see Figure 1-2 A hierarchical authorization control method for artificial intelligence models based on meta-learning includes the following steps: S1. Construct a multi-level generation task flow, define 3 authorization levels, match the corresponding latent vector processing method and generated image quality requirements for each level, and clarify the core objective of multi-task training; the 3 authorization levels include the highest level, the intermediate level and the lowest level; the highest level corresponds to the untransformed initial latent vector, generating a high-quality original image; the intermediate level corresponds to the initial latent vector after linear mapping transformation, generating an image with specific watermark features or retaining structure but with damaged details; the lowest level corresponds to the annotated latent vector after strong noise perturbation, generating a preview image that only retains the outline and is full of noise.

[0019] S2. Meta-learning training process is carried out through a two-layer training strategy of inner loop-outer loop.

[0020] S21. Obtain image data to get the training set, extract real image samples from the training set, and perform parameter synchronization operations at the beginning of each iteration of meta-learning to make the temporary model... Inherit global model All weights.

[0021] S22. Execute the inner loop operation to generate images of different authorization levels in parallel.

[0022] S221, from Initial latent vectors are generated in the 3D latent space through pseudo-random sampling. The expression is: ; ; In the formula, Represents the mean vector; Represents the unit covariance matrix; Indicates the first 1 initial potential vector, Indicates the index of the initial latent vector. ; Indicates the sign of the normal distribution; Indicates the transpose symbol; It represents the set of real numbers.

[0023] S222. The initial potential vectors are transformed through transformation paths corresponding to different authorization levels to obtain a set of transformation vectors corresponding to different authorization levels; For the highest level of transformation path, the initial latent vector is not processed in any way, preserving the most original feature distribution, as expressed in: ; In the formula, This represents the transformation vector obtained through the transformation path corresponding to the highest level. Represents the identity transformation function; The intermediate-level transformation path maps the initial latent vector to a specific subspace through a linear transformation layer. This transformation introduces implicit watermark features or a slight texture bias into the generated image, enabling the addition of copyright information while preserving most of the visual information. The expression is: ; In the formula, This represents the transformation vector obtained through the transformation path corresponding to the intermediate level. Represents the transformation weight matrix; Represents the bias vector; For the lowest level of transformation path, high-intensity random noise or nonlinear perturbation is injected into the initial latent vector. This random perturbation disrupts the high-frequency feature distribution of the initial latent vector, causing the generator to only be able to reproduce the coarse outline and tone of the image, unable to generate clear details. The expression is: ; In the formula, This represents the transformation vector obtained through the transformation path corresponding to the lowest level. This represents the injected random noise vector or nonlinear perturbation term; This indicates the intensity of the control disturbance under injected random noise or nonlinear disturbance conditions.

[0024] S223. Input the transformation vector obtained in S222 into the generator simultaneously to obtain three images with different license levels in parallel, including the original image, the watermarked image, and the noise image, which correspond to the highest, middle, and lowest license levels, respectively.

[0025] S23. Perform loss calculation; S231. Images of different authorization levels are scored by a discriminator. The discriminator compresses the scoring results using a Sigmoid activation function. Within the range, if the scoring result is in Within the specified interval, the discriminator considers the image to have extremely high realism, consistent with the distribution of high-quality original images; the scoring result is within... If the image quality is within the specified range, the discriminator will consider it to be of poor quality or a mismatch in features. S232. Calculate the comprehensive loss function. The comprehensive loss function considers both image realism and the discrimination between different levels. Including discriminator loss and generator loss ; The formula for calculating the discriminator loss is: ; In the formula, Represents a real image; Indicates a label; Represents the generated images at various levels; Indicates the discriminator network function; Indicates batch size; Indicates the index of the sample in the batch. ; The expression for calculating the generator loss is: ; In the formula, Indicates the number of authorization levels; Indicates the level of authority index. ; Represents the weighting coefficients for each level; Indicates the first The loss function for each authorization level satisfies: ; In the formula, Represents the mathematical expectation operator; Indicates the first Potential vector transformation path functions corresponding to each authorization level; S233. Based on the comprehensive loss, the temporary model is processed using the stochastic gradient descent method. All trainable parameters of the transposed convolutional kernel, linear mapping layer, and batch normalization layer in the process are performed. Updated in rotation.

[0026] S24. Execute the outer loop operation to aggregate parameters; the temporary model is complete. After the round parameters are updated, the outer loop aggregation is triggered to calculate the difference between the original parameters of the temporary model and the global model. Update the global model parameters along the direction of difference, enabling the model to "learn" how to handle generation tasks with different quality requirements simultaneously. The expression is: ; ; In the formula, Represents the original parameter vector of the global model; This represents the parameter vector of the temporary model after the inner loop completes the preset number of updates; This represents the parameter difference vector, indicating the direction of model evolution on a specific task; This represents the outer loop learning rate; This represents the updated global model parameter vector.

