Access control method, electronic device, storage medium and program product

By obtaining the aggregation model parameter gradient of AI applications, determining data access permissions and security levels, combining dynamic sandbox and local data training, the privacy leakage problem caused by AI applications accessing data indefinitely is solved, and dynamic permission control and data security are achieved.

CN120470623APending Publication Date: 2025-08-12ZTE CORP
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202510575386.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-04-30
Publication Date
2025-08-12

AI Technical Summary

Technical Problem

In the prior art, AI applications can access data indefinitely after obtaining authorization, resulting in a high risk of privacy data leakage, and the static authorization mechanism cannot dynamically adjust permissions, resulting in the problem of data residual leakage during over-acquisition of data and model training.

Method used

By obtaining the aggregation model parameter gradient of the target application, determining the data access permissions and security levels, dynamically generating access control policies based on the data access scope, using dynamic sandboxes to restrict data access of AI applications, realizing local data training and desensitization processing, and avoiding illegal access and over-acquisition.

Benefits of technology

Effectively prevent AI applications from over-accessing private data, improve the security of private data, prevent data leakage, ensure that data is not separated locally, and dynamically adjust permissions to meet the access needs of AI applications.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120470623A_ABST
    Figure CN120470623A_ABST
Patent Text Reader

Abstract

The invention provides an access control method, electronic equipment, a storage medium and a program product, relates to the technical field of computers, and is used for solving the problems that when an application accesses data, the risk of privacy data leakage exists, and the security of privacy data is relatively low. The method comprises the steps of obtaining a parameter gradient of an aggregation model corresponding to a target application, determining a data access permission corresponding to the target application based on the parameter gradient of the aggregation model, and indicating a security level of each data type allowed to be accessed by the target application through the data access permission; and further, determining an access control strategy of the target application based on the data access range and the data access permission of the target application, thereby indicating the data allowed to be accessed by the target application and / or the data forbidden to be accessed by the target application through the access control strategy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and in particular to an access control method, electronic device, storage medium, and program product. Background Art

[0002] With the popularity of artificial intelligence (AI) applications (such as AI chat, smart recommendations, AI agents, etc.), users face security risks of privacy data leakage when using AI applications. For example, when using AI smart albums to help users automatically classify photos in albums, it is necessary to authorize the application (i.e., AI smart albums) to access data (such as albums, contacts, location information, etc.). Once the application obtains authorization, the application can access the data anytime, anywhere, and even upload it to a third-party server. Therefore, when current applications access data, there is a risk of privacy data leakage, and the security of privacy data is low. Summary of the Invention

[0003] The embodiments of the present disclosure provide an access control method, an electronic device, a storage medium, and a program product, which are used to at least solve the problem of the risk of privacy data leakage and low security of privacy data when applications access data.

[0004] In one aspect, an access control method is provided, the method comprising: obtaining a parameter gradient of an aggregation model corresponding to a target application;

[0005] Determine the data access rights corresponding to the target application based on the parameter gradient of the aggregation model. The data access rights are used to indicate the security level of each data type that the target application is allowed to access.

[0006] Based on the data access scope and data access rights of the target application, an access control policy of the target application is determined, where the access control policy is used to indicate data that the target application is allowed to access and / or data that the target application is prohibited from accessing.

[0007] In another aspect, an access control device is provided, including an information acquisition module and an access control module.

[0008] An information acquisition module, used to obtain the parameter gradient of the aggregation model corresponding to the target application;

[0009] An access control module is used to determine the data access rights corresponding to the target application based on the parameter gradient of the aggregation model. The data access rights are used to indicate the security level of each data type that the target application is allowed to access;

[0010] The access control module is further used to determine the access control policy of the target application based on the data access scope and data access rights of the target application. The access control policy is used to indicate the data that the target application is allowed to access and / or the data that the target application is prohibited from accessing.

[0011] On the other hand, an electronic device is provided, comprising: a memory and a processor; the memory and the processor are coupled; the memory is used to store instructions executable by the processor; and the processor executes the access control method of any of the above embodiments when executing the instructions.

[0012] On the other hand, a computer-readable storage medium is provided, on which computer instructions are stored. When the computer instructions are executed on a computer, the computer executes the access control method of any of the above embodiments.

[0013] On the other hand, a computer program product is provided. When the computer program product is run on a computer, the computer is enabled to execute the access control method of any one of the above embodiments.

[0014] In an embodiment of the present disclosure, by obtaining the parameter gradient of the aggregation model corresponding to the target application, the data access rights corresponding to the target application can be determined to indicate the security level of each data type that the target application is allowed to access. In this way, when the security level of each data type that the target application is allowed to access is determined, the access control policy of the target application can be determined in combination with the data access scope of the target application. Thus, the target application's access to data is controlled by the data indicated by the access control policy, which data the target application is allowed to access, and / or the data that the target application is prohibited from accessing. In this way, based on the parameter gradient of the aggregation model corresponding to the target application, it is possible to accurately determine which data the target application is allowed to access, thereby accurately limiting the data that the target application can access when formulating the access control policy. This prevents the target application from excessively accessing other private data (i.e., data that the target application does not need to access) when accessing data, thereby avoiding the leakage of private data and improving the security of private data. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the present disclosure, the following briefly introduces the drawings required for use in some embodiments of the present disclosure. Obviously, the drawings described below are only drawings of some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0016] Figure 1 A schematic diagram of an access control system provided by an embodiment of the present disclosure;

[0017] Figure 2 A schematic diagram of a software system architecture of an electronic device provided in an embodiment of the present disclosure;

[0018] Figure 3 A flowchart of an access control method provided by an embodiment of the present disclosure;

[0019] Figure 4 A flowchart of another access control method provided by an embodiment of the present disclosure;

[0020] Figure 5 A schematic structural diagram of an access control device provided in an embodiment of the present disclosure;

[0021] Figure 6 A schematic structural diagram of an electronic device provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0022] The following will clearly and completely describe the technical solutions of this disclosure in conjunction with the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of this disclosure, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of this disclosure without making any creative efforts shall fall within the scope of protection of this disclosure.

[0023] It should be noted that in this disclosure, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described in this disclosure as "exemplary" or "for example" should not be construed as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0024] In the following, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the quantity of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the features.

[0025] In the description of this disclosure, unless otherwise specified, " / " means "or." For example, A / B can mean A or B. "And / or" in this document simply describes an association relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can mean: A exists alone, A and B exists simultaneously, and B exists alone. Furthermore, "at least one" means one or more, and "a plurality" means two or more.

[0026] Before introducing the access control method provided by the embodiment of the present disclosure in detail, the implementation environment and application scenarios of the embodiment of the present disclosure are first introduced.

[0027] First, the application scenarios of the embodiments of the present disclosure are introduced.

