Electromagnetic signal forgetting learning method based on mask distillation erasing

Abstract

The invention discloses an electromagnetic signal forgetting learning method based on mask distillation erasing, and belongs to the field of electromagnetic signal processing. According to the method, the problems of information leakage and low model performance retention degree caused by disastrous forgetting and incomplete forgetting when an existing electromagnetic signal recognition model forgets a specific category are solved, and according to the technical scheme, a pre-training model and an original data set are given and a to-be-forgotten data set and a to-be-reserved data set are divided; the method comprises the following steps: constructing a prototype perceptual playback set after extracting feature distribution by using a model, calculating distillation erasure loss and prototype perceptual playback loss, combining the loss into a total loss function, updating model parameters through back propagation until convergence, and realizing efficient, thorough and low-performance-damage knowledge erasure. The method is mainly applied to signal identification tasks in electromagnetic spectrum monitoring, wireless communication security and electronic countermeasure.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of electromagnetic signal processing, and particularly relates to an electromagnetic signal forgetting learning method based on mask distillation erasing. BACKGROUND

[0002] With the deep integration of artificial intelligence technology in the field of wireless communication, electromagnetic signal recognition technology based on machine learning plays an increasingly important role in key applications such as electromagnetic spectrum monitoring, wireless communication security and electronic countermeasures. By training complex neural network models on large-scale electromagnetic signal datasets, high-precision and automated recognition of signal modulation types, radiation source fingerprints and other features can be achieved. However, this data-driven technology path, while bringing significant performance advantages, also raises serious data security and compliance challenges. During the training process, deep learning models deeply memorize and internalize the key features of training samples in their vast parameters, which makes it insufficient to simply delete the original data samples from physical storage when responding to data deletion requests, because the model itself has become a potential carrier of sensitive information leakage. The openness and adversarial nature of the electromagnetic environment make it impossible to completely remove the impact of deleted data by simply deleting training samples or roughly updating the model.

[0003] To solve the above problems, the most direct technical solution in the field is to completely retrain the model after removing part of the training data. However, for deep signal recognition models with over one hundred million parameters and TB-level training datasets, the computational resource cost required for retraining is extremely high, and the time-consuming is long, which cannot meet the rapid response requirements of business systems to data deletion requests. The forgetting learning scheme provides a new technical path for electromagnetic signal recognition, aiming to eliminate the impact of specific data on the trained model at a much lower cost than retraining. However, there are few related studies on electromagnetic signals at present, and there are a series of technical challenges such as isochronous non-stationary signal characteristics, catastrophic forgetting phenomenon, and difficulty in balancing model accuracy and forgetting efficiency. Specifically, when trying to erase the knowledge of specific categories of signals, the model is prone to lose the recognition ability of the remaining categories; and simple parameter adjustment cannot ensure the complete removal of target knowledge, and there is a risk of residual information leakage. Therefore, the existing technology lacks an electromagnetic signal forgetting learning method that can accurately locate and completely erase the impact of target data, while balancing high efficiency, low cost and high security. SUMMARY

[0004] To solve the above technical problems, the present application provides an electromagnetic signal forgetting learning method based on mask distillation erasing to solve the problems existing in the prior art.

[0005] In a first aspect, to achieve the above object, the present application provides an electromagnetic signal forgetting learning method based on mask distillation erasing, comprising the following steps:

[0006] Given a pre-trained electromagnetic signal recognition model, an original training dataset and a class set, the original training dataset is divided into a to-be-forgotten dataset and a reserved dataset according to the requirement of the forgetting task;

[0007] Features of all classes are extracted by using the pre-trained model;

[0008] Based on the extracted feature distribution, a prototype perception playback set is constructed through unsupervised clustering analysis and sample selection;

[0009] A distillation erasure loss and a prototype perception playback loss are calculated, wherein the distillation erasure loss is used to guide the model to erase the knowledge of the to-be-forgotten classes, and the prototype perception playback loss is used to consolidate the knowledge of the reserved classes;

[0010] The distillation erasure loss and the prototype perception playback loss are combined into a total loss function, and the model parameters are updated through a back propagation algorithm;

[0011] The model parameters are repeatedly updated until the model converges, and a model with completed target class knowledge forgetting is obtained.