[0027] S3. Adaptive Quality Control: Based on FID scores, image quality is periodically evaluated, and transformation parameters are dynamically adjusted.

[0028] S31. Calculate the FID score for different authorization levels under the current outer loop iteration count, using the following expression: ; In the formula, A mean vector representing the distribution of features in a real image; This represents the mean vector of the generated image feature distribution; The covariance matrix representing the distribution of features in a real image; The covariance matrix represents the distribution of features in the generated image; Represents the difference between the means vectors The square of the norm; The trace of a matrix is ​​the sum of the elements on the main diagonal of the square matrix.

[0029] S32, Set safety threshold If in a certain authorization level This indicates that the image quality is too good, too close to the real image, and there is a risk of exceeding the weight limit. The system automatically triggers the inversion mechanism. The inversion mechanism forcibly reduces the quality of the generated image by increasing the perturbation intensity or adding a penalty term to the loss function. If the disturbance intensity is increased, let the transformation function parameters for this authorization level be... Then the expression for increasing the perturbation strength is: ; In the formula, This represents the updated disturbance intensity parameter; This represents the step adjustment factor (sensitivity coefficient). Indicates the safety threshold; Indicates the current FID score; By adding a penalty term to the loss function, the generator loss function for this authorization level is adjusted as follows: ; In the formula, This represents the punitive loss function after reversal; Indicates the penalty coefficient; Represents the binary cross-entropy loss function; This indicates the score given by the discriminator to the image generated for this authorization level; This represents the label of the actual sample.

[0030] S4. Entering the practical application stage, hierarchical authorization reasoning is performed based on the user's image generation request.

[0031] S41. Identity Authentication: Verify the user's identity category, including VIP users, regular users, and guests.

[0032] S42. Select a path based on the user's identity category; if the user is a VIP user, select the path corresponding to the highest level to generate a high-definition image; if the user is a regular user, select the path corresponding to the middle level to generate an image with noise or a watermark; if the user is a guest, select the path corresponding to the lowest level to generate a low-quality preview image.

[0033] S43. Output the images corresponding to different authorization levels.

[0034] The system of hierarchical authorization control method for artificial intelligence models based on meta-learning includes a latent vector transformation module, a generative network module, a discriminative network module, a meta-learning control module, and a quality assessment and feedback module.

[0035] The latent vector transformation module, located at the front end of the generator, receives the original random noise vector and performs data transformations along different paths based on the requested authorization level. These paths include three levels: Level 1, Level 2, and Level 3. In the Level 1 path, the random noise vector is not processed and is directly passed through. In the Level 2 path, a specific linear mapping or watermarking encoding transformation is applied to the random noise vector, giving the generated image implicit features or slight degradation. In the Level 3 path, Poisson noise or other strong perturbation transformations are applied to the random noise vector, resulting in an image with severe noise or blurring, intended only for previewing content outlines.

[0036] The generative network module employs a deep convolutional neural network architecture, receiving latent vectors output from the latent vector transformation module and mapping these latent vectors to RGB images. Due to differences in the distribution of input vectors, the same model will output images of different visual qualities.

[0037] The discriminant network module employs a classification network with spectral normalization technology to distinguish between real and generated samples and calculates the distribution distance (Wasserstein distance) to provide gradient information for generator optimization.

[0038] The meta-learning control module is used to control the internal and external loops of the training process. By simulating generation tasks of different authorization levels, it finds the optimal initialization point of parameters.

[0039] The quality assessment and feedback module extracts features based on the Inception network and calculates the FID index of the generated image and the real image; it dynamically adjusts the parameters or training strategy of the latent vector transformation module based on the FID index value.

[0040] Experimental setup and verification Dataset: The Bedroom subset of the LSUN (Large-scale Scene Understanding) dataset was used. Approximately 200,000 bedroom images were randomly selected from the dataset and uniformly cropped to 64×64 pixels.

[0041] Network architecture: The generator uses a deep convolutional network (ConvGenerator), and the discriminator uses a spectral normalization discriminator network (SNDiscriminator); latent space dimension .

[0042] Training strategy: The Reptile meta-learning algorithm is used, with the inner loop updating parameters 5 times and the outer loop aggregating parameters; an adaptive quality feedback mechanism (FID threshold) is also enabled. ).

[0043] Please refer to the visualization results. Figure 3The results show that the same model generates images of different qualities in parallel based on the input authorization level. VIP users receive images with clear details and high realism; ordinary users receive images with implicit watermark textures (traceable); and tourists receive images that only retain the outline and tone of the bedroom, with severe loss of detail, achieving the purpose of previewing.