[0028] The disclosed embodiments are applicable to AI application scenarios that require a high level of security for privacy data, and require the deployment of mobile devices with trusted execution environments, software systems that integrate federated learning frameworks and dynamic sandbox strategies, and network environments with decentralized blockchain nodes.

[0029] Currently, when an electronic device (terminal device) runs an AI application to perform an AI task, the AI application needs to access data information in the electronic device (such as photos in the album, contact information, location information, camera permissions, etc.).

[0030] For example, many users currently use AI smart albums to automatically categorize photos stored in the photo albums of electronic devices, such as photos by location information when they were taken, people included in the photos, and scene types (scenery, people, etc.). Therefore, it is necessary to authorize the AI smart album to access the data information in the electronic device (i.e., a static authorization mechanism), so that the AI smart album can categorize photos based on their content and generate classification labels.

[0031] However, after authorizing the AI Smart Album to access data on an electronic device, granting it data access rights, if the user does not adjust the AI Smart Album's access rights, the AI Smart Album may retain such rights for an extended period. This could allow the AI Smart Album to abuse its data access rights and, when the user does not need it to perform tasks, continue to scan photos in the background, obtain data such as the device's contacts, or even upload it to a third-party server, thereby leaking private data.

[0032] In addition, after the AI smart album is trained (federated learning) based on the data information in the electronic device, there may be a gradient reversal attack. That is, the encrypted gradient uploaded during the federated learning process may be reversed by the attacker to deduce the data information used during training (such as photo features, facial contours, etc.). In addition, the training data generated during the training process will remain in the graphics processing unit (GPU) memory or storage of the electronic device for a long time, resulting in residual data leakage. That is, the trained model parameters and temporary image tensors will remain in the GPU memory or storage, which may be stolen by malicious applications.

[0033] Understandably, the static authorization mechanism relies entirely on the native permissions system. Once a user grants an application access rights, the application has unlimited access to all authorized data. This mechanism cannot dynamically narrow the scope of data access based on AI task requirements (for example, limiting access to specific data, such as a specific album subdirectory), nor can it prevent the application from continuously scanning sensitive data in the background. Furthermore, the Trusted Execution Environment (TEE) only isolates and protects the model training process corresponding to the application, but does not impose hardware-level restrictions on data access at the application level. This allows attackers to exploit Android Binder vulnerabilities to bypass the isolation barrier and directly steal raw, unmasked data (such as raw image files). More seriously, intermediate data generated during model training (such as normalized image matrices and facial feature tensors) remains in public memory and lacks a secure erase mechanism. Attackers can extract this residual data from the GPU memory via Direct Memory Access (DMA), thereby fully recovering the user's private image features. These flaws collectively result in AI applications failing to guarantee data security after authorization.

[0034] It is understandable that the relevant technologies have the following problems: data centralization risk. Although federated learning realizes distributed training, gradient aggregation still relies on the central server, and there is a possibility that intermediate parameters will leak user privacy (for example, inferring the original data through gradients); static permission management defects. The permission system adopts a default authorization or dynamic authorization mode. Once the AI application is authorized, it can continue to access sensitive data information (such as photo albums, microphones, contacts, text messages, etc.), resulting in excessive data collection by AI applications; data residual threat: after the AI application completes the task, the temporarily cached data information is not completely cleared (such as image features stored in the GPU video memory), and the temporarily cached data information may be stolen by malicious processes.

[0035] This is due to the conflict between the protection of private data information and the utility of the model during model training. Over-reliance on encryption or noise injection leads to a significant decrease in model accuracy, while relaxing the protection of data information leads to the risk of data leakage. In addition, AI applications require short-term, high-frequency dynamic adjustment of access rights, and static authorization mechanisms cannot adapt in real time to the principle of minimized access of AI applications (access only to required data information, and not to other unnecessary data information). In addition, there is an imbalance between software protection and hardware attacks. Pure software solutions (such as container isolation) are difficult to defend against physical-level attacks (such as DMA stealing video memory data), forming a security protection gap. These problems together make it difficult to balance the protection of private data information with the needs of intelligence.

[0036] In order to solve the above problems, the embodiments of the present disclosure provide an access control method, and the access control method provided by the embodiments of the present disclosure is applied to electronic devices. The local data training of the model corresponding to the AI application is performed using the local security zone to generate intermediate parameters (i.e., the parameter gradient of the intermediate model), and the local data is desensitized so that users cannot directly obtain the source data. In this way, through the local training results, the parameter gradient of the intermediate model is uploaded to the cloud, and the local data is authenticated by security levels, to avoid illegal access to data or excessive data collection. Combined with the scenario of the AI application, the weight and security level of the data accessed by the AI application are determined, and the parameter gradient of the trained model is fitted in the cloud. The local electronic device dynamically restricts the access of the AI application to data information in the AI permissions dynamic sandbox according to the changes in the parameter gradient transmitted back from the cloud, thereby achieving training integrity and user privacy protection.

[0037] It should be noted that the AI permission dynamic sandbox can dynamically adjust the permission control strategy according to the model training status, and realize permission control with dynamic strategy generation, loading and recovery.

[0038] In an embodiment of the present disclosure, by obtaining the parameter gradient of the aggregation model corresponding to the target application, the data access rights corresponding to the target application can be determined to indicate the security level of each data type that the target application is allowed to access. In this way, when the security level of each data type that the target application is allowed to access is determined, the access control policy of the target application can be determined in combination with the data access scope of the target application. Thus, the target application's access to data is controlled by the data indicated by the access control policy, which data the target application is allowed to access, and / or the data that the target application is prohibited from accessing. In this way, based on the parameter gradient of the aggregation model corresponding to the target application, it is possible to accurately determine which data the target application is allowed to access, thereby accurately limiting the data that the target application can access when formulating the access control policy. This prevents the target application from excessively accessing other private data (i.e., data that the target application does not need to access) when accessing data, thereby avoiding the leakage of private data and improving the security of private data.

[0039] The implementation environment of the embodiments of the present disclosure is introduced below.

[0040] like Figure 1 , which is a schematic diagram of an access control system provided by an embodiment of the present disclosure, and is applied to an electronic device. The access control system may include: an electronic device 301 and a server 302 .

[0041] The electronic device 301 is configured to obtain a parameter gradient of an aggregation model corresponding to a target application.

[0042] Specifically, the electronic device 301 may obtain the parameter gradient of the aggregation model corresponding to the target application from the server 302 .

[0043] The electronic device 301 is further configured to determine data access permissions corresponding to the target application based on the parameter gradient of the aggregation model. The data access permissions are used to indicate the security level of each data type that the target application is allowed to access.

[0044] The electronic device 301 is further configured to determine an access control policy of the target application based on the data access scope and data access rights of the target application. The access control policy is configured to indicate data that the target application is allowed to access and / or data that the target application is prohibited from accessing.

