Method, system and equipment for realizing machine forgetting and readable storage medium
By using technologies such as meta-learning frameworks and relationship-aware entity embedding generators in the knowledge graph, the problem of poor generalization capabilities of existing machine forgetting methods in different scenarios is solved, and more efficient user privacy protection and model performance are achieved.
Patent Information
- Application Number
- CN202510622546.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-15
- Publication Date
- 2025-06-13
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing machine forgetting methods have poor generalization capabilities in different scenarios and cannot perform forgetting tasks efficiently, especially in data involving user privacy.
Using the meta-learning framework, by obtaining the data structure information of the target knowledge graph, extracting sub-graphs as different task scenarios, in which relationship-aware entity embedding generators and neighborhood enhancement modulators are trained, a knowledge graph embedding base model is constructed, and the base-default learner is used to remove the information that needs to be deleted.
By capturing the inherent laws of the knowledge graph and the internal logic in the forgetting task, the model can better adapt to different forgetting task scenarios, improve the ability to protect user privacy, and take into account the overall performance of the model and the forgetting effect of specific data.
Smart Images
Figure CN120146172A_ABST
Abstract
Description
Technical Field
[0001] The present invention discloses a method, system, device and readable storage medium for implementing machine unlearning, which relates to the technical field of artificial intelligence applications. Background Art
[0002] As an artificial intelligence model, knowledge graph embedding maps discrete entities and relationships in a knowledge graph into a low-dimensional vector space through mathematical expressions, so that these vectors can retain the semantic and structural information in the original graph. The knowledge graph embedding model simplifies the representation of the knowledge graph and lays a good foundation for the application of the knowledge graph in a variety of tasks.
[0003] Efficiently eliminating the influence of specific data in an artificial intelligence model, specific data, such as data related to user privacy, and retaining the overall performance of the model, such tasks are called machine unlearning tasks. With the help of machine unlearning, the influence of specific data in the model can be eliminated at a lower cost to achieve the protection of user privacy.
[0004] Machine unlearning is mainly applicable to performing unlearning tasks on a fixed and predefined knowledge graph, assuming that the entities and relationships are already known. Due to the need for privacy protection, the privacy content that users are concerned about is often consistent, such as personal information, detailed address, work unit, etc. These contents may have similarities in different unlearning tasks, but the existing methods do not take into account the internal connections between such tasks. As a result, the generalization ability of the existing methods is poor and they cannot efficiently perform unlearning tasks in different scenarios. Summary of the Invention
[0005] In view of the problems of the prior art, the present invention provides a method, system, device and readable storage medium for implementing machine unlearning. Based on a knowledge graph embedding model of metadata, with the help of machine unlearning, the influence of specific data in the model is eliminated at a lower cost to achieve the protection of user privacy.
[0006] The specific solution proposed by the present invention is as follows: The present invention provides a method for implementing machine unlearning, including: Step 1: Based on a meta-learning framework, obtain the data structure information of the target knowledge graph G, including the data of nodes, the number of relationships, and the triples describing the node resources. Step 2: Extract k subgraphs from the target knowledge graph G, regard the subgraphs as different task scenarios, and capture meta-knowledge in different task scenarios. Train the relationship-aware entity embedding generator and the neighborhood enhancement modulator in different task scenarios, where the initial embedding representation of the entity is generated according to the incoming relationship information and the outgoing relationship information of the entity, the initial embedding representation is optimized according to the multi-hop neighborhood information of the entity, an embedding integrator is added to the neighborhood enhancement modulator, and the concatenated high-dimensional initialization embedding representation is mapped to an entity embedding vector through the embedding integrator. Step 3: Construct a knowledge graph embedding base model according to the entity embedding vector. The knowledge graph embedding base model includes a base learner and a base de-learner, and optimize the base learner and the base de-learner. Step 4: For the information to be deleted, divide the tasks in the target knowledge graph according to the relationship, use the base de-learner to remove all relevant triples containing the information to be deleted, and fine-tune the relationship vectors of the remaining triples while keeping the relationship-aware entity embedding generator unchanged.