[0044] Regarding storage efficiency, traditional methods require training and deploying a separate model for each license level (such as a high-definition model, a watermark model, or a blur model), resulting in a total number of parameters that is three times that of a single model. This invention requires only one model, reducing the number of parameters by 66.7% and significantly lowering storage and deployment costs.

[0045] Regarding inference speed, compared with traditional attribute-based encryption (ABE) schemes, this invention does not require complex bilinear pairing decryption operations, but only one neural network forward propagation, thus improving inference speed by an order of magnitude (from seconds to milliseconds).

[0046] Regarding the security comparison, the lowest-level image was input into a super-resolution model (such as SRGAN) to attempt to recover high-definition details. The anti-attack capabilities of pixel-level mosaic and the latent spatial transformation of the present invention were compared. Pixel-level mosaic (blurred) could be partially recovered, while the present invention (strong noise path) could not be recovered, indicating that it has a stronger anti-reverse reconstruction capability.

[0047] Therefore, this invention adopts the above-mentioned meta-learning-based artificial intelligence model hierarchical authorization control method and system. By performing vector transformation from the source of the latent space, it forms fundamental differences in the image feature layer, which has strong anti-reverse reconstruction capability and can also introduce implicit watermarks to achieve copyright traceability. Relying on meta-learning, a single model can be adapted to multi-level generation tasks to avoid model collapse. Combined with the FID index dynamic feedback closed loop, it can accurately control the image quality of each level and prevent hierarchical authorization failure. The solution only requires a single model to output multiple parallel paths, which significantly reduces storage and deployment costs, improves inference speed by an order of magnitude, and eliminates post-pixel processing, simplifying the process. The three-level parallel authorization logic makes distribution more flexible and can be adapted to multiple scenarios such as digital collections and medical data sharing.

[0048] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A hierarchical authorization control method for artificial intelligence models based on meta-learning, characterized in that, Includes the following steps: S1. Construct a multi-level generation task flow, define 3 authorization levels, match the corresponding potential vector processing method and generated image quality requirements for each level, and clarify the core objective of multi-task training. S2. Meta-learning training process is carried out through a two-layer training strategy of inner loop-outer loop; S3. Adaptive Quality Control: Based on FID score, periodically evaluate image quality and dynamically adjust transformation parameters; S4. For the user's image generation request, perform hierarchical authorization reasoning and output the images corresponding to different authorization levels.

2. The hierarchical authorization control method for artificial intelligence models based on meta-learning according to claim 1, characterized in that, In S1: there are three authorization levels, including the highest level, intermediate level, and lowest level; The highest level corresponds to the untransformed initial latent vector, generating a high-quality original image; The intermediate level corresponds to the initial latent vector after linear mapping transformation, generating images with specific watermark features or images that retain structure but are damaged in detail; The lowest level corresponds to the annotated latent vector after being perturbed by strong noise, generating a preview image that preserves the contours and is full of noise.

3. The hierarchical authorization control method for artificial intelligence models based on meta-learning according to claim 2, characterized in that, In S2: S21. Obtain image data to get the training set, extract real image samples from the training set, and perform parameter synchronization operations at the beginning of each iteration of meta-learning to make the temporary model... Inherit global model All weights; S22. Execute the inner loop operation to generate images of different authorization levels in parallel. S23. Perform loss calculation; S24. Perform the outer loop operation to aggregate parameters.

4. The hierarchical authorization control method for artificial intelligence models based on meta-learning according to claim 3, characterized in that, In S22: S221, from Initial latent vectors are generated in the 3D latent space through pseudo-random sampling. The expression is: ； ； In the formula, Represents the mean vector; Represents the unit covariance matrix; Indicates the first 1 initial potential vector, Indicates the index of the initial latent vector. ; Indicates the sign of the normal distribution; Indicates the transpose symbol; Represents the set of real numbers; S222. The initial potential vectors are transformed through transformation paths corresponding to different authorization levels to obtain a set of transformation vectors corresponding to different authorization levels; For the highest level of transformation path, the initial latent vector is not processed in any way, preserving the most original feature distribution, as expressed in: ； In the formula, This represents the transformation vector obtained through the transformation path corresponding to the highest level. Represents the identity transformation function; The transformation path corresponding to the intermediate level maps the initial latent vector to a specific subspace through a linear transformation layer, expressed as: ； In the formula, This represents the transformation vector obtained through the transformation path corresponding to the intermediate level. Represents the transformation weight matrix; Represents the bias vector; For the lowest level of transformation path, random noise or nonlinear perturbation is injected into the initial latent vector, expressed as: ； In the formula, This represents the transformation vector obtained through the transformation path corresponding to the lowest level. This represents the injected random noise vector or nonlinear perturbation term; This indicates the intensity of the control disturbance under injected random noise or nonlinear disturbance conditions; S223. Input the transformation vector obtained in S222 into the generator simultaneously to obtain three images with different license levels in parallel, including the original image, the watermarked image, and the noise image, which correspond to the highest, middle, and lowest license levels, respectively.

