A Multi-Task Classification Method and System for Remote Sensing Images Based on Lightweight Hypernetworks

Abstract

This invention discloses a multi-task classification method and system for remote sensing images based on a lightweight hypernetwork. The trained lightweight hypernetwork generates compact conditional vectors for each task and preprocesses the image to obtain a standard tensor. Based on the conditional vectors, a task-adaptive convolutional kernel parameter set is dynamically generated. This parameter set is used to extract task-specific feature maps from the standard tensor. Adversarial training is performed through an embedded gradient inversion layer before classification is completed. This invention achieves feature separation for each kernel and task through a compact latent space, effectively solving gradient conflict and negative transfer problems with only a 3.2% increase in parameters. Deep collaboration between multi-task feature extraction and cross-domain adversarial training enhances the model's generalization ability to multi-source remote sensing data. An optimization framework combining KL distillation and dynamic weight scheduling ensures training stability and efficiency. This invention provides a high-precision, lightweight solution for resource-constrained edge computing scenarios.

Description

Technical Field

[0001] This application belongs to the field of image classification technology, specifically relating to a multi-task classification method and system for remote sensing images based on a lightweight hypernetwork. Background Technology

[0002] With the rapid development of remote sensing technology, the scale of multi-source Earth observation data, including optical, SAR, and hyperspectral data, continues to grow, driving the evolution of remote sensing image interpretation from single-task to multi-task collaborative processing. In resource-constrained edge scenarios such as satellite on-orbit analysis and UAV real-time monitoring, how to utilize a single model to simultaneously complete multiple tasks such as ground feature classification, scene recognition, and target detection has become crucial for improving the intelligence and efficiency of remote sensing systems.

[0003] Currently, multi-task learning of remote sensing images mainly relies on hard parameter sharing architectures, where multiple tasks share the same backbone network and static convolutional kernels. However, this method struggles to adapt to differences in feature distributions across different tasks, easily leading to gradient conflicts and negative transfer during training, thus limiting the overall model performance. Although some studies have introduced task-specific generators or multi-branch structures to achieve feature decoupling, these often result in a significant increase in parameter count and computational complexity, making it difficult to meet the deployment constraints of edge devices. Furthermore, existing methods generally lack effective domain adaptation mechanisms for multi-source, multi-sensor remote sensing data, causing a decline in the model's generalization ability when faced with data from different imaging conditions, resolutions, and bands. At the optimization level, most methods use fixed weights or simple heuristic strategies to balance multi-task losses, failing to dynamically adjust weights according to task learning progress, often resulting in the dominant task suppressing secondary tasks.

[0004] Therefore, there is an urgent need for a remote sensing image multi-task classification method that balances lightweight design, multi-task adaptability, cross-domain robustness, and training dynamics, so as to achieve efficient and reliable multi-task collaborative processing under limited resources. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this application provides a method and system for multi-task classification of remote sensing images based on a lightweight hypernetwork. The technical problem to be solved by this application is achieved through the following technical solution: A multi-task classification method for remote sensing images based on lightweight hypernetworks includes: S100 receives multiple classification tasks and the remote sensing images to be classified corresponding to each classification task; S200, execute each classification task by inputting the remote sensing image corresponding to each classification task into the trained lightweight supernetwork, so that the lightweight supernetwork can classify the remote sensing image to obtain a classification result. The process of the lightweight supernetwork classifying the remote sensing image is as follows: generate a conditional vector for each classification task and preprocess the corresponding remote sensing image to obtain a standard tensor; generate a convolution kernel parameter set based on the conditional vector; extract task-specific feature maps from the standard tensor based on the convolution kernel and parameter set; remove interference information related to the sensor domain from the task-specific feature map through gradient inversion adversarial training to obtain anti-interference features; and perform classification based on the anti-interference features to obtain a classification result.

