Method and apparatus for managing rights

Abstract

The application discloses a permission management method and device, and relates to the technical field of artificial intelligence. The method comprises the following steps: in the case that an AI agent triggers a first operation request, the first operation request is intercepted, and multi-dimensional context features corresponding to a terminal and request feature data of the first operation request are obtained; according to the multi-dimensional context features and the request feature data, an initial risk level corresponding to the first operation request is determined through a preset model; and the first operation request is subjected to permission management according to a permission control strategy corresponding to the initial risk level; wherein the multi-dimensional context features comprise at least two of the following: spatiotemporal context features, user context features, device context features and social context features.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence technology, specifically relating to a permission management method and device. Background Technology

[0002] With the increasing prevalence of agentic AI, AI agents can autonomously break down tasks and access system resources (such as location, network, and payment) across applications, thereby improving the intelligence level of electronic devices. However, agentic AI also brings new challenges to the permission management of electronic devices.

[0003] Currently, electronic devices can manage the operation permissions of AI agents through the static pop-up authorization mechanism built into the operating system. Specifically, when an AI agent requests sensitive resources, the kernel of the electronic device can intercept the AI ​​agent's operation request and pop up a dialog box, allowing the user to manually choose whether to authorize it.

[0004] However, the above solution has two drawbacks. First, frequent pop-ups can easily cause user "alarm fatigue," creating security vulnerabilities. Second, the fixed permission judgment mechanism of the static pop-up mechanism on the client side leads to low accuracy in authorizing AI agents. Therefore, how to manage AI agents securely, efficiently, and accurately is an urgent problem to be solved. Summary of the Invention

[0005] The purpose of this application is to provide a permission management method and apparatus that can manage AI agents securely, efficiently and accurately.

[0006] In a first aspect, embodiments of this application provide a permission management method, which includes: intercepting the first operation request when an AI agent triggers it, and obtaining multi-dimensional context features corresponding to the terminal and request feature data of the first operation request; determining the initial risk level corresponding to the first operation request through a preset model based on the multi-dimensional context features and request feature data; and managing the permission of the first operation request according to the permission control policy corresponding to the initial risk level; wherein the multi-dimensional context features include at least two of the following: spatiotemporal context features, user context features, device context features, and social context features.

[0007] Secondly, embodiments of this application provide a permission management method, which includes: receiving a risk assessment request sent by a terminal, the risk assessment request being used to request a risk assessment of a first operation request, the risk assessment request including: continuous operation logs of an AI agent within a first historical time period, request feature data of the first operation request, and multi-dimensional context features corresponding to the terminal; assessing the cloud risk assessment result corresponding to the first operation request based on the continuous operation logs, request feature data, and multi-dimensional context features; and sending the cloud risk assessment result to the terminal; wherein the multi-dimensional context features include at least two of the following: spatiotemporal context features, user context features, device context features, and social context features.

[0008] Thirdly, embodiments of this application provide a permission management device, which includes: a processing module; the processing module is configured to: intercept the first operation request when an artificial intelligence (AI) agent triggers the first operation request, and obtain the multi-dimensional context features corresponding to the terminal and the request feature data of the first operation request; determine the initial risk level corresponding to the first operation request through a preset model based on the multi-dimensional context features and the request feature data; and perform permission management on the first operation request according to the permission control strategy corresponding to the initial risk level; wherein the multi-dimensional context features include at least two of the following: spatiotemporal context features, user context features, device context features, and social context features.

[0009] Fourthly, embodiments of this application provide a permission management device, which includes: a receiving module, a processing module, and a sending module. The receiving module is used to receive a risk assessment request sent by a terminal. The risk assessment request requests a risk assessment of a first operation request, and includes: continuous operation logs of an AI agent within a first historical time period, request feature data of the first operation request, and multi-dimensional context features corresponding to the terminal. The processing module is used to assess the cloud-based risk assessment result corresponding to the first operation request based on the continuous operation logs, request feature data, and multi-dimensional context features in the risk assessment request received by the receiving module. The sending module is used to send the cloud-based risk assessment result processed by the processing module to the terminal. The multi-dimensional context features include at least two of the following: spatiotemporal context features, user context features, device context features, and social context features.

[0010] Fifthly, embodiments of this application provide a terminal including a processor and a memory, the memory storing programs or instructions executable on the processor, the programs or instructions, when executed by the processor, implementing the steps of the method described in the first aspect.

[0011] In a sixth aspect, embodiments of this application provide a server including a processor and a memory, the memory storing programs or instructions executable on the processor, the programs or instructions, when executed by the processor, implementing the steps of the method described in the second aspect.

[0012] In a seventh aspect, embodiments of this application provide a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method described in the first or second aspect.