5. The hierarchical authorization control method for artificial intelligence models based on meta-learning according to claim 4, characterized in that, In S23: S231. The discriminator scores images of different authorization levels, and the discriminator compresses the scoring results using the Sigmoid activation function. Within the range, if the scoring result is in Within the interval, the discriminator considers the image to conform to the distribution of the high-quality original image; the scoring result is within... If the image quality is within the specified range, the discriminator will consider it to be of poor quality or a mismatch in features. S232. Calculate the comprehensive loss function, and the comprehensive loss... Including discriminator loss and generator loss ; The formula for calculating the discriminator loss is: ； In the formula, Represents a real image; Indicates a label; Represents the generated images at various levels; Indicates the discriminator network function; Indicates batch size; Indicates the index of the sample in the batch. ; The expression for calculating the generator loss is: ； In the formula, Indicates the number of authorization levels; Indicates the level of authority index. ; Represents the weighting coefficients for each level; Indicates the first The loss function for each authorization level satisfies: ； In the formula, Represents the mathematical expectation operator; Indicates the first Potential vector transformation path functions corresponding to each authorization level; S233. Based on the comprehensive loss, the temporary model is processed using the stochastic gradient descent method. Trainable parameters in Updated in rotation.

6. The hierarchical authorization control method for artificial intelligence models based on meta-learning according to claim 5, characterized in that, In S24: Temporary model completed. After the round parameters are updated, the outer loop aggregation is triggered to calculate the difference between the original parameters of the temporary model and the global model. Update the global model parameters along the direction of difference, as expressed by: ； ； In the formula, Represents the original parameter vector of the global model; This represents the parameter vector of the temporary model after the inner loop completes the preset number of updates; This represents the parameter difference vector, indicating the direction of model evolution on a specific task; This represents the outer loop learning rate; This represents the updated global model parameter vector.

7. The hierarchical authorization control method for artificial intelligence models based on meta-learning according to claim 6, characterized in that, In S3: S31. Calculate the FID score for different authorization levels under the current outer loop iteration count, using the following expression: ； In the formula, A mean vector representing the distribution of features in a real image; This represents the mean vector of the generated image feature distribution; The covariance matrix representing the distribution of features in a real image; The covariance matrix represents the distribution of features in the generated image; Represents the difference between the means vectors The square of the norm; The trace of a matrix is ​​the sum of the elements on the main diagonal of the square matrix. S32, Set safety threshold If in a certain authorization level This indicates that the image quality is too good, posing a risk of exceeding the weight limit and triggering the inversion mechanism. The inversion mechanism forcibly reduces the generated image quality by increasing the perturbation intensity or adding a penalty term to the loss function. If the disturbance intensity is increased, let the transformation function parameters for this authorization level be... Then the expression for increasing the perturbation strength is: ； In the formula, This represents the updated disturbance intensity parameter; Indicates the step adjustment factor; Indicates the safety threshold; Indicates the current FID score; By adding a penalty term to the loss function, the generator loss function for this authorization level is adjusted as follows: ； In the formula, This represents the punitive loss function after reversal; Indicates the penalty coefficient; Represents the binary cross-entropy loss function; This indicates the score given by the discriminator to the image generated for this authorization level; This represents the label of the actual sample.

8. The hierarchical authorization control method for artificial intelligence models based on meta-learning according to claim 7, characterized in that, In S4: S41. Identity Authentication: Verify the user's identity category, including VIP users, regular users, and guests; S42. Select the path based on the user's identity category; If you are a VIP user, select the path corresponding to the highest level to generate a high-definition image; For regular users, selecting the path corresponding to the middle level will generate an image with noise or a watermark; for guests, selecting the path corresponding to the lowest level will generate a low-quality preview image. S43. Output the images corresponding to different authorization levels.

9. A system for a hierarchical authorization control method for artificial intelligence models based on meta-learning as described in any one of claims 1-8, characterized in that, include: The latent vector transformation module, located at the front end of the generator, is used to receive the original random noise vector and perform data transformation for different paths according to the requested authorization level. The generative network module adopts a deep convolutional neural network architecture, receives the latent vectors output from the latent vector transformation module, and maps the latent vectors to RGB images; The discriminant network module employs a classification network with spectral normalization technology to distinguish between real samples and generated samples, and calculates the distribution distance to provide gradient information for generator optimization. The meta-learning control module is used to control the internal and external loops of the training process. By simulating generation tasks of different authorization levels, it finds the optimal initialization point of parameters. The quality assessment and feedback module extracts features based on the Inception network and calculates the FID index of the generated image and the real image; it dynamically adjusts the parameters or training strategy of the latent vector transformation module based on the FID index value.

Images