[0012] Optionally, the process of dividing the original training dataset into the to-be-forgotten dataset and the reserved dataset includes that a union of the to-be-forgotten dataset and the reserved dataset is equal to the original training dataset.

[0013] Optionally, the process of extracting features of all classes by using the pre-trained model includes that for each class, a feature extraction network of the pre-trained model is used to process all samples of the class to obtain a feature vector set of the class.

[0014] Optionally, the process of constructing the prototype perception playback set includes:

[0015] For each class in the reserved dataset, a feature vector set of all samples of the class is obtained;

[0016] Unsupervised clustering analysis is performed on the feature vector set, the K-means algorithm is used to divide it into a preset number K of clusters, and a center vector of each cluster is calculated;

[0017] For each center vector, a sample with a minimum Euclidean distance from the center vector is found in an original input space; all found samples are collected to form the prototype perception playback set.

[0018] Optionally, the process of calculating the distillation erasure loss includes:

[0019] A masking vector is used to shield logical values of the to-be-forgotten classes in the model output;

[0020] calculating a KL divergence between a current model output and an ideal model output, wherein the ideal model output represents a uniform distribution or a zero-knowledge state for the to-forget classes;

[0021] The process of calculating the prototype-aware replay loss includes calculating a cross-entropy loss using the prototype-aware replay set.

[0022] Optionally, the process of combining the distillation erasure loss and the prototype-aware replay loss into a total loss function includes linearly weighting the distillation erasure loss and the prototype-aware replay loss.

[0023] The process of updating the model parameters by the backpropagation algorithm includes updating the model parameters using a gradient descent optimizer with a model learning rate.

[0024] In a second aspect, the present application further provides a mask-distillation erasure based electromagnetic signal forgetting learning system for implementing a mask-distillation erasure based electromagnetic signal forgetting learning method, the system comprising:

[0025] A data division module is configured to divide an original training data set into a to-forget data set and a reserved data set according to forgetting task requirements.

[0026] A feature extraction module is configured to extract feature distributions of all classes using a pre-trained model.

[0027] A replay set construction module is configured to construct a prototype-aware replay set based on the extracted feature distributions through unsupervised clustering analysis and sample selection.

[0028] A loss calculation module is configured to calculate a distillation erasure loss and a prototype-aware replay loss.

[0029] A parameter updating module is configured to combine the distillation erasure loss and the prototype-aware replay loss into a total loss function, and update model parameters by a backpropagation algorithm until the model converges.

[0030] In a third aspect, the present application further provides a computer terminal device, comprising:

[0031] One or more processors;

[0032] A memory coupled to the processor, configured to store one or more programs;

[0033] When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the mask-distillation erasure based electromagnetic signal forgetting learning method of the first aspect described above.

[0034] In a fourth aspect, the present application also provides a computer readable storage medium having stored thereon a computer program, which, when executed by a processor, implements the steps of the method for electromagnetic signal forgetting learning based on mask distillation erasing in the first aspect.

[0035] In a fifth aspect, the present application also provides a computer program product comprising a computer program which, when executed by a processor, implements the steps of the method for electromagnetic signal forgetting learning based on mask distillation erasing in the first aspect.

[0036] Compared with the prior art, the present application has the following advantages and technical effects:

[0037] The method for electromagnetic signal forgetting learning based on mask distillation erasing provided by the present application can effectively suppress the catastrophic forgetting problem by constructing a prototype perception replay set and introducing a replay loss in the forgetting training, and ensure that the recognition performance of the model on non-target categories is maintained at a high level after completing knowledge erasing. Combined with the knowledge erasing mechanism based on mask distillation, the present application can guide the model to approach a zero-knowledge state for the forgotten categories, realize deep and complete erasing of target knowledge, significantly improve data security and effectively resist privacy attacks. Through the weighted balance of the prototype perception replay loss and the mask distillation erasing loss, the present application provides a flexible and controllable joint optimization framework, which can adjust the trade-off between forgetting strength and retention priority according to actual needs. The entire forgetting process does not need to completely retrain the entire data, but only needs to operate on a small-scale forgetting set and a replay set, which has the advantages of low computational overhead and significantly reduced time cost, and maximizes the application value of the model while ensuring data compliance and security. BRIEF DESCRIPTION OF DRAWINGS

[0038] The accompanying drawings, which form a part of this application, are intended to provide further understanding of the application and are incorporated herein in their entirety. The schematic embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0039] Figure 1 The machine forgetting learning flowchart of the embodiment of the present application;

[0040] Figure 2 The flowchart of the embodiment of the present application;

[0041] Figure 3 The different method forgetting performance comparison diagram of the embodiment of the present application. DETAILED DESCRIPTION

[0042] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0043] It is noted that the steps shown in the flowcharts of the drawings can be performed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowcharts, in some cases the steps shown or described can be performed in an order different from that shown here.