[0045] Among them, the electronic device 301 can be a mobile phone, a tablet computer, a laptop computer, an ultra-mobile personal computer (UMPC), a handheld computer, a netbook, a personal digital assistant (PDA), a wearable electronic device, a vehicle-mounted device, etc., including a device with a naked-eye 3D display screen, and the embodiments of the present application do not impose any restrictions on this.

[0046] Exemplarily, the electronic device 301 may include a processor, a display screen, a camera, etc. The processor may include one or more processing units, for example, an application processor, a modem processor, a graphics processor, an image signal processor, a controller, a video codec, a digital signal processor, a baseband processor, and / or a neural network processor, etc. Different processing units may be independent devices or integrated into one or more processors.

[0047] Of course, the electronic device 301 provided in the embodiment of the present application may also include one or more devices such as a battery, an antenna, a mobile communication module, a wireless communication module, an audio module, a speaker, a receiver, a microphone, a headphone jack, a button, a motor, an indicator, and a SIM card interface, and the embodiment of the present application does not impose any restrictions on this.

[0048] In one possible implementation, Figure 2 As shown, it is a schematic diagram of the software system architecture of an electronic device provided by an embodiment of the present disclosure. The layered architecture of the electronic device includes: an application layer, a policy control layer, a dynamic sandbox layer, a federated learning layer and a hardware security layer.

[0049] Among them, the application layer can include multiple AI applications, such as AI smart album, AI smart camera, AI voice assistant, etc.

[0050] The policy control layer can include a dynamic permissions allocation engine, a federated learning task scheduler, and a privacy compliance audit module. The policy control layer interacts with the server to provide feedback on the access control policy for AI applications. The policy control layer and the server are connected via a blockchain gradient aggregation network.

[0051] Specifically, the dynamic permission allocation engine is used to dynamically allocate data access rights for AI applications based on the parameter gradient of the aggregation model; the federated learning task scheduler is used to train the model of AI applications during the federated learning phase; and the privacy compliance audit module is used to determine the degree of privacy of data information in electronic devices.

[0052] The dynamic sandbox layer may include: a hardware driver-level isolation container, a runtime permission interception module, and a data flow monitoring agent module. Among them, the hardware driver-level isolation container may specifically include: a camera sandbox, a microphone sandbox, a file system sandbox (Data Center Infrastructure Management (DCIM) area), etc. The camera sandbox is used to control the access rights of the camera, the microphone sandbox is used to control the access rights of the microphone, and the file system sandbox is used to control the access rights of the stored data information. The runtime permission interception module is used to control the access rights of the AI application based on the access control policy of the determined AI application, and to determine whether to intercept the AI application's access to data information when the AI application is running. The data flow monitoring agent module is used to monitor the data information accessed by the AI application in real time.

[0053] The federated learning layer can include a real-time database and an encrypted transmission channel. The real-time database includes a noise injection module and a model quantization and compression engine. The noise injection module is used to inject noise into the model during the federated learning phase, and the model quantization and compression engine is used to quantize and compress the trained model during the federated learning phase. The encrypted transmission channel is used to encrypt the parameter gradients of the trained model.

[0054] The hardware security layer may include: ARM TrustZone chip, Titan chip, Neural-network Processing Unit (NPU) (acceleration unit), and secure storage partitions. The ARM TrustZone chip is used to isolate the hardware and software resources in the system into secure areas and ordinary areas. Ordinary area components cannot access secure area resources, thereby ensuring that peripherals such as secure memory and encryption blocks are protected from software attacks. The Titan chip is used to establish a hardware root of trust for the platform, providing a basic security layer to prevent physical attacks and threats to user data. The NPU is used for model inference and is optimized for the characteristics of neural networks. It can efficiently process common neural network operations such as convolution, pooling, and full connection. It has powerful parallel computing capabilities and can process multiple data elements at the same time, greatly improving computing efficiency. The secure storage partition can be an embedded Multi-Media Card (eMMC).

[0055] After introducing the application scenario and implementation environment of the embodiment of the present disclosure, the access control method provided by the embodiment of the present disclosure is described in detail below in combination with the above implementation environment.

[0056] The access control method provided by the embodiments of the present disclosure can be applied to AI application scenarios such as smart terminal photo albums / contact data privacy protection / smart doctors, to prevent AI applications from excessively collecting user data, dynamically control and ensure that data does not leave the local area, and avoid leakage of user privacy data.

[0057] The present disclosure provides an access control method, which is applied to electronic devices, such as Figure 3 As shown, the access control method may include: S301-S303.

[0058] S301: Obtain the parameter gradient of the aggregation model corresponding to the target application.

[0059] In the embodiment of the present disclosure, the electronic device may obtain the parameter gradient of the aggregation model corresponding to the target application from the server, or may obtain the parameter gradient of the aggregation model corresponding to the target application locally on the electronic device.

[0060] In a possible implementation, the electronic device may locally train an aggregation model corresponding to the target application, and thereby determine a parameter gradient corresponding to the aggregation model based on the trained aggregation model.

[0061] In one possible implementation, the electronic device can also locally train the intermediate model corresponding to the target application, and send the parameter gradient of the intermediate model to the server, so that the server can train the aggregate model based on the parameter gradients of the multiple intermediate models received, and then return the parameter gradient of the aggregate model to the electronic device.

[0062] It should be noted that the parameter gradient of the aggregation model is used to indicate the highest security level of each data type included in the electronic device that the target application needs to access when performing a task.

[0063] Specifically, after obtaining the parameter gradient of the aggregation model from the cloud server, the electronic device can modify the local parameters based on the parameter gradient of the aggregation model, thereby confirming the data access rights corresponding to the target application according to the parameter gradient of the aggregation model returned by the cloud server.

[0064] S302: Determine the data access permission corresponding to the target application based on the parameter gradient of the aggregation model.

[0065] The data access permission is used to indicate the security level of each data type that the target application is allowed to access.

[0066] In a possible implementation, the data information included in the electronic device can be specifically divided into multiple data types, for example, an image data type, a contact data type, a short message data type, a positioning information data type, and the like.

[0067] In one possible implementation, data of each data type may be specifically divided into multiple security levels, so that when the target application has access rights to a certain data type at different security levels, different data information in the certain data type may be accessed.

[0068] For example, for image data, the image data can be divided into three security levels (level 1, level 2, and level 3). Level 1 refers to the facial feature information included in the image data, level 2 refers to the geographic location information when the image data was generated (shot), and level 3 refers to the background information (scene, color, etc.) of the image data.

[0069] It should be noted that among the data security levels, the first level security level represents the security level with the highest degree of privacy. Of course, the first level security level can also represent the security level with the lowest degree of privacy. The embodiment of this application takes the first level security level as the security level with the highest degree of privacy as an example for explanation.