[0013] Eighthly, embodiments of this application provide a chip, the chip including a processor and a communication interface, the communication interface being coupled to the processor, the processor being used to run programs or instructions to implement the methods described in the first or second aspect.

[0014] Ninthly, embodiments of this application provide a computer program product stored in a storage medium, which is executed by at least one processor to implement the method as described in the first or second aspect.

[0015] In this embodiment, when an AI agent triggers a first operation request, the first operation request can be intercepted, and the multi-dimensional context features corresponding to the terminal and the request feature data of the first operation request can be obtained. Based on the multi-dimensional context features and the request feature data, an initial risk level corresponding to the first operation request is determined through a preset model. Based on the first permission control policy corresponding to the initial risk level, permission management is performed on the first operation request. The multi-dimensional context features include at least two of the following: spatiotemporal context features, user context features, device context features, and social context features. In this solution, when the first operation request of the AI ​​agent is intercepted, since the risk level corresponding to the first operation request can be dynamically determined through a preset model based on the multi-dimensional context features corresponding to the terminal and the request data features of the first operation request, compared to a static pop-up permission mechanism where the user decides whether to grant permission, the permission management method provided in this embodiment can improve the dynamism, accuracy, and intelligence of permission management. Attached Figure Description

[0016] Figure 1 This is one of the flowcharts illustrating a permission management method provided in an embodiment of this application;

[0017] Figure 2 This is one of the flowcharts illustrating a permission management method provided in an embodiment of this application;

[0018] Figure 3 This is one of the flowcharts illustrating a permission management method provided in an embodiment of this application;

[0019] Figure 4 This is one of the structural schematic diagrams of a permission management device provided in the embodiments of this application;

[0020] Figure 5 This is a second schematic diagram of the structure of a permission management device provided in an embodiment of this application;

[0021] Figure 6 This is a schematic diagram of the structure of a communication device provided in an embodiment of this application;

[0022] Figure 7 This is a schematic diagram of the structure of a terminal provided in an embodiment of this application. Detailed Implementation

[0023] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0024] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0025] The terms "at least one," "at least one of," etc., used in the specification and claims of this application refer to any one, any two, or a combination of two or more of the included items. For example, at least one of a, b, and c can mean: "a," "b," "c," "a and b," "a and c," "b and c," and "a, b, and c," where a, b, and c can be single or multiple. Similarly, "at least two" refers to two or more items, and its meaning is similar to that of "at least one."

[0026] The nouns and terms used in the embodiments of this application are explained below.

[0027] AI Agent: An intelligent system with autonomous perception, decision-making, and execution capabilities, capable of dynamically planning paths to achieve its goals.

[0028] On-Device: Computation and processing are performed directly on the terminal, offering advantages such as low latency, privacy protection, and offline availability.

[0029] Context: A collection of environmental, user, and device states related to a mobile terminal, encompassing multi-dimensional information such as time and space, users, devices, and social interactions.

[0030] Task Intent: The ultimate goal that the AI ​​agent wants to achieve, and it is the core basis for judging the rationality of the AI ​​agent's behavior.

[0031] Action Chain: A series of consecutive (usually across applications) actions performed by an AI agent to complete a task.

[0032] Edge-cloud collaboration: Computing tasks are intelligently allocated between the edge and the cloud, with the edge responsible for routine low-risk verification and the cloud responsible for complex high-risk in-depth analysis.

[0033] The permission management method and apparatus provided in this application will be described in detail below with reference to the accompanying drawings, through specific embodiments and application scenarios.

[0034] The permission management method provided in this application embodiment is applied to scenarios involving permission management of AI agents.

[0035] In the permission management method provided in this application embodiment, when an AI agent triggers a first operation request, the first operation request can be intercepted, and the multi-dimensional context features corresponding to the terminal and the request feature data of the first operation request can be obtained. Based on the multi-dimensional context features and the request feature data, the initial risk level corresponding to the first operation request is determined through a preset model. Permission management is then performed on the first operation request according to the first permission control strategy corresponding to the initial risk level. The multi-dimensional context features include at least two of the following: spatiotemporal context features, user context features, device context features, and social context features. In this solution, when the first operation request of the AI ​​agent is intercepted, since the risk level corresponding to the first operation request can be dynamically determined through a preset model based on the multi-dimensional context features corresponding to the terminal and the request data features of the first operation request, compared to the static pop-up permission mechanism where the user decides whether to grant permission, the permission management method provided in this application embodiment can improve the dynamism, accuracy, and intelligence of permission management.

[0036] The execution subject of the permission management method provided in this application embodiment is a permission management device, which can be an electronic device, or a functional module or entity in an electronic device. This application embodiment does not limit this. The following will use a permission management device as an example to illustrate the permission management method provided in this application embodiment.

[0037] This application provides a permission management method, such as... Figure 1 As shown, the permission management method may include the following steps 101 to 103.