[0006] A remote sensing image multi-task classification system based on a lightweight hypernetwork includes: The receiving module is configured to receive multiple classification tasks and the remote sensing images to be classified corresponding to each classification task. The execution module is configured to perform each classification task, inputting the remote sensing image corresponding to each task into a trained lightweight supernetwork. The lightweight supernetwork then classifies the remote sensing image to obtain a classification result. The classification process of the lightweight supernetwork is as follows: a conditional vector is generated for each classification task, and the corresponding remote sensing image is preprocessed to obtain a standard tensor; a convolutional kernel parameter set is generated based on the conditional vector; task-specific feature maps are extracted from the standard tensor based on the convolutional kernel and parameter set; anti-interference features are obtained by removing sensor-domain related interference information from the task-specific feature maps through gradient inversion adversarial training; and classification is performed based on the anti-interference features to obtain the classification result. Beneficial effects: 1. This application proposes a task-adaptive convolution kernel generation mechanism with a compact latent space, which achieves accurate task feature separation with extremely low parameter overhead (3.2%) and solves the gradient conflict problem in multi-task learning.

[0007] 2. This application realizes deep collaborative optimization of multi-task feature extraction and cross-domain adversarial processing. By embedding progressive gradient inversion layers, it effectively aligns the distribution of multi-source remote sensing data and improves the model's generalization ability.

[0008] 3. This application designs a KL distillation-enhanced multi-task dynamic optimization framework, which combines adaptive weight scheduling and mixed precision training to ensure the stability and efficiency of multi-task learning.

[0009] 4. This application constructs a system-level lightweight solution that comprehensively considers resource constraints from algorithm design to deployment optimization, providing practical technology for edge computing scenarios.

[0010] The present application will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of a remote sensing image multi-task classification method based on a lightweight hypernetwork provided in this application; Figure 2 This is a schematic diagram of the lightweight hypernetwork provided in this application; Figure 3 This is a schematic diagram of the forward and backward propagation processes provided in this application; Figure 4 This is the training flowchart provided in this application. Detailed Implementation

[0012] The present application will be described in further detail below with reference to specific embodiments, but the implementation of the present application is not limited thereto.

[0013] like Figure 1 As shown, this application provides a multi-task classification method for remote sensing images based on a lightweight hypernetwork, including: S100 receives multiple classification tasks and the remote sensing images to be classified corresponding to each classification task; S200, execute each classification task by inputting the remote sensing image corresponding to each classification task into the trained lightweight supernetwork, so that the lightweight supernetwork can classify the remote sensing image to obtain a classification result. The process of the lightweight supernetwork classifying the remote sensing image is as follows: generate a conditional vector for each classification task and preprocess the corresponding remote sensing image to obtain a standard tensor; generate a convolution kernel parameter set based on the conditional vector; extract task-specific feature maps from the standard tensor based on the convolution kernel and parameter set; remove interference information related to the sensor domain from the task-specific feature map through gradient inversion adversarial training to obtain anti-interference features; and perform classification based on the anti-interference features to obtain a classification result.

[0014] refer to Figure 2The lightweight supernetwork includes an intelligent task encoding module, a lightweight supernetwork module, a task-adaptive feature extraction network, a domain adversarial alignment module, and a multi-task classification head. The output of the intelligent task encoding module is connected to the inputs of the lightweight supernetwork module and the task-adaptive feature extraction network, respectively. The outputs of the lightweight supernetwork module and the task-adaptive feature extraction network are connected to the domain adversarial alignment module, and the output of the domain adversarial alignment module is connected to the input of the multi-task classification head. The intelligent task encoding module generates a conditional vector for each classification task and preprocesses the corresponding input image to obtain a standard tensor. The lightweight supernetwork module generates a convolutional kernel parameter set based on the conditional vector. The task-adaptive feature extraction network extracts task-specific feature maps from the standard tensor based on the convolutional kernels and the parameter set. The domain adversarial alignment module uses gradient inversion adversarial training to remove sensor-domain-related interference information from the task-specific feature maps to obtain anti-interference features. The multi-task classification head performs classification based on the anti-interference features to obtain the classification result.

[0015] The task-adaptive feature extraction network is used to apply the convolution and parameter set to a standard tensor, and then batch processing is performed using grouped convolution to obtain a task-specific feature map; the domain adversarial alignment module performs gradient inversion adversarial training through a gradient inversion layer to remove interference information related to the sensor domain from the task-specific feature map to obtain anti-interference features.