[0070] It can be understood that when the parameter gradient of the aggregation model indicates the highest security level of each data type included in the electronic device that the target application needs to access when performing a task, based on the parameter gradient of the aggregation model, the highest security level of data that the target application can access when accessing data information of each data type can be determined.

[0071] For example, if the parameter gradient of the aggregation model indicates that the highest security level of the image data type is the second security level, then when the target application accesses the data of the image data type included in the electronic device, the target application is allowed to access the geographic location information when the image data is generated, as well as the background information (scene, color, etc.) of the image data; and the target application is prohibited from accessing the feature information of the human face included in the image data.

[0072] S303: Determine an access control policy for the target application based on the data access scope and data access permissions of the target application.

[0073] The access control policy is used to indicate data that the target application is allowed to access and / or data that the target application is prohibited from accessing.

[0074] In one possible implementation, the data access scope of the target application can be determined by semantically parsing the task request when the target application performs the AI task, thereby parsing the permissions declared in the task request of the AI task and the data that may be accessed.

[0075] For example, when the target application is an AI smart album, and the AI task is to intelligently classify all photos included in the album (for example, classification based on the location information when the photos were taken, or classification based on the color of the photos, etc.), the data access scope of the target application can be determined to be image data, without accessing data information of types such as contact data and short message data. In this way, based on the data access scope and data access rights of the target application, it is possible to determine the data that the target application is allowed to access and / or the data that the target application is prohibited from accessing.

[0076] For example, if the AI task is to classify based on the location information when the photo was taken, the data access permission can be determined to be a level 2 security level. Therefore, the target application can be allowed to access the geographic location information when the image data was generated and the background information of the image data, while the target application is prohibited from accessing the facial feature information included in the image data. If the AI task is to classify based on the color of the photo, the data access permission can be determined to be a level 3 security level. Therefore, the target application can be allowed to access the background information of the image data when the image data was generated, while the target application is prohibited from accessing the geographic location information when the image data was generated and the facial feature information included in the image data.

[0077] It should be noted that a classification model can be used to classify and label data that a target application may access. This allows images (photos) to be fed into a trained classification model, where the geographic location of each image when it was generated is determined and labeled, as well as the background information (color) of each image. Based on each image's label, it is then determined whether the target application is allowed to access the image, adjusting the target application's access permissions for each image and generating an access control policy for the target application.

[0078] In one possible implementation, an access control knowledge graph can be constructed based on the OWL (Web Ontology Language) ontology to determine the data range that the target application is allowed to access (such as the storage directory of data of the image data type) and the data range that the target application is prohibited from accessing (such as the contact data type (contact database), short message data type (SMS database)).

[0079] Specifically, after an access control policy (eg, a Security-Enhanced Linux (SE Linux) policy) of a target application is generated and dynamically loaded, the data range that the target application can access can be restricted based on the access control policy.

[0080] For example, after generating the access control policy of the target application, the access control policy of the target application can be compiled into an SE Linux module, and then the legitimacy of the SE Linux module can be verified by signature. If the SE Linux module is legal, the SE Linux module can be dynamically loaded into the kernel of the electronic device, thereby monitoring the access to data information when the target application performs a task based on the SE Linux module, and unloading the SE Linux module after the task is completed.

[0081] In an embodiment of the present disclosure, by obtaining the parameter gradient of the aggregation model corresponding to the target application, the data access rights corresponding to the target application can be determined to indicate the security level of each data type that the target application is allowed to access. In this way, when the security level of each data type that the target application is allowed to access is determined, the access control policy of the target application can be determined in combination with the data access scope of the target application. Thus, the target application's access to data is controlled by the data indicated by the access control policy, which data the target application is allowed to access, and / or the data that the target application is prohibited from accessing. In this way, based on the parameter gradient of the aggregation model corresponding to the target application, it is possible to accurately determine which data the target application is allowed to access, thereby accurately limiting the data that the target application can access when formulating the access control policy. This prevents the target application from excessively accessing other private data (i.e., data that the target application does not need to access) when accessing data, thereby avoiding the leakage of private data and improving the security of private data.

[0082] In some embodiments, as Figure 4 As shown, the access control method may further include: S401-S402.

[0083] S401 : Based on an access request of a target application to execute a target task, determine target data currently accessed by the target application.

[0084] In one possible implementation, when a target application executes a target task (AI task), it sends an access request to the electronic device, requesting access to the data information required to execute the target task. Therefore, the access request may include indication information indicating the target data currently being accessed by the target application. The electronic device can then determine the target data currently being accessed by the target application based on the indication information indicating the target data.

[0085] S402: Determine whether to allow the target application to access the target data based on the access control policy.

[0086] In one possible implementation, the AI dynamic sandbox can limit the target application's access to data based on the access control policy (the highest security level for each data type) to isolate, restrict, or cancel access to data that the target application is prohibited from accessing.

[0087] It should be noted that the AI dynamic sandbox performs access control based on data access within independent target applications. For a specific target application's data access rights, when the target application accesses different data, the data is divided into more fine-grained categories, and each data type is marked with a security level. When a target application accesses data of a certain data type, the security scope (security level) for accessing data of other data types is dynamically adjusted, thereby limiting the target application's excessive access (collection) of data and preventing the leakage of private data caused by the target application accessing irrelevant data information.

[0088] In one possible implementation, a differential privacy noise injection unit can be integrated at the NPU chip level to dynamically perturb the inference output. Furthermore, the F2FS file system can be modified to add a privacy tag field to the metadata. This allows for the activation of a virtual camera within the sandbox, which returns a blank image, immediately terminates access to the image, and records and reports the attack signature when a malicious application attempts to steal or access data (e.g., image data).

[0089] In one possible implementation, when it is determined that there is unauthorized or illegal access to the target data accessed by the target application (such as a request to turn on the camera), it is necessary to intercept the target application from accessing the prohibited data in the target data.

[0090] Optionally, the electronic device can modify the Android Binder driver and insert a permission check hook at the kernel layer. Specifically, the function entry and condition judgment are first required. When the Binder driver receives the input and output control (ioctl) system call request, it enters the Binder input and output control processing function. Check whether the command (cmd) parameter is a preset instruction (such as the camera turn-on instruction) to determine whether the current request involves access operations on the camera device. Enter the sandbox policy query, call the get current policy function (i.e., access control policy), and obtain the sandbox policy structure of the current process. Based on the allowed access device set field contained in the structure (expressed in the form of a binary bit mask), identify the set of device types that the current process is allowed to access. For example: if the value of the allowed access device set is binary 1010, it means that access to the devices corresponding to the 1st and 3rd bits is allowed (counting from the 0th bit). Finally, a permission bit check is performed, and a bit operation is used to check whether the camera device flag bit is included in the set of devices allowed to access. If access is allowed, the bit operation result is non-zero (that is, the policy allows access to the camera device), and subsequent operations are continued. If access is denied, the bit operation result is zero, and an error code indicating that the operation is not authorized (-EPERM) is returned to prevent the camera device from being turned on.