[0007] Further, in step 2 of the method for implementing machine forgetting, regarding the subgraphs as different task scenarios includes: Represented by the formula: ; Where represents the set of entities in the subgraph , represents the set of relationships in the subgraph , and represents the set of triples in the subgraph . Take some triples in the subgraph as the support set to enable the relationship-aware entity embedding generator to generate appropriate embedding representations; take another part of the triples as the query set to evaluate the quality of the generated embedding representations.
[0008] Further, in step 2 of the method for implementing machine forgetting, capturing meta-knowledge in different task scenarios includes: Using the formula: ; Where represents the task distribution, and each task is a specific task sampled from the task distribution. The function is an abstract representation of the meta-knowledge component in the model, and it is expected to perform well on the query set after observing the support set ; represents the loss function of the query set, represents the updated parameters after learning from the support set , represents the input value of the function , represents the label value.
[0009] Furthermore, in step 2 of the method for implementing machine forgetting, it includes: According to the incoming relationship information and outgoing relationship information of the entity, using the formula: ; Generate the initial embedding representation of the entity , where represents the outgoing relationship embedding matrix, represents the incoming relationship embedding matrix, represents the embedding of a specific relationship , represents the entity 's set of incoming relationships, represents the entity 's set of outgoing relationships.
[0010] The present invention also provides a system for implementing machine forgetting, including a data acquisition module, a meta-knowledge capture module, a base model management module, and a forgetting module. The data acquisition module, based on the meta-learning framework, acquires the data structure information of the target knowledge graph G, including the data of nodes, the number of relationships, and the triples describing the node resources. The meta-knowledge capture module extracts k subgraphs from the target knowledge graph G, regards the subgraphs as different task scenarios, and captures meta-knowledge in different task scenarios. And trains the relationship-aware entity embedding generator and the neighborhood enhancement modulator in different task scenarios, where the initial embedding representation of the entity is generated according to the incoming relationship information and outgoing relationship information of the entity, the initial embedding representation is optimized according to the multi-hop neighborhood information of the entity, an embedding integrator is added to the neighborhood enhancement modulator, and the concatenated high-dimensional initial embedding representation is mapped to an entity embedding vector through the embedding integrator. The base model management module constructs a knowledge graph embedding base model according to the entity embedding vector. The knowledge graph embedding base model includes a base learner and a base de-learner, and optimizes the base learner and the base de-learner. The forgetting module divides the tasks according to the relationships for the information to be deleted in the target knowledge graph, uses the base de-learner to remove all relevant triples containing the information to be deleted, and fine-tunes the relationship vectors of the remaining triples while keeping the relationship-aware entity embedding generator unchanged.
[0011] Furthermore, the meta-knowledge capture module of the system for implementing machine forgetting regards the subgraphs as different task scenarios, including: Represented by the formula: ; where represents the subgraph The set of entities in represents a sub-graph The set of relationships in, and represents the sub-graph The set of triples of. A part of the triples in the sub-graph is used as the support set to enable the relationship-aware entity embedding generator to generate appropriate embedding representations; another part of the triples is used as the query set to evaluate the quality of the generated embedding representations.
[0012] Furthermore, the meta-knowledge capture module for implementing the machine forgetting system captures meta-knowledge in different task scenarios, including: Using the formula: ; where represents the task distribution, and each task is a specific task sampled from the task distribution. The function is an abstract representation of the meta-knowledge component in the model, and it is expected to perform well on the query set after observing the support set ; represents the loss function of the query set, represents the parameters updated after learning from the support set ; represents the input value of the function ; represents the label value.
[0013] Furthermore, the meta-knowledge capture module for implementing the machine forgetting system uses the formula according to the incoming relationship information and outgoing relationship information of the entity: ; to generate the initial embedding representation of the entity, where represents the outgoing relationship embedding matrix, represents the incoming relationship embedding matrix, represents the embedding of a specific relationship ; represents the entity 's incoming relationship set, represents the entity 's outgoing relationship set.
[0014] The present invention also provides a device for implementing machine forgetting, including: at least one memory and at least one processor; The at least one memory is used to store machine-readable programs; The at least one processor is used to call the machine-readable program and execute the method for implementing machine forgetting described above.
[0015] The present invention also provides a readable storage medium, on which computer instructions are stored. When the computer instructions are executed by a processor, the processor is caused to execute the described method for implementing machine forgetting.