[0044] Embodiment One

[0045] The embodiment provides an electromagnetic signal forgetting learning method based on mask distillation erasing, comprising:

[0046] Given a pre-trained electromagnetic signal recognition model, an original training data set and a category set, the original training data set is divided into a forgetting data set and a reserved data set according to the forgetting task requirement;

[0047] The pre-trained model is used to extract the feature distribution of all categories;

[0048] Based on the extracted feature distribution, an original prototype perception playback set is constructed through unsupervised clustering analysis and sample selection;

[0049] The distillation erasing loss and the prototype perception playback loss are calculated, wherein the distillation erasing loss is used to guide the model to erase the knowledge of the forgetting category, and the prototype perception playback loss is used to consolidate the knowledge of the reserved category;

[0050] The distillation erasing loss and the prototype perception playback loss are combined into a total loss function, and the model parameters are updated through a back propagation algorithm;

[0051] The model parameters are repeatedly updated until the model converges, and a model with completed target category knowledge forgetting is obtained.

[0052] As an optional implementation manner in the embodiment, the specific steps of the above process include:

[0053] Step one, given a pre-trained electromagnetic signal recognition model , and an original complete training data set used to train the model and a category set . According to the forgetting task requirement, the original data set is divided into a forgetting data set and a reserved data set .

[0054] Step two, the pre-trained model is used to extract and represent the feature distribution of all categories.

[0055] Step three, based on the features extracted in step two, an original prototype perception playback set is constructed through unsupervised clustering analysis and sample selection .

[0056] Step four, to guide the model for accurate knowledge erasing and preserving, respectively calculate distillation erasing loss and prototype-aware replay loss .

[0057] Step five, combine the loss function defined above, and update the model parameters through the back propagation algorithm until the forgetting target is reached.

[0058] Step six, repeat the parameter update process in step five until the model converges or reaches the preset number of training rounds. After training, the model parameters obtained are the final model that has completed the forgetting of the target class knowledge.

[0059] As an embodiment in this embodiment, the process of dividing the original training data set into a to-be-forgotten data set and a preserved data set includes:

[0060] The union of the to-be-forgotten data set and the preserved data set is equal to the original training data set.

[0061] As an optional embodiment in this embodiment, the above process corresponds to step one, and the relationship between the original complete training data set, the to-be-forgotten data set, and the to-be-preserved data set is as follows:

[0062] (1),

[0063] Formula (1), represents the original training data set, and represent the to-be-forgotten data set and the preserved data set, respectively.

[0064] As an embodiment in this embodiment, the process of extracting the feature distribution of all classes using a pre-trained model includes:

[0065] For each class, use the feature extraction network of the pre-trained model to process all samples of the class to obtain a set of feature vectors of the class.

[0066] As an optional embodiment in this embodiment, the above process corresponds to step two, and the process of extracting and representing the feature distribution of all classes using a pre-trained model :

[0067] (2),

[0068] (3),

[0069] In formulas (2) and (3), represents the feature extraction network defined by the parameters, represents a single input signal sample, represents a training sample set of category c, represents a set of feature vectors of all samples of category c.

[0070] As an embodiment in this embodiment, the process of constructing the prototype perception playback set includes:

[0071] For each category in the reserved data set, obtain the feature vector set of all samples of the category;

[0072] Perform unsupervised clustering analysis on the feature vector set, divide it into a preset number K clusters using the K-means algorithm, and calculate the center vector of each cluster;

[0073] For each of the center vectors, find the sample with the smallest Euclidean distance to the center vector in the original input space;

[0074] Collect all the found samples to form the prototype perception playback set.