[0038] Step 101: When the AI ​​agent triggers the first operation request, the terminal intercepts the first operation request and obtains the multi-dimensional context features of the terminal and the request feature data of the first operation request.

[0039] In some embodiments of this application, the terminal can deploy an interception component at the operating system layer to intercept operation requests triggered by the AI ​​agent in real time. It should be noted that intercepting an operation request means that the security module actively captures and prevents the operation request from continuing to transmit before it reaches the target resource (such as a file, network, or sensor). After interception, the operation request can be analyzed, modified, rejected, or allowed. Interception is a proactive control action that emphasizes "interrupting" the operation flow and does not automatically retain the ability to recover later.

[0040] In some embodiments of this application, the first operation request can be any operation request initiated by the AI ​​agent. In other words, for each operation request initiated by the AI ​​agent, the terminal can perform interception, risk assessment, and decide whether to allow it based on the risk assessment result.

[0041] In some embodiments of this application, the multidimensional context features corresponding to the terminal include at least two of the following: spatiotemporal context features, user context features, device context features, and social context features.

[0042] For example, spatiotemporal context features may include at least one of the following: the terminal's GPS geographic location, precise timestamp, and motion acceleration information obtained from the gyroscope.

[0043] For example, user context features may include at least one of the following: screen on / off state, and active application in the foreground.

[0044] For example, device context features may include at least one of the following: remaining battery percentage, network connection type.

[0045] For example, social context features can be features that reflect a user's social habits, such as the time period during which they contact friends, the communication tools used to contact friends, and the dating apps they habitually browse.

[0046] For example, social context features can also be a class of features that can reflect information such as the relationship attributes, interaction patterns, trust levels, and social network structure between the AI ​​agent and its interactive objects.

[0047] In some embodiments of this application, the multi-dimensional context features corresponding to the terminal can be the multi-dimensional context features corresponding to the terminal within the most recent second time period. The first time period can be 5 minutes, 10 minutes, 15 minutes, 3 minutes, etc., and can be determined according to actual usage requirements; this application does not limit the specific duration.

[0048] For example, the spatiotemporal context may include: the terminal's GPS geographic location information within the last 5 minutes, the absolute time point within the last 5 minutes, such as 10:30~10:35, and motion acceleration information within the last 5 minutes.

[0049] Thus, since spatiotemporal context features, user context features, device context features, and social context features can assess the terminal's environmental state from different dimensions, the preset model set in the terminal can more accurately assess the risk of the first operation request, thereby enabling more accurate and dynamic management of the AI ​​agent's permissions.

[0050] In some embodiments of this application, the "obtaining the multi-dimensional context features corresponding to the terminal" in step 101 above may include the following steps: the terminal calls the terminal's sensor and system status manager in parallel to collect the multi-dimensional context information corresponding to the terminal, and performs feature encoding on the multi-dimensional context information to obtain a multi-dimensional context feature vector of a unified dimension.

[0051] In some embodiments of this application, the terminal can use feature engineering algorithms to preprocess the multidimensional context information, and then perform feature encoding on the preprocessed multidimensional context information to obtain a multidimensional context feature vector of a unified dimension.

[0052] For example, the terminal can use feature engineering algorithms to perform Min-Max normalization on continuous numerical data and One-hot encoding on discrete categorical data. Then, through embedding layer fusion, the preprocessed multidimensional context information is losslessly transformed into a fixed-length (e.g., 128-dimensional) tensor feature vector.

[0053] The following example demonstrates how a terminal obtains its corresponding multi-dimensional contextual features in a real-world scenario.

[0054] Example 1: The system backend interface simultaneously collected the following data: the current time is 02:30 AM, the GPS coordinates are located in the user-defined home address area, the gyroscope shows the device has been stationary for a long time, and the device's current state is locked, black screen, and charging. These physical parameters, after being encoded in a dimensionality reduction manner, generate feature vectors that, in a mathematical space model, extremely accurately represent the specific scenario environment in which "the user is very likely to be in deep sleep."

[0055] In some embodiments of this application, the "obtaining the request feature data of the first operation request" in step 101 above can be achieved through the following step 1B:

[0056] Step 1B: The terminal performs feature extraction processing on the execution context information corresponding to the first operation request to obtain request feature data.

[0057] The aforementioned request feature data may include at least one of the following: agent identity information, execution parameters, and task intent information.

[0058] The execution parameters can be the call parameters or operation parameters carried by the AI ​​agent when it initiates the first operation request. These call parameters can include the operation type and the operation object.