[0016] The advantages of the present invention are as follows: The present invention circumvents the problem of insufficient capture of the internal connection information in the forgetting task in the existing knowledge graph embedding learning method, and obtains the internal logic in the forgetting task and the internal rules of the knowledge graph through a meta-learning framework, enabling the model to better adapt to different forgetting task scenarios.
[0017] The present invention uses a GNN to capture features in the knowledge graph structure, uses a relation-aware entity embedding generator to capture the type features of entities, and uses a neighborhood enhancement modulator to capture the instance features of entities. Compared with the existing learning only for specific entities, the present invention can better handle task scenarios with unknown entities or dynamic knowledge graph scenarios.
[0018] The present invention can integrate multiple base models to better generalize the model to dynamic forgetting task scenarios, and balance the overall performance of the model and the forgetting effect on specific data. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] Figure 1 It is a schematic flowchart of the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0020] Forgetting set: In the machine forgetting scenario, the forgetting set represents the data set that the machine learning model should forget or the data set that the machine learning model should reduce its influence on, manifested as the poor performance of the machine learning model on these data sets.
[0021] Retention set: In the machine forgetting scenario, the retention set represents the data set that the machine learning model should "retain", manifested as the good performance of the machine learning model on these data sets.
[0022] Retraining: In the machine forgetting scenario, retraining means that after removing the forgetting set from the original data set, training is carried out again on this data set to expect to eliminate the influence of the data in the forgetting set, which is the simplest and most crude machine forgetting method.
[0023] Forgetting task: The main purpose of performing a forgetting task in a machine learning model is to eliminate some sensitive information in the model to protect user privacy and meet the requirements of relevant privacy protection laws.
[0024] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, so that those skilled in the art can better understand the present invention and be able to implement it, but the embodiments given are not intended to limit the present invention.
[0025] Embodiment 1: The present invention provides a method for implementing machine forgetting, including: Step 1: Based on the meta-learning framework, obtain the data structure information of the target knowledge graph G, including the data of nodes, the number of relationships, and the triples describing the node resources.
[0026] Step 2: Extract k subgraphs from the target knowledge graph G, and regard the subgraphs as different task scenarios, including: represented by the formula: ; Where represents the entity set in the subgraph , represents the relationship set in the subgraph , and represents the triple set of the subgraph . Part of the triples in the subgraph are used as the support set to enable the relationship-aware entity embedding generator to generate appropriate embedding representations; another part of the triples are used as the query set to evaluate the quality of the generated embedding representations.
[0027] And calculate the training loss function: ; According to the target knowledge graph construct a data set , including a training set and a test set . After completing the construction of the meta-training data set for the target knowledge graph, the model regards each subgraph as a task and trains according to the following meta-learning objective equation to achieve the capture of meta-knowledge across task scenarios: ; Where represents the task distribution, and each task is a specific task sampled from the task distribution. The function is an abstract representation of the meta-knowledge component in the model, and it is expected to perform well on the query set after observing the support set ; represents the loss function of the query set, represents the parameters updated after learning from the support set , represents the input value of the function , Represents the tag value.
[0028] And train the relationship-aware entity embedding generator and the neighborhood enhancement modulator in different task scenarios: generate the initial embedding representation of the entity according to the incoming relationship information and outgoing relationship information of the entity, optimize the initial embedding representation according to the multi-hop neighborhood information of the entity, add an embedding integrator in the neighborhood enhancement modulator, and map the concatenated high-dimensional initialized embedding representation to an entity embedding vector through the embedding integrator. Among them, according to the incoming relationship information and outgoing relationship information of the entity, using the formula: ; Generate the initial embedding representation of the entity , where Represents the outgoing relationship embedding matrix, Represents the incoming relationship embedding matrix, Represents a specific relationship Embedding, Represents the entity Incoming relationship set of, Represents the entity Outgoing relationship set of.