[0091] In one possible implementation, it is necessary to monitor the data scope accessed by the target application in real time. The access service of the target application to perform the target task can be intercepted by Hook, the service status of the target application to perform the target task can be detected, and the Linux control group (cgroups) can be used to limit the CPU quota of the background process to limit the resources occupied by the target application when performing the target task.

[0092] In the disclosed embodiments, when a target application executes a target task, the target data currently being accessed by the target application can be determined based on the access request. Based on the access control policy, it is determined whether the target application is allowed to access the target data. When the target application accesses data, it is determined whether the data being accessed is authorized based on the predetermined access control policy, thereby ensuring the security of private data.

[0093] In some embodiments, the access control method may further include: upon completion of the target task, deleting temporary data generated when executing the target task, and clearing the access control policy.

[0094] It is understood that after the target application completes its target task, the memory and temporary data can be erased. Specifically, a standard erase can be performed on the temporary data stored in the GPU memory, and a secure trim command can be triggered at the eMMC storage chip level to ensure physical data deletion.

[0095] In the embodiment of the present disclosure, after the target task is completed, by deleting the temporary data generated when executing the target task and clearing the access control policy, it is possible to further prevent the privacy data from being leaked and ensure the security of the privacy data.

[0096] In some embodiments, the parameter gradient of the aggregation model is used to indicate the highest security level of each data type that the target application needs to access; the above S302 may specifically include:

[0097] S3021. Determine the highest security level of each data type indicated by the parameter gradient of the aggregation model as the security level of each data type that the target application is allowed to access.

[0098] In a possible implementation, based on the highest security level of each data type indicated by the parameter gradient of the aggregation model, only data information corresponding to the highest security level may be determined as data information that is allowed to be accessed by the target application.

[0099] For example, for data of three security levels (level one security level, level two security level, and level three security level) corresponding to the image data type, assuming that the highest security level of the image data type indicated by the parameter gradient of the aggregation model is level two security level, the target application may only be allowed to access the data information corresponding to the level two security level (i.e., the geographic location information when the image data is generated), while the target application may be prohibited from accessing the data information corresponding to the level one security level (i.e., the feature information of the face included in the image data) and the data information corresponding to the level three security level (i.e., the background information of the image data).

[0100] In a possible implementation, based on the highest security level of each data type indicated by the parameter gradient of the aggregation model, data information with a security level lower than the highest security level may be determined as data information allowed to be accessed by the target application.

[0101] For example, for data of three security levels (level one security level, level two security level, and level three security level) corresponding to the image data type, assuming that the highest security level of the image data type indicated by the parameter gradient of the aggregation model is level two security level, the target application can be allowed to access the data information corresponding to the level two security level (i.e., the geographic location information when the image data is generated) and the data information corresponding to the level three security level (i.e., the background information of the image data), while the target application is only prohibited from accessing the data information corresponding to the level one security level (i.e., the feature information of the face included in the image data).

[0102] In the embodiment of the present disclosure, the highest security level of each data type indicated by the parameter gradient of the aggregation model is used to determine the data range that the target application is allowed to access. When the data accessed by the target application exceeds the corresponding data range, the target application can be restricted from accessing the data that exceeds the data range, thereby preventing the privacy data from being leaked and ensuring the security of the privacy data.

[0103] In some embodiments, the above S303 may specifically include:

[0104] S3031. Determine the data type of the data that the target application needs to access based on the data access scope of the target application.

[0105] In a possible implementation, the data access scope of the target application includes the data types of data that the target application needs to access when executing the target task.

[0106] In a possible implementation, the data access scope of the target application includes the specific data that the target application needs to access when executing the target task, so that the data type of the data that the target application needs to access can be determined based on the data that needs to be accessed.

[0107] S3032: Determine an access control policy for the target application based on the data type and data access permission of the data that the target application needs to access.

[0108] It can be understood that after determining the data type of data that the target application needs to access and the security level of each data type that the target application is allowed to access as indicated by the data access permission, the access control policy of the target application can be formulated.

[0109] In the disclosed embodiment, based on the determined data access scope of the target application, the data type that the target application needs to access when performing the target task can be determined first. Then, combined with the security level of each data type that the target application is allowed to access as indicated by the data access permission, the data corresponding to each data type that the target application is allowed to access can be determined, thereby accurately formulating the access control policy for the target application. This allows the target application to accurately control the data accessed by the target application when performing the target task, avoid excessive data access by the target application, prevent the leakage of private data, and ensure the security of private data.

[0110] In some embodiments, the data access rights include the highest security level corresponding to each of the multiple data types; the above S3032 can be specifically implemented by the following steps:

[0111] S3032a. Based on the data type of the data that the target application needs to access and the highest security level corresponding to each data type, determine the data that the target application is allowed to access and / or the data that the target application is prohibited from accessing from the data of each data type.

[0112] In one possible implementation, after determining the data type that the target application needs to access, based on the security level of each data type that the target application is allowed to access as indicated by the data access rights corresponding to the determined target application, the data that the target application is actually allowed to access (authorized data) can be determined from the data of each data type that the target application is allowed to access.

[0113] Furthermore, after determining the data that the target application is actually allowed to access (authorized data), an access control policy for the target application can be formulated based on the authorized data, thereby allowing the target application to access authorized data and prohibiting the target application from accessing unauthorized data.

[0114] In some embodiments, the aggregate model is obtained by:

[0115] S501: Obtain model training data corresponding to the target application and determine the security level of the model training data.

[0116] In one possible implementation, the basic model corresponding to the target application can be trained locally on the electronic device based on local data to obtain a trained intermediate model, and then the parameter gradient of the intermediate model can be uploaded to the server, so that the server can train the aggregate model based on the parameter gradients of multiple intermediate models uploaded by multiple electronic devices, determine the parameter gradient of the aggregate model, and return it to the electronic device.

[0117] In one possible implementation, the electronic device may obtain model training data corresponding to the target application from a local secure storage area, and obtain a basic model from a cloud server, so as to train the basic model using the model training data to obtain an intermediate model.

[0118] For example, taking the target application as an AI smart album, the electronic device can obtain album source data from the local secure data storage area as model training data.

[0119] Optionally, after obtaining the model training data corresponding to the target application, the security level of the model training data can be determined based on the privacy level of the data. For example, facial feature information included in the image data can be determined as level one security level, geographic location information when the image data was generated (shot) can be determined as level two security level, and background information of the image data can be determined as level three security level.

[0120] It should be noted that the local secure storage area of an electronic device can be divided into a secure partition (Secure World) and a normal partition (Normal World) at the hardware level based on ARM TrustZone. The secure partition is used to store sensitive data such as biometrics and location trajectories, while the normal partition is used to store non-sensitive data such as user behavior logs.