[0016] The intelligent task encoding module of this application adopts a unified optical / SAR / hyperspectral processing pipeline, which can achieve differentiated data augmentation based on task characteristics and finally output a 128-dimensional conditional vector. The lightweight hypernetwork module of this application uses a channel dimension multiplexing matrix to achieve parameter sharing and employs dynamic kernel calculation compatible with grouped convolution, supporting dynamic task addition / removal. The domain adversarial alignment module of this application is deeply integrated with residual block feature extraction, adopting progressive adversarial scheduling, with λ(t) linearly increasing from 0.2 to 0.9, combined with a multi-granularity feature alignment strategy, namely, shallow texture and deep semantic alignment.

[0017] This invention discloses a multi-task classification method for remote sensing images based on a lightweight hypernetwork, belonging to the fields of computer vision and deep learning. It achieves precise feature separation for each task by dynamically generating task-adaptive 3×3 convolutional kernels using compact 128-dimensional latent vectors. It innovatively integrates gradient inversion layers with multi-task feature extraction to achieve cross-sensor domain alignment. An optimization strategy combining an integrated Adam optimizer, KL distillation regularization, and dynamic task weight scheduling balances multi-task learning conflicts. By integrating mixed-precision training, gradient clipping, and test-time augmentation techniques, it achieves an average accuracy of 94.16% with only a 3.2% increase in parameters, representing a 3.715% improvement over traditional methods and a 7.1% reduction in gradient conflicts. This invention is particularly suitable for resource-constrained remote sensing image processing scenarios, providing an efficient solution for satellite on-orbit processing and real-time UAV monitoring.

[0018] In one specific embodiment of this application, the training process of the lightweight hypernetwork includes: a. Obtain the training set, which includes multiple training samples, each training sample corresponding to a remote sensing sample image, the true category of the remote sensing sample image, and the identification task identifier; b. For each training sample, the corresponding recognition task identifier and remote sensing sample image are input into a preset lightweight supernetwork, so that the lightweight supernetwork generates a training condition vector and the corresponding remote sensing sample image for preprocessing to obtain a standard tensor; a convolution kernel parameter set is generated based on the training condition vector; a task-specific feature map is extracted from the standard tensor based on the convolution and parameter set; through gradient inversion adversarial training, interference information related to the sensor domain in the task-specific feature map is removed to obtain anti-interference features; classification is performed based on the anti-interference features to obtain the predicted classification result for each task; Combination Figure 3 and Figure 4 The training process of this application is divided into four stages: stage 1 data preprocessing and embedding stage, stage 2 feature generation, stage 3 training optimization, and stage 4 testing and evaluation.

[0019] In Phase 1 of this application, the intelligent task encoding module maps the task identifier of each task to a 128-dimensional condition vector through an embedding layer. .

[0020] In stage 2, the feature generation input is a 128-dimensional conditional vector. The lightweight hypernetwork transforms this conditional vector into a 3×3 convolutional kernel parameter set through a two-stage linear transformation. Based on the convolutions and parameter set, task-specific feature maps are extracted from the standard tensor. The two-stage linear transformation is expressed as:

[0021]

[0022] In the formula, , , These represent the number of output channels and the number of input channels, respectively, and k=3 represents the filter size. Indicates the vector Remodeling The convolution kernel vector, It is the intermediate feature vector after the first linear transformation and ReLU activation. It is the weight matrix of the first-level linear transformation. It is the weight matrix of the second-level linear transformation, and It is the bias vector associated with that layer; In stage 3, the network hyperparameters are input, the total loss function is calculated, and backpropagation training is performed. To further improve cross-domain robustness, this application integrates a gradient reversal layer after the final residual block, connected to a multi-task classification head for adversarial training. The forward and backward propagation formulas for the gradient reversal layer are as follows:

[0023]

[0024] in The adaptive coefficient increases linearly from 0.2 to 0.9. Represents the identity matrix.

[0025] c. For each recognition task, the total loss function is calculated using the predicted classification result and the true category corresponding to the recognition task; the total loss function includes the classification loss, domain classification loss and regularization loss for each task. f, in the direction that makes the total loss function decrease, adjust the parameters of the lightweight supernetwork using a dynamic task weight adjustment strategy until the training cutoff condition is met, and obtain the trained lightweight supernetwork.