[0075] As an optional embodiment in this embodiment, the above process corresponds to step three, based on the features extracted in step two, a small and representative playback set is constructed through clustering analysis and sample selection (selecting the original sample closest to the cluster center in the original input space) :

[0076] (4),

[0077] (5),

[0078] In formulas (4) and (5) represents performing K-means clustering algorithm on the feature set to divide it into K clusters and obtain the center vector (i.e. prototype) of each cluster, represents the L2 norm (Euclidean distance), represents the reserved category set, and K represents the number of clustering centers, the prototype perception playback set constructed.

[0079] As an embodiment in this embodiment, the process of calculating the distillation erasure loss includes:

[0080] Use the mask vector to shield the logical value of the to-be-forgotten category in the model output;

[0081] Calculate the KL divergence between the current model output and the ideal model output, wherein the ideal model output represents a uniform distribution or zero knowledge state for the to-be-forgotten category;

[0082] The process of calculating the prototype-aware replay loss comprises:

[0083] calculating a cross-entropy loss using the prototype-aware replay set.

[0084] As an optional implementation in the embodiment, the above process comprises in detail:

[0085] To guide the precise knowledge erasing and preserving of the model, respectively calculate the distillation erasing loss and the prototype-aware replay loss :

[0086] (6),

[0087] (7),

[0088] (8),

[0089] (9),

[0090] In formulas (6) and (7) and respectively represent the logits distribution of the ideal network model output and the logits distribution of the current network model output, is a temperature parameter, represents a mask vector, is a normalization function. In formulas (8) and (9) represents a KL divergence function, represents a cross-entropy loss function.

[0091] As an implementation in the embodiment, the process of combining the distillation erasing loss and the prototype-aware replay loss into a total loss function comprises:

[0092] linearly weighting the distillation erasing loss and the prototype-aware replay loss;

[0093] The process of updating the model parameters by the back propagation algorithm comprises:

[0094] updating the model parameters using a gradient descent optimizer with a model learning rate.

[0095] As an optional implementation in the embodiment, the above process corresponds to step five, linearly weights the distillation erasing loss and the prototype-aware replay loss to obtain a final total loss function, and updates the model parameters by back propagation.

[0096] (10),

[0097] (11),

[0098] Equations (10), (11) is a balance coefficient, which is a hyperparameter for balancing the importance of forgetting strength and knowledge retention, represents the model learning rate, represents a momentum-based gradient descent optimizer.

[0099] Finally, in step six, the parameter update process in step five is repeatedly performed until the model converges or reaches a preset number of training rounds, and the final base model parameters are .

[0100] Taking WiFi radiation source identification forgetting learning as an example, the training data contains 50 categories of WiFi radiation source signals (with different levels of noise added), and the test data only contains WiFi signal samples consistent with the semantic labels of the training set but different:

[0101] First, the training data set is used to train the network from scratch, and the cross-entropy loss function is minimized to obtain ;

[0102] Then, the training set is input into the model , steps two and three are performed to obtain the replay set and save it (no need to repeat in the future);

[0103] In the test phase, assuming that the user requests the model to forget multiple categories (such as 23, 7, 34, 15, and 42), the model performs distillation erasing operation based on the forgetting data set and the replay set , the detailed steps are shown in Figure 2 , and the results are shown in Figure 3 . Figure 3 The baseline (original model, retraining, and random label) and the method of the present application are respectively shown for WiFi radiation source identification multi-class forgetting performance, and the present application achieves performance close to retraining under the condition of greatly reducing the computational time complexity, wherein the forgetting set accuracy is 0%, and the retention set accuracy is 97.90%.

[0104] Based on this, the electromagnetic signal forgetting learning method based on mask distillation erasing provided by the embodiment of the present application has the following beneficial effects:

[0105] 1. The application realizes accurate and efficient consolidation of the core knowledge of the reserved categories by constructing a prototype-aware replay set and introducing a replay loss in the forgetting training. Through feature space clustering, the mechanism intelligently selects a few representative samples, simulates the "review" process of key knowledge during the forgetting process, effectively suppresses the "catastrophic forgetting" problem commonly existing in traditional forgetting methods, and ensures that the recognition performance of the model for non-target categories can be maintained at a high level after completing the knowledge erasure.