[0121] In one possible implementation, at the file system layer, the file system can be modified to add privacy tags to the metadata, that is, to determine the security level of each data information.

[0122] S502: Perform data desensitization processing on the model training data to obtain desensitized model training data.

[0123] In some embodiments, the model training data includes first data and second data, and the security level of the second data is higher than the security level of the first data. The above S502 can be specifically implemented by the following steps:

[0124] S5021. Eliminate the second data in the model training data to obtain desensitized model training data.

[0125] For example, when the model training data is image data, the image data includes: facial feature information, geographic location information when the image data was generated, background information of the image data, etc. Therefore, the facial feature information and geographic location information when the image data was generated included in the image data can be determined as the second data, and the background information of the image data can be determined as the first data.

[0126] Based on this, when training the base model so that the trained model can classify images based on color, the base model can be trained only based on the background information of the image data, without having to train the base model using the facial features included in the image data and the geographic location information when the image data was generated. Therefore, the second data in the model training data can be eliminated, and only the first data in the model training data can be retained, thereby obtaining the desensitized model training data.

[0127] In this way, by desensitizing the model training data, obfuscating the features of the original training data, and restricting access to data of other data types (such as contacts, text messages, etc.), only the required functional features (such as background information of image data) are retained, while the traceability features (such as the feature information of the face included in the image data) are destroyed.

[0128] S503: Based on the desensitized model training data, differential training is performed on the basic model to obtain an intermediate model.

[0129] S504: Determine an aggregation model based on the parameter gradient of the intermediate model.

[0130] In one possible implementation, during differential training of the base model, lightweight model training (MobileNetV3-Small) can be performed within the TEE, and differential training can be achieved by injecting Laplace noise.

[0131] Specifically, during differential training, sensitivity assessment is required to analyze and determine the impact (sensitivity) of changes in model training data on parameter gradients. Specifically, this involves the maximum possible fluctuation range of the base model's parameter gradients when a piece of data is added or removed from the model training data. Noise generation and parameter control are then performed, and the noise distribution characteristics are determined based on the sensitivity and the preset privacy protection strength (privacy budget).

[0132] It's important to note that a smaller privacy budget leads to higher protection levels and a larger noise amplitude. Noise follows a symmetrical distribution centered on zero, with the probability of its occurrence rapidly decreasing as its absolute value increases. This ensures that most noise is concentrated within a small range, effectively interfering with private information while avoiding excessive disruption to model training.

[0133] Furthermore, gradient noise protection is performed, where the generated noise is superimposed on the original parameter gradients of the underlying model, masking the contribution of a single data point to the parameter gradient. For example, if a data point increases a dimension of the parameter gradient by 0.3, the dimension may become 0.42 or -0.28 after adding noise, making it impossible for an attacker to distinguish whether the change in the parameter gradient is due to real data or random noise.

[0134] In one possible implementation, you can also use TensorFlow Lite's Micro Interpreter for memory optimization so that during the training of the basic model (i.e., the federated learning phase), the memory usage is controlled within a preset memory value (e.g., 50MB).

[0135] In some embodiments, the above S504 may be implemented by the following steps:

[0136] S5041. Send the parameter gradient of the intermediate model to the server.

[0137] The server is used to determine the aggregation model based on the parameter gradients of the obtained multiple intermediate models.

[0138] In one possible implementation, after the intermediate model is trained and the parameter gradient of the intermediate model is determined, the parameter gradient of the intermediate model needs to be homomorphically encrypted, and then the encrypted parameter gradient of the intermediate model is uploaded to the cloud server.

[0139] For example, a semi-homomorphic encryption algorithm (or a fully homomorphic encryption algorithm) can be used to process the parameter gradients of the intermediate model, thereby obtaining an encrypted parameter gradient c = (g^m*r^n) mod n², which is then uploaded to the server for decentralized gradient aggregation via a blockchain smart contract. Here, g represents the parameter gradient before encryption, m represents the key (public and / or private), r represents a random number, and n represents the key (public and / or private).

[0140] S5042. Obtain an aggregation model from the server.

[0141] In one possible implementation, the server can obtain the parameter gradients of the intermediate model obtained after each electronic device trains the basic model from multiple electronic devices, and then perform aggregate training on the model based on the obtained multiple parameter gradients to obtain an aggregate model.

[0142] In the disclosed embodiments, the parameter gradients of the intermediate models are sent to a server, allowing the server to train an aggregate model based on the parameter gradients of multiple intermediate models, and then return the parameter gradients of the aggregate model to the electronic device. Based on the parameter gradients of the aggregate model, the electronic device can more accurately determine the data access permissions corresponding to the target application.

[0143] In some embodiments, the desensitized model training data includes: first model training data and second model training data, and the intermediate model includes: a first intermediate model and a second intermediate model; the above S503 can be specifically implemented by the following steps:

[0144] S5031. Perform differential training on the basic model based on the first model training data to obtain a first intermediate model.

[0145] S5032. Perform differential training on the first intermediate model based on the second model training data to obtain a second intermediate model.

[0146] In one possible implementation, the desensitized model training data can be divided into multiple training data to train the basic model (or the trained intermediate model) multiple times.

[0147] That is, the basic model can be differentially trained using a set of training data (for example, the first model training data) to obtain a first intermediate model; then, based on a second set of training data (for example, the second model training data), the model obtained by the first training (that is, the first intermediate model) can be differentially trained again to obtain a second intermediate model.

[0148] It should be noted that, further, based on the third set of training data (for example, the third model training data), the model obtained by the second training (i.e., the second intermediate model) can be differentially trained again to obtain the third intermediate model, until all the training data are completed, and the final intermediate model can be obtained.

[0149] In the embodiment of the present disclosure, the final intermediate model is obtained by performing differential training on the basic model (or intermediate model) multiple times, which can improve the accuracy of the obtained intermediate model.

[0150] In some embodiments, the method may further include:

[0151] S601: Determine a corresponding gradient change between a parameter gradient of a first intermediate model and a parameter gradient of a second intermediate model.

[0152] S602: Based on the gradient change amount, adjust the highest access level of the target application to each data type.

[0153] In one possible implementation, in the process of differentially training the first intermediate model based on the second model training data to obtain the second intermediate model, when new training data is added to the model training (or there is an adversarial sample attack), the parameter gradient of the intermediate model may change significantly.

[0154] It should be noted that during model training, the gradient change ΔW of the intermediate model's parameter gradient is inversely proportional to the permission opening threshold (i.e., the highest security level). When the gradient change ΔW is large, it indicates that the model is still in a stage of intensive updating and may require more training data to promote model convergence. Therefore, the permission opening threshold is low, allowing higher permission levels (i.e., security levels) to be opened. Conversely, when the gradient change ΔW is small, it indicates that the model is tending to converge and permissions need to be tightened to reduce the risk of data leakage. At this time, the opening threshold is high and the permission opening level is lowered to avoid excessive opening of access rights to private data.