[0026] The dynamic task weight adjustment strategy of this application is expressed as follows:

[0027] In the formula, Indicates the first i The baseline for calculating the weight of each task. This indicates the maximum extent to which the weights will be adjusted based on task performance. Indicates the first i The current classification accuracy for each task. This represents the maximum current classification accuracy.

[0028] This application employs the AdamW optimizer, with the learning rate scheduled using a one-cycle policy. Mixed-precision training and gradient pruning (threshold = 1.0) are implemented during training to ensure stability. The weights corresponding to tasks 1 and 2 are... , How it changes with the number of training steps t:

[0029] In the formula, t represents the current training step number. This indicates the total number of training steps.

[0030] In stage 3, this application introduces a gradient stability enhancement module during the training process to train the ultra-lightweight network through adaptive gradient clipping, mixed precision training, and gradient noise injection; wherein, the threshold set by the adaptive gradient clipping is dynamically adjusted according to the exponential moving average of the gradient norm.

[0031] This application uses a gradient norm-based exponential moving average to dynamically adjust the pruning threshold. Mixed-precision training optimization uses FP16 precision for forward propagation and FP32 precision for gradient accumulation, while setting a loss scaling factor. Gradient noise injection regularization is represented as adding small-amplitude Gaussian noise during the later stages of training. N(0, ²), ² = 0.01·(1-t / T), enhancing the robustness of the model; This application employs a multi-objective optimization strategy combining KL distillation regularization and dynamic task weight scheduling. The total loss function is then expressed as: ； In the formula, Represents the total loss function. Indicates task weight, Indicates the first i Classification loss for each task, Indicates the first i KL distillation regularization loss for each task Represents the coefficient. Representation domain classification loss, This indicates hyperparameters that are greater than 0. This represents the loss due to regularization and knowledge transfer. With training steps change.

[0032] The intelligent task encoding module of this application encodes multi-task requirements into a compact 128-dimensional latent space representation, supporting dynamic task expansion. The lightweight supernetwork module uses a two-level linear transformation network with shared channel dimension parameters to generate task-adaptive convolutional kernels with a parameter cost of 3.2%. The domain adversarial alignment module integrates adaptive gradient clipping, mixed-precision training, and gradient noise injection to achieve three-level gradient alignment and conflict mitigation at the task layer, network layer, and optimizer layer. This application employs a strategy of dynamically adjusting task weights and optimization based on learning progress.

[0033] This application employs KL distillation for knowledge transfer between tasks. A dynamic weight scheduling strategy is used for learning progress-aware weight allocation. Mixed precision and gradient pruning are employed to ensure stable training. Finally, the TTA enhancement pipeline can perform four geometric transformations, and the model integrates multi-checkpoint weight averaging and a high-reliability prediction screening mechanism for confidence calibration, which can improve inference accuracy.

[0034] In Phase 4, this application conducts testing and evaluation. During the testing phase, two strategies are employed to improve performance: (1) applying horizontal flip, vertical flip, and 90-degree rotation to the input image and averaging the prediction results of each enhancement version; (2) loading multiple training checkpoints and integrating the prediction results of each model. This application supports test-time enhancement, multi-model ensemble, and high-confidence sample analysis. During inference, test-time enhancement and model ensemble strategies are employed to improve classification accuracy.

[0035] This application also provides a remote sensing image multi-task classification system based on a lightweight hypernetwork, including: The receiving module is configured to receive multiple classification tasks and the remote sensing images to be classified corresponding to each classification task. The execution module is configured to perform each classification task by inputting the remote sensing image corresponding to each classification task into a trained lightweight supernetwork. The lightweight supernetwork then classifies the remote sensing image to obtain a classification result. The classification process of the lightweight supernetwork is as follows: a conditional vector is generated for each classification task, and the corresponding remote sensing image is preprocessed to obtain a standard tensor; a convolutional kernel parameter set is generated based on the conditional vector; a task-specific feature map is extracted from the standard tensor based on the convolutional kernel and parameter set; anti-interference features are obtained by stripping sensor-domain-related interference information from the task-specific feature map through gradient inversion adversarial training; and classification is performed based on the anti-interference features to obtain the classification result.