[0106] 2. The application combines the knowledge distillation mechanism based on logits mask and KL divergence constraint to guide the model to approach the zero knowledge state for the prediction of the forgotten categories, realizing the deep and thorough erasure of the target knowledge. Compared with simple parameter fine-tuning or gradient penalty method, the data security of the method is higher, which can effectively resist privacy attacks such as member reasoning. At the same time, since the whole process avoids complete retraining of the entire data, only small-scale forgetting set and replay set are operated, so the method has the advantages of high algorithm efficiency, low computational overhead and significantly reduced time cost.

[0107] 3. The core of the application is to propose a joint optimization strategy for electromagnetic signal recognition, which integrates knowledge reservation and erasure. Through the weighted balance of prototype-aware replay loss and mask distillation erasure loss, a flexible and controllable framework is provided for the model forgetting task. The design not only can adjust the forgetting strength and the trade-off between reservation priority according to the actual demand (for example, the strictness of regulatory compliance or the importance of business performance), but also can ensure the optimal overall performance of the model after forgetting, thereby maximizing the application value of the model while ensuring data compliance and security.

[0108] Embodiment two

[0109] In this embodiment, a computer terminal device is provided, comprising:

[0110] one or more processors;

[0111] a memory coupled to the processor, for storing one or more programs;

[0112] When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the above-mentioned mask distillation erasure-based electromagnetic signal forgetting learning method.

[0113] In this embodiment, a computer readable storage medium having a computer program stored thereon is also provided, and the computer program is executed by a processor to implement the steps of the above-mentioned mask distillation erasure-based electromagnetic signal forgetting learning method.

[0114] In the embodiment, an electronic device is also provided, comprising a memory and a processor, the memory stores a computer program, and the processor is configured to run the computer program to perform the steps of the above-mentioned mask distillation erasing-based electromagnetic signal forgetting learning method.

[0115] In the embodiment, a computer program product is also provided, comprising a computer program, which, when executed by a processor, implements the steps of the above-mentioned mask distillation erasing-based electromagnetic signal forgetting learning method.

[0116] The above-mentioned program can be run in a processor, or can also be stored in a memory (or a computer readable medium), which includes permanent and non-permanent, removable and non-removable media, and can be realized by any method or technology to store information. The information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic tape disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible by a computing device.

[0117] These computer programs can also be loaded into a computer or other programmable data processing device, so that a series of operation steps are performed on the computer or other programmable device to generate a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more flows Figure 1 The steps of the functions specified in one or more flows or one or more blocks Figure 1 The steps of the functions specified in one or more flows or one or more blocks

[0118] In the embodiment, such a device or system is provided. The system is called a mask distillation erasing-based electromagnetic signal forgetting learning system, which comprises:

[0119] A data division module is configured to divide the original training data set into a forgetting data set and a reserved data set according to the forgetting task requirements;

[0120] A feature extraction module is configured to extract the feature distribution of all categories by using a pre-trained model;

[0121] A playback set construction module is configured to construct a prototype perception playback set by unsupervised clustering analysis and sample selection based on the extracted feature distribution.

[0122] a loss calculation module configured to calculate a distillation erasure loss and a prototype-aware replay loss;

[0123] a parameter updating module configured to combine the distillation erasure loss and the prototype-aware replay loss into a total loss function, and update model parameters through a back propagation algorithm until the model converges.

[0124] As an implementation form of the embodiment, the data division module comprises a division verification unit configured to ensure that a union of the to-be-forgotten data set and the reserved data set is equal to the original training data set.

[0125] As an implementation form of the embodiment, the feature extraction module comprises a feature calculation unit configured to, for each category, process all samples of the category using a feature extraction network of the pre-trained model to obtain a feature vector set of the category.

[0126] As an implementation form of the embodiment, the replay set construction module comprises a clustering analysis unit configured to perform unsupervised clustering analysis on the feature vector set; a center vector calculation unit configured to calculate a center vector of each cluster; a sample selection unit configured to find, in an original input space, a sample having a minimum Euclidean distance from the center vector; and a replay set formation unit configured to collect all found samples to form the prototype-aware replay set.

[0127] As an implementation form of the embodiment, the loss calculation module comprises a mask processing unit configured to shield logical values of to-be-forgotten categories in model output using a mask vector; a distillation loss calculation unit configured to calculate a KL divergence between a current model output and an ideal model output; and a replay loss calculation unit configured to calculate a cross-entropy loss using the prototype-aware replay set.