[0059] For example, a user gives a voice command: "Book me a flight to Beijing tomorrow." The AI ​​agent, after breaking down the task in the background, attempts to make an unauthorized silent background call to the system interface to read the SMS message and obtain the booking verification code. The underlying probe accurately captures this unauthorized request, and the parsed request feature data set is as follows: Agent identity: Travel Assistant; Task description (i.e., task intent information): "Book a flight to Beijing"; Operation target: READ_SMS; Operation parameter: Read the latest verification code SMS message.

[0060] In some embodiments of this application, the aforementioned task intent information is used to indicate the natural language task intent of the first operation request. The task intent information can be obtained by parsing the user input content in the execution context information corresponding to the first operation request.

[0061] In some embodiments of this application, the execution context information corresponding to the first operation request refers to a set of information stored in the execution stack of the AI ​​agent at the moment the AI ​​agent initiates the first operation request, which describes the state of the AI ​​agent when the first operation request is initiated, including but not limited to agent identity, calling parameters, user input content, etc.

[0062] In some embodiments of this application, the terminal can deploy an interception component at the terminal's operating system layer to intercept operation requests triggered by the AI ​​agent in real time, and extract and parse the execution context information corresponding to the intercepted operation requests to obtain the request feature data of the operation requests.

[0063] Specifically, the terminal can deploy interception probes at the terminal's operating system kernel layer or application framework layer to intercept operation requests initiated by AI agents, and extract features from the execution context information corresponding to the intercepted operation requests to obtain request feature data containing agent identity, operation target, underlying execution parameters and natural language task intent.

[0064] For example, the underlying interception mechanism uses Linux kernel-level Extended Berkeley PacketFilter (eBPF) technology or the Android system's Xposed framework to implement seamless hooking interception of application programming interfaces (APIs) that are critical to the operating system, such as accessing the network, reading files, and sending SMS messages. The parsing module then utilizes device-side Natural Language Processing (NLP) technology to extract and understand the declared natural language task intent from the AI ​​agent's runtime stack.

[0065] The following example demonstrates how a terminal obtains request feature data for the first operation request, using a real-world scenario.

[0066] Example 2: Real-world scenario: A user gives a voice command: "Book me a flight to Beijing tomorrow." The AI ​​agent, after breaking down the task in the background, attempts an unauthorized, silent background call to the system interface to read the SMS message and obtain the booking verification code. The underlying probe accurately captures this unauthorized request, and the parsed request feature data set is as follows: Agent identity: Travel Assistant; Task description: "Book a flight to Beijing"; Operation target: READ_SMS; Operation parameter: Read the latest verification code SMS message.

[0067] Thus, since at least one of the task intent, agent identity, and execution parameters of the operation request triggered by the AI ​​agent can be extracted as the request feature parameters of the operation request, the security of the operation request can be evaluated from the perspectives of agent identity, intent, and execution parameters, thereby improving the comprehensiveness and accuracy of permission management for the AI ​​agent.

[0068] Step 102: Based on the multi-dimensional context features and the request feature data, the terminal determines the initial risk level corresponding to the first operation request using a preset model.

[0069] In some embodiments of this application, the aforementioned preset model can be a lightweight risk assessment model deployed within the terminal.

[0070] In some embodiments of this application, the lightweight risk assessment model mounted on the device side adopts a miniaturized pre-trained language model based on the Transformer architecture. Furthermore, by employing knowledge distillation and INT8 model weight compression and quantization techniques, the size limit of the miniaturized and trained language model is compressed to less than 50MB, ensuring that local inference can be completed on mainstream mobile chips with an extremely low latency of less than 100ms.

[0071] Understandably, the core logic of the lightweight risk assessment model lies in using a cross-attention mechanism to calculate the logical correlation between the "operational goal" requested by the AI ​​agent and the "task intent" declared by the user, and to keenly investigate whether there is a serious application conflict between the action and the "multi-dimensional contextual features" currently corresponding to the terminal.

[0072] In some embodiments of this application, step 102 may include steps 102A and 102B as described below.

[0073] Step 102A: The terminal inputs the request feature data and the multi-dimensional context feature vector into a lightweight risk assessment model pre-deployed locally on the terminal for inference calculation to obtain the initial risk score of the first operation request.

[0074] Step 102B: The terminal will logically compare the initial risk score with the preset dynamic risk classification threshold to obtain the initial risk level determination result for the first operation request, i.e., the initial risk level.

[0075] In some embodiments of this application, the request feature data is jointly concatenated with the multidimensional context feature vector, and the concatenated features are input into a lightweight risk assessment model pre-deployed locally on the terminal for inference calculation to obtain an initial risk score that measures the degree of risk of the current operation.

[0076] Then, through the strategy diversion decision engine, the initial risk score is compared with the multi-level risk judgment threshold dynamically set by the system to obtain the initial risk level judgment result of low risk, medium risk or high risk, so that the terminal can execute the hierarchical response mechanism corresponding to the initial risk level judgment result for the first operation request.