[0036] This application can be applied to real-time on-orbit processing of satellites in resource-constrained scenarios, requiring simultaneous cloud detection, ground feature classification, and change detection. A single model accomplishes multiple tasks, reducing data downlink volume by 60%.

[0037] This application can also be applied to autonomous drone monitoring, with the scenario requirement being real-time farmland monitoring (crop classification, pest and disease detection, and growth assessment). It can also be applied to disaster emergency response, with the scenario requirement being rapid disaster assessment (flood extent, damaged buildings, and road accessibility). The implementation results show that a multi-dimensional disaster assessment can be completed within 10 minutes; in actual participation in three flood disaster monitoring operations, the assessment accuracy rate was >90%.

[0038] This application proposes a task-adaptive convolutional kernel generation mechanism with a compact latent space, achieving accurate task feature separation with extremely low parameter overhead (3.2%), and resolving the gradient conflict problem in multi-task learning. It implements deep collaborative optimization of multi-task feature extraction and cross-domain adversarial learning, effectively aligning the distribution of multi-source remote sensing data and improving model generalization ability through progressive gradient inversion layer embedding. A KL distillation-enhanced multi-task dynamic optimization framework is designed, combining adaptive weight scheduling and mixed-precision training to ensure the stability and efficiency of multi-task learning. A system-level lightweight solution is constructed, comprehensively considering resource constraints from algorithm design to deployment optimization, providing practical technology for edge computing scenarios. A complete trusted inference enhancement mechanism is provided, improving reliability in practical applications through test-time enhancement, model ensemble, and confidence calibration. It also includes an S100 that receives multiple classification tasks and the corresponding remote sensing images to be classified for each task.

[0039] It is worth noting that the terms "first" and "second" in this application are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0040] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of this application and should not be construed as limiting the specific implementation of this application to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of this application, and all such modifications or substitutions should be considered within the scope of protection of this application.

Claims

1. A multi-task classification method for remote sensing images based on lightweight hypernetworks, characterized in that, include: S100 receives multiple classification tasks and the remote sensing images to be classified corresponding to each classification task; S200, execute each classification task, input the remote sensing image corresponding to each classification task into the trained lightweight supernetwork, so that the lightweight supernetwork can classify the remote sensing image to obtain the classification result. The process of the lightweight supernetwork classifying the remote sensing image is as follows: generate a conditional vector for each classification task and preprocess the corresponding remote sensing image to obtain a standard tensor. A convolution kernel parameter set is generated based on the conditional vector; a task-specific feature map is extracted from the standard tensor based on the convolution and parameter set; and anti-interference features are obtained by stripping sensor domain-related interference information from the task-specific feature map through gradient inversion adversarial training. The classification result is obtained based on the aforementioned anti-interference features.

2. The remote sensing image multi-task classification method based on lightweight hypernetworks according to claim 1, characterized in that, The training process of the lightweight hypernetwork includes: a. Obtain the training set, which includes multiple training samples, each training sample corresponding to a remote sensing sample image, the true category of the remote sensing sample image, and the identification task identifier; b. For each training sample, the corresponding recognition task identifier and remote sensing sample image are input into a preset lightweight supernetwork, so that the lightweight supernetwork generates a training condition vector and the corresponding remote sensing sample image for preprocessing to obtain a standard tensor; a convolution kernel parameter set is generated based on the training condition vector; a task-specific feature map is extracted from the standard tensor based on the convolution and parameter set; through gradient inversion adversarial training, interference information related to the sensor domain in the task-specific feature map is removed to obtain anti-interference features; classification is performed based on the anti-interference features to obtain the predicted classification result for each task; c. For each recognition task, calculate the total loss function using the predicted classification result and the true category corresponding to the recognition task; f, in the direction that makes the total loss function decrease, adjust the parameters of the lightweight supernetwork using a dynamic task weight adjustment strategy until the training cutoff condition is met, and obtain the trained lightweight supernetwork.

3. The remote sensing image multi-task classification method based on lightweight hypernetworks according to claim 2, characterized in that, The total loss function includes the classification loss for each task, the domain classification loss, and the regularization loss; the total loss function is expressed as: ； In the formula, Represents the total loss function. Indicates task weight, Indicates the first i Classification loss for each task, Indicates the first i KL distillation regularization loss for each task Represents the coefficient. Representation domain classification loss, This indicates hyperparameters that are greater than 0. This represents the loss due to regularization and knowledge transfer. With training steps change.