[0155] In one possible implementation, an adversarial sample generator can be deployed to inject interference signals into the input layer of the intermediate model (for example, adding adversarial noise patterns to the facial features included in the image). When a malicious application injects adversarial samples into the model (for example, adding special noise patterns) in an attempt to bypass detection or forge identity authentication, the model's recognition confidence for continuous input of the same face will jump from the normal value to a lower value. At this time, the gradient parameter in the federated learning process will increase, causing fluctuations. At the same time, when the target application (such as an AI smart album) accesses other secure data (such as contacts or location information), when abnormal fluctuations in the model output are detected, the circuit breaker mechanism can be automatically triggered to terminate permission authorization and dynamically revoke permissions to prevent data from leaving the local area.

[0156] In the embodiment of the present disclosure, during the model training process, based on the gradient change between the parameter gradients of the intermediate model obtained from two adjacent trainings, the highest access level of the target application to each data type can be accurately adjusted, thereby ensuring the security of privacy data.

[0157] It can be understood that the disclosed embodiment builds a privacy protection closed loop through the deep coupling of federated learning and dynamic permission dynamic sandbox. Localized federated learning is implemented in a trusted execution security environment, combined with differential privacy noise injection and homomorphic encryption technology to ensure that data does not need to leave the device and the transmission of gradient parameters is irreversible. Fine-grained access policies are dynamically generated based on the semantic parsing model, and hardware-level real-time permission interception is achieved by transforming the Android kernel driver (such as Binder). Data access boundaries are dynamically defined according to task requirements (such as limiting album subdirectories). Moreover, permissions are automatically reclaimed after the task is completed and NIST standard secure erase is triggered to completely clear residual data in GPU memory and storage chips. In addition, a gradient change and permission policy linkage mechanism is designed. When abnormal fluctuations in model parameters are detected, the data access scope is automatically reduced to balance privacy protection and model accuracy. Dynamic sandbox replaces static authorization architecture, TEE hardware isolation enhances traditional software protection, and active erasure defends against physical attacks. The contradiction between privacy protection and AI performance is systematically resolved, providing a complete technology chain of "data available but invisible" for high-sensitivity data scenarios.

[0158] It is understandable that, in order to implement the above functions, the electronic device includes hardware structures and / or software modules corresponding to the execution of each function. Those skilled in the art should easily realize that, in combination with the algorithm steps of each example described in the embodiments of the present disclosure, the present disclosure can be implemented in the form of hardware or a combination of hardware and computer software. Whether a function is executed in the form of hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present disclosure.

[0159] The embodiment of the present disclosure can divide the data sending end and the data receiving end into functional modules according to the above-mentioned method embodiment. For example, each functional module can be divided corresponding to each function, or two or more functions can be integrated into one functional module. The above-mentioned integrated module can be implemented in the form of hardware or software. It should be noted that the division of modules in the embodiment of the present disclosure is schematic and is only a logical function division. There may be other division methods in actual implementation. The following is an example of dividing each functional module corresponding to each function.

[0160] Figure 5This is a schematic diagram of the structure of an access control device provided by an embodiment of the present disclosure, which is applied to electronic devices. The access control device 500 can execute the above method embodiment. Figure 3 The access control method shown in Figure 5 As shown, the access control device 500 includes: an information acquisition module 501 and an access control module 502 .

[0161] Information acquisition module 501, used to obtain the parameter gradient of the aggregation model corresponding to the target application;

[0162] An access control module 502 is configured to determine data access permissions corresponding to a target application based on a parameter gradient of an aggregation model, where the data access permissions indicate a security level of each data type that the target application is allowed to access.

[0163] The access control module 502 is further configured to determine an access control policy for the target application based on the data access scope and data access rights of the target application. The access control policy is configured to indicate data that the target application is allowed to access and / or data that the target application is prohibited from accessing.

[0164] Optionally, the parameter gradient of the aggregation model is used to indicate the highest security level of each data type that the target application needs to access; the access control module 502 is specifically used to determine the highest security level of each data type indicated by the parameter gradient of the aggregation model as the security level of each data type that the target application is allowed to access.

[0165] Optionally, the access control module 502 is specifically used to determine the data type of data that the target application needs to access based on the data access scope of the target application; the access control module 502 is specifically used to determine the access control policy of the target application based on the data type of data that the target application needs to access and data access rights.

[0166] Optionally, data access permissions include the highest security level corresponding to each data type among multiple data types; the access control module 502 is specifically used to determine the data that the target application is allowed to access and / or the data that the target application is prohibited from accessing from the data of each data type based on the data type of the data that the target application needs to access and the highest security level corresponding to each data type.

[0167] Optionally, the access control module 502 is further used to determine the target data currently accessed by the target application based on the access request of the target application to perform the target task: the access control module 502 is further used to determine whether the target application is allowed to access the target data based on the access control policy.

[0168] Optionally, the access control module 502 is further configured to delete temporary data generated when executing the target task and clear the access control policy when the target task is completed.

[0169] Optionally, the information acquisition module 501 is also used to obtain the model training data corresponding to the target application and determine the security level of the model training data; the information acquisition module 501 is also used to perform data desensitization on the model training data to obtain desensitized model training data; the information acquisition module 501 is also used to perform differential training on the basic model based on the desensitized model training data to obtain an intermediate model; the information acquisition module 501 is also used to determine the aggregation model based on the parameter gradient of the intermediate model.

[0170] Optionally, the model training data includes first data and second data, and the security level of the second data is higher than the security level of the first data; the information acquisition module 501 is specifically used to eliminate the second data in the model training data to obtain desensitized model training data.

[0171] Optionally, the information acquisition module 501 is specifically used to send the parameter gradient of the intermediate model to the server, and the server is used to determine the aggregation model based on the obtained parameter gradients of multiple intermediate models; the information acquisition module 501 is specifically used to obtain the aggregation model from the server.

[0172] Optionally, the desensitized model training data includes: first model training data and second model training data, and the intermediate model includes: a first intermediate model and a second intermediate model; the information acquisition module 501 is specifically used to perform differential training on the basic model based on the first model training data to obtain a first intermediate model; the information acquisition module 501 is specifically used to perform differential training on the first intermediate model based on the second model training data to obtain a second intermediate model.

[0173] Optionally, the access control module 502 is also used to determine the corresponding gradient change between the parameter gradient of the first intermediate model and the parameter gradient of the second intermediate model; the access control module 502 is also used to adjust the highest access level of the target application to access each data type based on the gradient change.

[0174] In the case of implementing the functions of the above-mentioned integrated modules in the form of hardware, the embodiment of the present disclosure provides a structural diagram of an electronic device of the access control device involved in the above-mentioned embodiment. Figure 6 As shown, the electronic device 600 includes a memory 601 and a processor 602. Optionally, the electronic device 600 may further include a communication interface 603 and a bus 604.