[0077] The following example demonstrates how to determine the initial risk level corresponding to the first operation request, using a real-world scenario.

[0078] Example 3, following Example 2 above, combines the specific action of "booking a flight and reading SMS messages" with the contextual feature of "2:30 AM and the user being in deep sleep" into the edge model. Through cross-attention inference, the model discovers that in the middle of the night, without the user's awareness or active triggering, the AI ​​automatically reads private SMS messages and executes the ticket booking and payment action, severely deviating from normal human sleep patterns and safety rules, posing a very high risk of loss of control. Therefore, the output calculation result is: initial risk score = 0.85 (falling into the high-risk range), and the risk category is determined as "high-risk of unexpected charges and privacy theft".

[0079] Step 103: The terminal performs permission management on the first operation request mentioned above according to the permission control policy corresponding to the initial risk level.

[0080] In some embodiments of this application, the access control policy corresponding to the above-mentioned initial risk level may include any of the following:

[0081] Strategy 1: Allow low-risk operation requests;

[0082] Strategy 2: Allow users to decide whether to grant operation requests at the medium-risk level;

[0083] Strategy 3: For high-risk operation requests, implement the permission management policy corresponding to the cloud risk assessment results determined by the server.

[0084] In some embodiments of this application, when the initial risk level is medium risk, the terminal can display a prompt window, which includes a prompt message, an "Allow" control, and a "Deny" control. The prompt message is used to prompt the user whether to allow the first operation request, the "Allow" control is used to confirm allowing the first operation request, and the "Deny" control is used to confirm terminating the first operation request.

[0085] It is understood that this application adopts an edge-cloud collaborative permission management architecture. Specifically, a lightweight risk assessment model is deployed on the edge to handle high-frequency, low-risk operation requests, while the risk assessment of complex, high-risk operation requests is only delegated to the cloud. This achieves extremely low latency, such as less than 100ms, while ensuring offline availability. Furthermore, because the lightweight risk assessment model has a small computational load, it ensures that the increase in terminal power consumption is extremely low, typically less than or equal to 5%.

[0086] In this way, since the AI ​​agent's operation requests can be automatically allowed when the client-side assessment is low-risk, allowed by the user when the risk is medium-risk, and then submitted to the server for a second assessment when the risk level is high, low latency and offline availability can be guaranteed, and the accuracy of permission management for the AI ​​agent can be ensured.

[0087] It is understood that this application can use a lightweight risk assessment model deployed on the client side to assess the risk level of an operation request based on the request feature data of the operation request triggered by the AI ​​agent and the multi-dimensional context features corresponding to the terminal. Based on the client-side assessment results, it can execute a hierarchical permission management response to the operation request, thereby filling the gap in effective, real-time, and intelligent permission management of autonomous decision-making AI agents in mobile resource-constrained environments.

[0088] In the permission management method provided in this application embodiment, a lightweight risk assessment model deployed on the terminal side is used to handle high-frequency, low-risk requests, while only complex, high-risk operations are handled by the cloud. That is, through the terminal-cloud collaborative AI agent permission management architecture, the lightweight risk assessment model is deployed on the terminal side, multi-dimensional context real-time analysis, intent understanding, and cross-application operation chain security analysis are realized.

[0089] In some embodiments of this application, the terminal can use the policy triage decision engine to compare the initial risk score with the multi-level risk judgment threshold dynamically set by the system to obtain the level judgment result of low risk, medium risk or high risk, and trigger the execution of the corresponding graded response mechanism.

[0090] In some embodiments of this application, the system internally sets a first risk threshold (e.g., 0.2) and a second risk threshold (e.g., 0.7), and executes the following three control branch flows:

[0091] (1) Determined as low risk (initial risk score < 0.2): A clearance instruction is received, triggering the green channel mechanism. The underlying system directly grants clearance and silently records security audit logs in the background, without disturbing the user at all. For example: When the AI ​​agent calls the weather API to obtain the temperature during daytime working hours in a secure Wi-Fi environment at the company, the model score is 0.05, and clearance is automatically granted without being noticed.

[0092] (2) If the risk level is determined to be medium (0.2 ≤ initial risk score < 0.7): A confirmation instruction is received, triggering the yellow warning mechanism. The system temporarily blocks the operation and requests the user to manually review and confirm via a pop-up window in the front-end system UI.

[0093] For example, when an AI agent requests to use a third-party payment component to buy coffee for a user, it receives a score of 0.55. A pop-up message appears asking the user, "Your AI assistant has requested to use the payment component, which is expected to cost 35 yuan. Do you want to allow it? Options: [Allow this time] / [Always allow] / [Deny]." This process then proceeds to step 107, which collects feedback to close the loop.