[0128] As an implementation form of the embodiment, the parameter updating module comprises a loss combination unit configured to linearly weight the distillation erasure loss and the prototype-aware replay loss; and a gradient optimization unit configured to update model parameters using a gradient descent optimizer with a model learning rate.

[0129] The system or device is used to realize the functions of the method in the above-mentioned embodiments. Each module in the system or device corresponds to each step in the method, and has been described in the method and will not be described here.

[0130] Through the above-mentioned embodiments, the problem of electromagnetic signal forgetting learning based on mask distillation erasure in the related art is solved, so that the problem existing in the prior art can be solved.

[0131] The above merely describes the preferred embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for electromagnetic signal forgetting learning based on mask distillation erase, characterized in that, The method comprises the following steps: Given a pre-trained electromagnetic signal recognition model, an original training dataset and a category set, the original training dataset is divided into a to-be-forgotten dataset and a reserved dataset according to the requirement of a forgetting task; Features of all categories are extracted by using the pre-trained model; Based on the extracted features, a prototype-aware replay set is constructed through unsupervised clustering analysis and sample selection; A distillation erasure loss and a prototype-aware replay loss are calculated, wherein the distillation erasure loss is used to guide the model to erase the knowledge of the to-be-forgotten categories, and the prototype-aware replay loss is used to consolidate the knowledge of the reserved categories; The distillation erasure loss and the prototype-aware replay loss are combined into a total loss function, and the model parameters are updated through a back propagation algorithm; The model parameters are repeatedly updated until the model converges, and a model with completed target category knowledge forgetting is obtained.

2. The method of claim 1, wherein, The process of dividing the original training dataset into the to-be-forgotten dataset and the reserved dataset comprises: The union of the to-be-forgotten dataset and the reserved dataset is equal to the original training dataset.

3. The method of claim 1, wherein, The process of extracting features of all categories by using the pre-trained model comprises: For each category, all samples of the category are processed by using a feature extraction network of the pre-trained model to obtain a feature vector set of the category.

4. The method of claim 1, wherein, The process of constructing the prototype-aware replay set comprises: For each category in the reserved dataset, a feature vector set of all samples of the category is obtained; Unsupervised clustering analysis is performed on the feature vector set, the K-means algorithm is used to divide the feature vector set into a preset number K of clusters, and a center vector of each cluster is calculated; For each center vector, a sample with a minimum Euclidean distance from the center vector in the original input space is found; All found samples are collected to form the prototype-aware replay set.

5. The method of claim 1, wherein, The process of calculating the distillation erasure loss comprises: A mask vector is used to shield logical values of the to-be-forgotten categories in the model output; A KL divergence between a current model output and an ideal model output is calculated, wherein the ideal model output represents a uniform distribution or a zero-knowledge state for the to-be-forgotten categories. The process of calculating the prototype-aware replay loss comprises: A cross-entropy loss is calculated by using the prototype-aware replay set.

6. The method of claim 1, wherein, The process of combining the distillation erasure loss and the prototype-aware replay loss into a total loss function comprises: The distillation erasure loss and the prototype-aware replay loss are linearly weighted. The process of updating the model parameters through the back propagation algorithm comprises: A gradient descent optimizer is used to update the model parameters at a model learning rate.

7. A system for electromagnetic signal forgetting learning based on mask distillation erasing, characterized in that, The system for implementing the method in any one of claims 1-6 comprises: A data division module for dividing an original training dataset into a to-be-forgotten dataset and a reserved dataset according to the requirement of a forgetting task; A feature extraction module for extracting features of all categories by using a pre-trained model; A replay set construction module for constructing a prototype-aware replay set through unsupervised clustering analysis and sample selection based on the extracted features; A loss calculation module for calculating a distillation erasure loss and a prototype-aware replay loss; A parameter updating module is configured to combine the distillation erasure loss and the prototype-aware playback loss into a total loss function, and update model parameters through a back propagation algorithm until the model converges.

8. A computer terminal device, characterized by The method comprises the steps of: one or more processors; a memory coupled to the processors, for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the method according to any one of claims 1-6.

9. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method according to any one of claims 1-6.

10. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method according to any one of claims 1-6.

Images