[0029] Further optimize the initial embedding representation according to the multi-hop neighborhood information of the entity , specifically as shown in the following formula: ; Where Represents the feature representation of the node At the th layer, Represents the neighbor node set connected to the node Through the relationship , Is a normalization factor for dealing with node degree imbalance, usually defined as . Is the th layer, the weight matrix of the relationship , Is the self-loop weight matrix for dealing with the own features of the node . Represents the activation function, usually ReLU is used. And the input representation of NEEM is set to To more flexibly apply low - order and high - order neighborhood information to select the most appropriate neighborhood information for each entity, a Hierarchical Embedding Integrator (HEI) is added to the Neighbor - Enhanced Modulator (NEEM), and its formula is as follows: ; where represents the final embedding of the entity; represents the Hierarchical Embedding Integrator, which is used to hierarchically integrate all embeddings from layer 0 to layer ; represents the concatenation operation between levels; represents the transformation matrix in HEI, which maps the concatenated high - dimensional representation to the final entity embedding vector.
[0030] Step 3: Construct a knowledge graph embedding base model based on the entity embedding vectors. The knowledge graph embedding base model includes a base learner and a base de - learner, and optimize the base learner and the base de - learner.
[0031] The base learner can be optimized using the following formula: ; where represents the base learner, represents the weight assigned to the th base learner, satisfying , represents the number of base learners, represents the th base learner's loss function on the query set.
[0032] The base de - learner can be optimized using the following formula: ; represents the base de - learner, represents the weight assigned to the th base de - learner, satisfying , represents the number of base de - learners, represents the th base de - learner's loss function on the query set.
[0033] Step 4: For the information to be deleted, divide the tasks in the target knowledge graph according to the relationships, use the base de - learner to remove all relevant triples containing the information to be deleted, and fine - tune the relationship vectors of the remaining triples while keeping the relationship - aware entity embedding generator unchanged.
[0034] Embodiment 2: The present invention also provides a machine forgetting system implementation, including a data acquisition module, a meta-knowledge capture module, a base model management module, and a forgetting module. Based on the meta-learning framework, the data acquisition module obtains the data structure information of the target knowledge graph G, including the data of nodes, the number of relationships, and the triples describing the node resources. The meta-knowledge capture module extracts k subgraphs from the target knowledge graph G, regards the subgraphs as different task scenarios, and captures meta-knowledge in different task scenarios. And trains the relationship-aware entity embedding generator and the neighborhood enhancement modulator in different task scenarios. Among them, according to the in-coming relationship information and out-coming relationship information of the entity, the initial embedding representation of the entity is generated, and the initial embedding representation is optimized according to the multi-hop neighborhood information of the entity. An embedding integrator is added to the neighborhood enhancement modulator, and the concatenated high-dimensional initialized embedding representation is mapped to an entity embedding vector through the embedding integrator. The base model management module constructs a knowledge graph embedding base model according to the entity embedding vector. The knowledge graph embedding base model includes a base learner and a base de-learner, and optimizes the base learner and the base de-learner. For the information to be deleted, the forgetting module divides tasks according to relationships in the target knowledge graph, uses the base de-learner to remove all relevant triples containing the information to be deleted, and fine-tunes the relationship vectors of the remaining triples, keeping the relationship-aware entity embedding generator unchanged.
[0035] Regarding the information interaction and execution process between the above modules in the system, since they are based on the same concept as the method embodiment of the present invention, the specific content can be referred to the description in the method embodiment of the present invention, and will not be elaborated here.
[0036] Similarly, the advantages of the system of the present invention are: It avoids the problem of insufficient capture of the internal connection information in the forgetting task in the existing knowledge graph embedding de-learning method, and obtains the internal logic in the forgetting task and the internal rules of the knowledge graph through the meta-learning framework, enabling the model to better adapt to different forgetting task scenarios.
[0037] It uses GNN to capture the features in the knowledge graph structure, uses the relationship-aware entity embedding generator to capture the type features of entities, and uses the neighborhood enhancement modulator to capture the instance features of entities. Compared with the existing de-learning that only targets specific entities, the present invention can better handle task scenarios with unknown entities or dynamic knowledge graph scenarios.
[0038] By integrating multiple base models, the model can be better generalized to dynamic forgetting task scenarios, and the overall performance of the model and the forgetting effect on specific data can be taken into account.
[0039] It should be noted that not all steps and modules in the above processes and system architectures are necessary, and some steps or modules can be ignored according to actual needs. The execution order of each step is not fixed and can be adjusted as required. The system architectures described in the above embodiments can be physical architectures or logical architectures, that is, some modules may be implemented by the same physical entity, or some modules may be implemented by multiple physical entities separately, or some components in multiple independent devices may be jointly implemented.