[0094] (3) Determined as high risk (initial risk score ≥ 0.7): A high-risk interception instruction is received. The system immediately and forcibly intercepts and freezes the suspicious request, while automatically initiating a cloud-based high-computing-power in-depth analysis process. For example, for the aforementioned behavior of stealing SMS messages for late-night ticket booking with a score of 0.85, the terminal can freeze the operation request for stealing SMS messages and send a risk assessment request to the server.

[0095] In some embodiments of this application, step 103 described above can be implemented by steps 103A to 103C described below.

[0096] Step 103A: If the initial risk level is high, the terminal freezes the first operation request and sends a risk assessment request to the server.

[0097] The aforementioned risk assessment request can be used to request a risk assessment of the first operation request. The risk assessment request includes the historical API call logs of the AI ​​agent within the first historical time period, the request feature data, and the multidimensional context features.

[0098] Step 103B: The terminal receives the cloud-based risk assessment results sent by the server.

[0099] Step 103C: The terminal executes the permission control policy corresponding to the cloud risk assessment result for the first operation request.

[0100] In some embodiments of this application, the access control policy corresponding to the above-mentioned cloud risk assessment results may include any of the following:

[0101] Allow low-risk operation requests;

[0102] Forcefully terminate high-risk operation requests and output security warning information;

[0103] The execution parameters of operation requests at the medium-risk level are downgraded before being allowed.

[0104] In some embodiments of this application, when the terminal assesses the first operation request as high-risk, the terminal can package and send the recently intercepted continuous operation sequence logs, current request feature data and context features to the cloud server through a high-strength encrypted communication channel. The large language model and graph neural network of the cloud computing power are used to perform deep intent tracing and cross-application operation chain detection to obtain accurate cloud-based deep assessment decision instructions.

[0105] Specifically, given the limited computing power and field of view of the lightweight risk assessment model on the terminal, while the server deploys a large-scale language model with tens of billions of parameters to handle extremely complex long texts and deep spoofing logic, the specific assessment process on the server side is described in the relevant description in the server-side method embodiments below. To avoid repetition, it will not be repeated here.

[0106] In some embodiments of this application, the terminal can receive the cloud risk assessment result sent by the server through the terminal's end-to-cloud collaborative communication interface, parse the cloud risk assessment result, and, based on the parsed content instructions, execute the corresponding forced termination, downgrade modification, or confirmation release action on the first operation request through the system's underlying security framework to obtain the final physical execution result of the first operation request.

[0107] For example, (1) the cloud risk assessment result confirms that the first operation request is a high-risk request: if the terminal receives a cloud risk assessment result of "confirmed as cross-application privacy theft" ruling instruction, the highest level of countermeasure will be triggered. Specifically, the operating system will forcibly terminate its malicious AI agent main process, physically cut off all its network and IO permissions, and pop up a conspicuous red warning in the center of the screen to notify the user: "Security warning: a high-risk malicious behavior of an assistant stealing the address book has been forcibly blocked."

[0108] (2) The cloud-based risk assessment results confirm that the first operation request is an operation request whose risk can be downgraded by modifying the execution parameter level. That is, the server suggests modifying the execution parameters to intelligently downgrade and allow the request: If an AI agent intends to find a nearby high-quality coffee shop for a user, it requests to call the "precise GPS positioning satellite API". The server analyzes that its business intention is compliant, but believes that finding nearby shops only requires "rough positioning at the business district level". Therefore, the server can send a "suggest downgrade and modify parameters" instruction to the terminal. In this way, when the terminal receives the instruction, it can parse the instruction and give the parsing result to the interception component or interception probe in the terminal. The interception component can automatically adjust the execution parameters of the request message (or execution context information) of the operation request from requesting precise location (ACCESS_FINE_LOCATION) to requesting fuzzy base station location (ACCESS_COARSE_LOCATION), and then return the first operation request after parameter downgrading to the underlying data for release. It can be understood that this permission management strategy, while ensuring that the AI ​​agent can successfully complete its intended work, strictly deprives unnecessary privacy from the underlying layer and implements the principle of least privilege security.

[0109] Thus, since the AI ​​agent's operation requests can be automatically allowed when assessed as low-risk in the cloud, instructed to lower the level of execution parameters before allowing the request when assessed as medium-risk, or forcibly terminated when assessed as high-risk, fine-grained and dynamic control over the AI ​​agent's operation requests is achieved. This ensures the security of the terminal while maximizing the efficiency and continuity of the AI ​​agent's task execution, avoiding service interruptions or user experience degradation caused by a single blocking strategy. At the same time, the hierarchical response mechanism that links cloud-based intelligent assessment with terminal execution effectively reduces the frequency of passive authorization and decision-making burden for users, and reduces unnecessary local resource consumption and privacy exposure risks.