4. The remote sensing image multi-task classification method based on lightweight hypernetworks according to claim 2, characterized in that, The dynamic task weight adjustment strategy is expressed as follows: In the formula, Indicates the first i The baseline for calculating the weight of each task. This indicates the maximum extent to which the weights will be adjusted based on task performance. Indicates the first i The current classification accuracy for each task. This represents the maximum current classification accuracy.

5. The remote sensing image multi-task classification method based on lightweight hypernetworks according to claim 2, characterized in that, A gradient stability enhancement module is introduced during training to train the ultra-lightweight network through adaptive gradient clipping, mixed-precision training, and gradient noise injection. The threshold set by the adaptive gradient clipping is dynamically adjusted according to the exponential moving average of the gradient norm.

6. The remote sensing image multi-task classification method based on lightweight hypernetworks according to claim 1 or 2, characterized in that, The lightweight hypernetwork includes an intelligent task encoding module, a lightweight hypernetwork module, a task-adaptive feature extraction network, a domain adversarial alignment module, and a multi-task classification head. The output of the intelligent task encoding module is connected to the inputs of the lightweight hypernetwork module and the task-adaptive feature extraction network, respectively. The outputs of the lightweight hypernetwork module and the task-adaptive feature extraction network are connected to the domain adversarial alignment module, and the output of the domain adversarial alignment module is connected to the input of the multi-task classification head.

7. The remote sensing image multi-task classification method based on lightweight hypernetworks according to claim 3, characterized in that, The intelligent task encoding module generates a conditional vector for each classification task and preprocesses the corresponding input image to obtain a standard tensor; the lightweight supernetwork module generates a set of convolutional kernel parameters based on the conditional vector; the task adaptive feature extraction network extracts task-specific feature maps from the standard tensor based on the convolution and parameter set; and the domain adversarial alignment module removes interference information related to the sensor domain from the task-specific feature map to obtain anti-interference features through gradient inversion adversarial training. The multi-task classification head is used to perform classification based on the anti-interference features to obtain classification results.

8. The remote sensing image multi-task classification method based on lightweight hypernetworks according to claim 7, characterized in that, The intelligent task encoding module maps the task identifier of each task to a 128-dimensional condition vector through the embedding layer, and the lightweight supernetwork converts the condition vector into a 3×3 convolution kernel parameter set through two-level linear transformation.

9. The remote sensing image multi-task classification method based on lightweight hypernetworks according to claim 7, characterized in that, The two-level linear transformation is expressed as: In the formula, , , These represent the number of output channels and the number of input channels, respectively, and k=3 represents the filter size. Indicates the vector Remodeling The convolution kernel vector, It is the intermediate feature vector after the first linear transformation and ReLU activation. It is the weight matrix of the first-level linear transformation. It is the weight matrix of the second-level linear transformation, and It is the bias vector associated with that layer; The task-adaptive feature extraction network is used to apply the convolution and parameter set to a standard tensor, and then batch processing is performed using grouped convolution to obtain a task-specific feature map; the domain adversarial alignment module performs gradient inversion adversarial training through a gradient inversion layer to remove interference information related to the sensor domain from the task-specific feature map to obtain anti-interference features.

10. A remote sensing image multi-task classification system based on a lightweight hypernetwork, characterized in that, include: The receiving module is configured to receive multiple classification tasks and the remote sensing images to be classified corresponding to each classification task. The execution module is configured to execute each classification task by inputting the remote sensing image corresponding to each classification task into the trained lightweight supernetwork, so that the lightweight supernetwork can classify the remote sensing image to obtain a classification result. The process of the lightweight supernetwork classifying the remote sensing image is as follows: generating a conditional vector for each classification task and preprocessing the corresponding remote sensing image to obtain a standard tensor. A convolution kernel parameter set is generated based on the conditional vector; a task-specific feature map is extracted from the standard tensor based on the convolution and parameter set; and anti-interference features are obtained by stripping sensor domain-related interference information from the task-specific feature map through gradient inversion adversarial training. The classification result is obtained based on the aforementioned anti-interference features.

Images