[0040] Embodiment 3: The present invention also provides a device for implementing machine forgetting, including: at least one memory and at least one processor; The at least one memory is used to store machine-readable programs; The at least one processor is used to call the machine-readable program and execute the method for implementing machine forgetting.
[0041] Regarding the content such as information interaction and execution process between the processor and the memory in the above device, since it is based on the same concept as the method embodiment of the present invention, the specific content can be referred to the description in the method embodiment of the present invention and will not be elaborated here.
[0042] Similarly, the advantages of the device of the present invention are: It avoids the problem of insufficient capture of the internal connection information in the forgetting task in the existing knowledge graph embedding learning method, and obtains the internal logic in the forgetting task and the internal rules of the knowledge graph through the meta-learning framework, so that the model can better adapt to different forgetting task scenarios.
[0043] It uses GNN to capture the features in the knowledge graph structure, uses a relation-aware entity embedding generator to capture the type features of entities, and uses a neighborhood enhancement modulator to capture the instance features of entities. Compared with the existing learning only for specific entities, the present invention can better handle task scenarios with unknown entities or dynamic knowledge graph scenarios.
[0044] By integrating multiple base models, the model can be better generalized to dynamic forgetting task scenarios, and the overall performance of the model and the forgetting effect on specific data can be taken into account.
[0045] Embodiment 4: The present invention also provides a readable storage medium, on which computer instructions are stored. When the computer instructions are executed by a processor, the processor executes the method for implementing machine forgetting. Specifically, a system or device equipped with a storage medium can be provided. On this storage medium, software program codes for implementing the functions of any one of the above embodiments are stored, and the computer (or CPU or MPU) of the system or device is made to read and execute the program codes stored in the storage medium.
[0046] In this case, the program code read from the storage medium itself can implement the functions of any one of the above embodiments, so the program code and the storage medium storing the program code constitute a part of the present invention.
[0047] Examples of the storage medium for providing the program code include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD+RW), magnetic tapes, non-volatile memory cards, and ROMs. Optionally, the program code can be downloaded from a server computer via a communication network.
[0048] Furthermore, it should be clear that not only can the functions of any one of the above embodiments be realized by executing the program code read by the computer, but also by causing an operating system or the like operating on the computer based on the instructions of the program code to complete part or all of the actual operations.
[0049] In addition, it can be understood that the program code read from the storage medium is written into the memory provided in the expansion board inserted into the computer or into the memory provided in the expansion unit connected to the computer, and then based on the instructions of the program code, the CPU or the like installed on the expansion board or the expansion unit is caused to execute part or all of the actual operations, thereby realizing the functions of any one of the above embodiments.
[0050] The above-described embodiments are merely preferred embodiments given to fully illustrate the present invention, and the protection scope of the present invention is not limited thereto. Equivalent substitutions or transformations made by those skilled in the art on the basis of the present invention are all within the protection scope of the present invention. The protection scope of the present invention is subject to the claims.
Claims
1. A method for realizing machine forgetting, characterized in that include: Step 1: Based on the meta-learning framework, obtain the data structure information of the target knowledge graph G, including the data of the nodes, the number of relationships, and the triples of the node resource description. Step 2: Extract k subgraphs based on the target knowledge graph G, treat the subgraphs as different task scenarios, and capture meta-knowledge in different task scenarios. The relation-aware entity embedding generator and the neighborhood enhancement modulator are trained in different task scenarios, wherein an initial embedding representation of the entity is generated according to the inbound and outbound relation information of the entity, the initial embedding representation is optimized according to the multi-hop neighborhood information of the entity, an embedding integrator is added to the neighborhood enhancement modulator, and the concatenated high-dimensional initialized embedding representation is mapped to an entity embedding vector through the embedding integrator. Step 3: Construct a knowledge graph embedding base model based on the entity embedding vector. The knowledge graph embedding base model includes a base learner and a base delearner. Optimize the base learner and the base delearner. Step 4: For the information that needs to be deleted, divide the tasks according to the relations in the target knowledge graph, use the basis learner to remove all related triplets containing the information that needs to be deleted, fine-tune the relation vectors of the remaining triplets, and keep the relation-aware entity embedding generator unchanged.