[0110] Thus, when the terminal side assesses the first operation request as a high-risk request, the permission management device can report the AI ​​agent's historical API call logs within the first historical time period, the request feature data of the first operation request, and the multi-dimensional context features corresponding to the terminal to the cloud server, requesting the server to conduct a more comprehensive and accurate risk assessment, and execute the permission control policy corresponding to the cloud risk assessment result of the server on the first operation request. This not only improves the accuracy of permission management for high-risk operation requests, but also reduces the amount of data exchanged between the terminal and the cloud, and avoids the abuse of users' sensitive data.

[0111] In the permission management method provided in this application embodiment, when the AI ​​agent triggers the first operation request, the first operation request of the AI ​​agent can be intercepted, and the risk level corresponding to the first operation request can be dynamically determined according to the multi-dimensional context features of the terminal and the request data features of the first operation request through a preset model. Therefore, compared with the static pop-up permission mechanism where the user decides whether to allow it, the permission management method provided in this application embodiment can improve the dynamism, accuracy and intelligence of permission management.

[0112] In some embodiments of this application, after step 104 above, the permission management method provided in the embodiments of this application may further include the following step 105.

[0113] Step 105: If the initial risk level is medium risk, the terminal will use the user's confirmation result of the first operation request as a feedback sample to update the model parameters of the preset model.

[0114] In some embodiments of this application, for operation requests that the terminal determines to be of medium risk level, the terminal can capture the user's actual click behavior (agree or refuse) on the warning interaction window through the system front-end, and convert the monitored click behavior into a numerical reward and punishment gradient signal of the reinforcement learning algorithm. Then, through the numerical reward and punishment gradient signal, the terminal performs backpropagation weight parameter update on the lightweight risk assessment model deployed locally on the terminal to obtain a continuously evolving personalized strategy model.

[0115] Specifically, the system employs an online reinforcement learning framework, either a Deep Q-Network (DQN) or Proximal Policy Optimization (PPO), running locally on the terminal to protect privacy. The captured "specific combination of contextual feature vectors" is defined as the state space for model exploration, and the system's "interception / allowance decision" is defined as the action space. If the user manually clicks "Always Allow Current Task" in the prompt window, the system determines its decision is too conservative and receives a positive reward signal (e.g., Reward = +1). If the system previously adopted a direct allowance strategy, but the user later determines that the operation contradicts their actual intention and manually withdraws the permission in the system settings, the system will automatically receive a severe, high negative penalty feedback (e.g., Reward = -2). The system uses these continuously accumulating gradient feedbacks daily to fine-tune the bias weights of the nodes within the local lightweight neural network in real time.

[0116] For example, the lightweight risk assessment model on the device becomes increasingly intelligent with use, better understanding the evolving needs of users: A senior executive of a company heavily relies on a smart in-car assistant and habitually has the AI ​​agent automatically read aloud and draft replies to subordinates' work WeChat messages during his daily commute. Initially, due to the highly sensitive nature of automatic cross-application messaging, the device model defaulted to classifying this "voice-activated WeChat messaging" operation as having a medium-to-high risk (score of 0.65), frequently triggering yellow warnings on the car's infotainment screen, forcibly interrupting driving and requiring confirmation. Because it was indeed a convenient feature he needed, the executive repeatedly clicked "Always Allow" in the pop-up windows. The reinforcement learning algorithm accurately accumulated these strong positive confirmation feedbacks (+1 reward for each consecutive acquisition), and began to quickly reduce the local model's estimated risk weight penalty for this specific combination of features in the scenario. A week later, when the exact same context (high-quality Bluetooth connection already paired + vehicle GPS speed > 20km / h + clear voice response indicating the intended action) occurred again, the evolved local risk assessment model quickly converged and dropped the initial risk score to 0.15, directly breaking through the low-risk clearance threshold and smoothly transitioning to the "green channel" for silent execution. At this point, the access control system truly possessed the dynamic security evolution capability of "becoming more understanding of the owner's habits the more it's used, and becoming more intelligent and non-intrusive," thus concluding the entire business logic loop.

[0117] It is understandable that by using a personalized reinforcement learning mechanism driven by user feedback, the risk assessment model on the device can continuously optimize the risk assessment strategy based on the user's "allow / deny / modify" feedback, thereby achieving adaptive evolution of the risk assessment strategy, making the assessment results more in line with the user's personalized habits, and continuously improving the defense against new types of attacks.

[0118] Thus, by using user confirmations of medium-risk operation requests as feedback samples to update the parameters of the preset model deployed on the terminal, the preset model can continuously adapt to individual user behavior and risk preferences, gradually improving the accuracy of identifying medium-risk scenarios. Furthermore, with the accumulation of feedback samples, the preset model can effectively reduce false positives and false negatives caused by fixed thresholds, reducing unnecessary confirmation burdens for users in subsequent similar operations. At the same time, it can achieve self-evolution of the preset model without relying on cloud retraining, thereby achieving personalized and dynamic permission management effects in environments with privacy protection and limited resources.