[0175] The memory 601 may be a read-only memory (ROM) or other type of static storage device that can store static information and instructions, a random access memory (RAM) or other type of dynamic storage device that can store information and instructions, or an electrically erasable programmable read-only memory (EEPROM), a disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto.

[0176] Processor 602 may implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the embodiments of this disclosure. Processor 602 may be a central processing unit, a general-purpose processor, a digital signal processor, an application-specific integrated circuit, a field-programmable gate array, or other programmable logic device, a transistor logic device, a hardware component, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the embodiments of this disclosure. Processor 602 may also be a combination that implements computing functions, such as a combination of one or more microprocessors, or a combination of a DSP and a microprocessor.

[0177] The communication interface 603 is used to connect to other devices via a communication network, such as Ethernet, wireless access network, wireless local area network (WLAN), etc.

[0178] The bus 604 may be an extended industry standard architecture (EISA) bus, etc. The bus 604 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 6 Only one thick line is used in the diagram, but this does not mean that there is only one bus or one type of bus.

[0179] As a possible implementation, the memory 601 can exist independently of the processor 602. The memory 601 can be connected to the processor 602 via a bus 604 to store instructions or program codes. When the processor 602 calls and executes the instructions or program codes stored in the memory 601, the access control method provided in the embodiment of the present disclosure can be implemented.

[0180] In another possible implementation, the memory 601 may also be integrated with the processor 602 .

[0181] Some embodiments of the present disclosure provide a computer-readable storage medium (e.g., a non-transitory computer-readable storage medium) storing computer program instructions. When the computer program instructions are executed on a computer, the computer executes an access control method as described in any of the above embodiments.

[0182] Exemplarily, the above-mentioned computer-readable storage media may include, but are not limited to: magnetic storage devices (e.g., hard disks, floppy disks, or magnetic tapes, etc.), optical disks (e.g., compact disks (CDs), digital versatile disks (DVDs), etc.), smart cards, and flash memory devices (e.g., erasable programmable read-only memories (EPROMs), cards, sticks, or key drives, etc.). The various computer-readable storage media described in the present disclosure may represent one or more devices and / or other machine-readable storage media for storing information. The term "machine-readable storage medium" may include, but is not limited to, wireless channels and various other media capable of storing, containing, and / or carrying instructions and / or data.

[0183] An embodiment of the present disclosure provides a computer program product containing instructions. When the computer program product is run on a computer, the computer is enabled to execute the access control method of any one of the above embodiments.

[0184] The above are only specific embodiments of the present disclosure, but the scope of protection of the present disclosure is not limited thereto. Any changes or replacements within the technical scope disclosed in the present disclosure should be included in the scope of protection of the present disclosure. Therefore, the scope of protection of the present disclosure should be based on the scope of protection of the claims.

Claims

1. An access control method, characterized in that: The method comprises: Obtain the parameter gradient of the aggregation model corresponding to the target application; determining, based on a parameter gradient of the aggregation model, a data access permission corresponding to the target application, the data access permission being used to indicate a security level of each data type that the target application is allowed to access; An access control policy of the target application is determined based on the data access scope of the target application and the data access permission, where the access control policy is used to indicate data that the target application is allowed to access and / or data that the target application is prohibited from accessing.

2. The method according to claim 1, characterized in that The parameter gradient of the aggregation model is used to indicate the highest security level of each data type that the target application needs to access; The determining, based on the parameter gradient of the aggregation model, the data access permission corresponding to the target application includes: The highest security level of each data type indicated by the parameter gradient of the aggregation model is determined as the security level of each data type that the target application is allowed to access.

3. The method according to claim 1, characterized in that The determining the access control policy of the target application based on the data access scope and the data access permission of the target application includes: Determining the data type of the data that the target application needs to access based on the data access scope of the target application; An access control policy for the target application is determined based on the data type of the data that the target application needs to access and the data access permission.

4. The method according to claim 3, characterized in that The data access permission includes the highest security level corresponding to each data type among the multiple data types; The determining of the access control policy of the target application based on the data type of the data that the target application needs to access and the data access permission includes: Based on the data type of the data that the target application needs to access and the highest security level corresponding to each data type, data that the target application is allowed to access and / or data that the target application is prohibited from accessing is determined from the data of each data type.

5. The method according to claim 1, wherein The method further comprises: Determining target data currently accessed by the target application based on an access request for executing a target task by the target application: Based on the access control policy, it is determined whether the target application is allowed to access the target data.

6. The method according to claim 5, characterized in that The method further comprises: When the target task is completed, the temporary data generated when the target task is executed is deleted, and the access control policy is cleared.

7. The method according to claim 1, characterized in that The aggregation model is obtained in the following way: Obtaining model training data corresponding to the target application and determining a security level of the model training data; Performing data desensitization processing on the model training data to obtain desensitized model training data; Based on the desensitized model training data, differential training is performed on the basic model to obtain an intermediate model; The aggregate model is determined based on the parameter gradient of the intermediate model.

8. The method according to claim 7, characterized in that The model training data includes first data and second data, and the security level of the second data is higher than the security level of the first data; The performing data desensitization processing on the model training data to obtain desensitized model training data includes: The second data in the model training data is eliminated to obtain the desensitized model training data.

9. The method according to claim 7, characterized in that The determining of the aggregation model based on the parameter gradient of the intermediate model includes: Sending the parameter gradient of the intermediate model to a server, wherein the server is configured to determine the aggregate model based on the obtained parameter gradients of the multiple intermediate models; The aggregation model is obtained from the server.

10. The method according to claim 7, characterized in that The desensitized model training data includes: first model training data and second model training data, and the intermediate model includes: a first intermediate model and a second intermediate model; The method of performing differential training on the basic model based on the desensitized model training data to obtain an intermediate model includes: Performing the differential training on the basic model based on the first model training data to obtain the first intermediate model; The first intermediate model is subjected to the differential training based on the second model training data to obtain the second intermediate model.

11. The method according to claim 10, characterized in that The method further comprises: Determining a corresponding gradient change between a parameter gradient of the first intermediate model and a parameter gradient of the second intermediate model; Based on the gradient change amount, the highest access level of the target application to each data type is adjusted.

12. An electronic device, characterized in that: include: memory and processor; Memory and processor coupling; The memory is used to store instructions executable by the processor; When the processor executes the instructions, the method according to any one of claims 1 to 11 is performed.

13. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer instructions, and when the computer instructions are executed on a computer, the computer is caused to execute the method according to any one of claims 1 to 11.

14. A computer program product, characterized in that When the computer program product is run on a computer, the computer is caused to perform the method according to any one of claims 1 to 11.

Citation Information

Cited By

  • Identity authentication strategy determination method and device, electronic equipment, medium and product

    CN121508893A