2. A method for realizing machine forgetting according to claim 1, characterized in that In step 2, the subgraphs are considered as different task scenarios, including: Using the formula: ； in Representative subgraph The entity set in Representative subgraph The set of relations in , and Represents a subgraph A set of triples in the subgraph is used as a support set to allow the relation-aware entity embedding generator to generate appropriate embedding representations; and another part of the triples is used as a query set to evaluate the quality of the generated embedding representation.
3. The method for realizing machine forgetting according to claim 1, characterized in that In step 2, meta-knowledge is captured in different task scenarios, including: Using the formula: ； in Represents the task distribution, and each task are specific tasks sampled from the task distribution, function It is an abstract representation of the meta-knowledge component in the model, which is expected to be After that, in the query set Good performance on represents the loss function of the query set, Represents the support set The parameters updated after learning, Representation function The input value of Indicates a tag value.
4. A method for achieving machine forgetting according to claim 3, characterized in that Step 2 includes: According to the entity's inbound and outbound relationship information, use the formula: ； Generate an initial embedding representation of the entity ,in represents the outgoing relation embedding matrix, represents the incoming relation embedding matrix, Indicates a specific relationship Embedded, Representing Entities The incoming relationship set of Representing Entities The set of outgoing relations.
5. A system for realizing machine forgetting, characterized in that It includes data acquisition module, meta-knowledge capture module, base model management module and forgetting module. The data acquisition module is based on the meta-learning framework to obtain the data structure information of the target knowledge graph G, including the data of the nodes, the number of relationships, and the triples of the node resource description. The meta-knowledge capture module extracts k sub-graphs based on the target knowledge graph G, regards the sub-graphs as different task scenarios, and captures meta-knowledge in different task scenarios. The relation-aware entity embedding generator and the neighborhood enhancement modulator are trained in different task scenarios, wherein an initial embedding representation of the entity is generated according to the inbound and outbound relation information of the entity, the initial embedding representation is optimized according to the multi-hop neighborhood information of the entity, an embedding integrator is added to the neighborhood enhancement modulator, and the concatenated high-dimensional initialized embedding representation is mapped to an entity embedding vector through the embedding integrator. The base model management module constructs a knowledge graph embedding base model according to the entity embedding vector. The knowledge graph embedding base model includes a base learner and a base delearner, and optimizes the base learner and the base delearner. For the information that needs to be deleted, the forgetting module divides the task according to the relationship in the target knowledge graph, uses the basis learner to remove all related triplets containing the information that needs to be deleted, fine-tunes the relationship vectors of the remaining triplets, and keeps the relationship-aware entity embedding generator unchanged.
6. A system for realizing machine forgetting according to claim 5, characterized in that The meta-knowledge capture module considers subgraphs as different task scenarios, including: Using the formula: ； in Representative subgraph The entity set in Representative subgraph The set of relations in , and Represents a subgraph A set of triples in the subgraph is used as a support set to allow the relation-aware entity embedding generator to generate appropriate embedding representations; and another part of the triples is used as a query set to evaluate the quality of the generated embedding representation.
7. The system for realizing machine forgetting according to claim 5, characterized in that The meta-knowledge capture module captures meta-knowledge in different task scenarios, including: Using the formula: ； in Represents the task distribution, and each task are specific tasks sampled from the task distribution, function It is an abstract representation of the meta-knowledge component in the model, which is expected to be After that, in the query set Good performance on represents the loss function of the query set, Represents the support set The parameters updated after learning, Representation function The input value of Indicates a tag value.
8. The system for realizing machine forgetting according to claim 5, characterized in that The meta-knowledge capture module uses the formula based on the inbound and outbound relationship information of the entity: ； Generate an initial embedding representation of the entity ,in represents the outgoing relation embedding matrix, represents the incoming relation embedding matrix, Indicates a specific relationship Embedded, Representing Entities The incoming relationship set of Representing Entities The set of outgoing relations.
9. A device for achieving machine forgetting, characterized in that include: at least one memory and at least one processor; The at least one memory is used to store a machine-readable program; The at least one processor is used to call the machine-readable program to execute a method for realizing machine forgetting as described in any one of claims 1 to 4.
10. A readable storage medium, characterized in that The readable storage medium stores computer instructions, and when the computer instructions are executed by a processor, the processor executes a method for realizing machine forgetting as described in any one of claims 1 to 4.