[0119] This application also provides a permission management method, such as... Figure 2 As shown, the permission management method provided in this application embodiment includes the following steps 201 to 203.

[0120] 201. The server receives a risk assessment request sent by the terminal. The risk assessment request is used to request a risk assessment of the first operation request. The risk assessment request includes: the continuous operation log of the AI ​​agent in the first historical time period, the request feature data of the first operation request, and the multi-dimensional context features corresponding to the terminal.

[0121] For a description of the request feature data and multidimensional context features, please refer to the relevant descriptions in the above terminal-side method embodiments.

[0122] Step 202: The server evaluates the cloud risk assessment result corresponding to the first operation request based on the continuous operation log, request feature data, and multi-dimensional context features.

[0123] The aforementioned multidimensional context features may include at least two of the following: spatiotemporal context features, user context features, device context features, and social context features. For further descriptions of multidimensional context features, please refer to the relevant descriptions in the above terminal method embodiments; to avoid repetition, they will not be repeated here.

[0124] In some embodiments of this application, the first operation request is triggered by an AI agent and is determined by the terminal to be an operation request of a high risk level.

[0125] In some embodiments of this application, when the terminal detects that the first operation request is a high-risk operation request, it can use a high-strength encrypted communication channel to package and send the continuous operation sequence logs, request feature data, and multi-dimensional context features of the AI ​​agent recently intercepted locally by the terminal within the first historical time period to the cloud server. Using the large-scale language model and graph neural network on the server side, based on the continuous operation sequence logs, request feature data, and multi-dimensional context features, the first operation request is subjected to deep intent tracing and cross-application operation chain detection to obtain accurate cloud-based deep evaluation and decision instructions.

[0126] In one embodiment of this application, step 202 can be implemented by the following steps 202A and 202B.

[0127] Step 202A: Based on the continuous operation logs, the server constructs a cross-application operation chain graph using a graph neural network model, with the applications or APIs called by the AI ​​agent in the past as nodes and the data flow direction as directed edges.

[0128] Step 202B: The server determines the cloud risk assessment result corresponding to the first operation request through a large language model based on the cross-application operation chain diagram, request feature data, and multi-dimensional context features.

[0129] In some embodiments of this application, given the limited computing power and field of view of the terminal, the system deploys a large language model (LLM) with hundreds of billions of parameters in the cloud to handle extremely complex long texts and deep camouflage logic. Simultaneously, the cloud security center specifically utilizes graph neural networks (GNNs), such as graph attention networks (GANs), to reconstruct the historical API call records of the AI ​​agent over a period of time (e.g., 10 minutes) into a macroscopic "cross-application operation chain directed graph" (nodes in the graph represent applications or operation APIs, and edges represent hidden data flows). Through intelligent matching of the massive graph topology patterns, it investigates whether there are advanced persistent threats (APTs) or complex Trojan data flows that utilize data obfuscation and transfer through intermediate nodes.

[0130] For example, cross-application data leak detection: The terminal intercepted a cleanup-type AI agent attempting to perform an external network upload action, and reported its recent historical logs to the cloud for verification. The cloud-based GNN network reconstructed its complete operation graph and found that: three minutes ago, the AI ​​agent first silently read all contacts from the Contacts app (node ​​A), then encrypted and packaged them and wrote them to the system's public clipboard (node ​​B). Immediately afterwards, it woke up a hidden backdoor browser app (node ​​C) and attempted to read the clipboard and POST the data to an unknown overseas node. Individually, "reading contacts" or "calling the browser" are easy to evade rule detection, but once a continuous A->B->C graph link is formed, the cloud-based large model instantly and accurately determines that this is an extremely malicious "paraphrasing-style cross-application privacy theft composite attack," and immediately generates a "confirm high risk" final decision command in the cloud and issues it.

[0131] For example, if the large language model confirms that the first operation request is of a high-risk level, the server can send a ruling instruction to the terminal stating "confirmed as cross-application privacy theft," instructing the terminal to trigger the highest level of countermeasures. Specifically, this ruling instruction can instruct the terminal to forcibly terminate its malicious AI agent main process, physically cut off all its network and I / O permissions, and display a prominent red warning in the center of the screen to notify the user: "Security Warning: A high-risk malicious act of an assistant stealing contacts has been forcibly blocked."

[0132] For example, the large language model confirms that the first operation request is one that can be intelligently downgraded and allowed by suggesting modifications to the execution parameters: If an AI agent intends to find a nearby high-quality coffee shop for a user, it requests to call the "precise GPS positioning satellite API." The cloud-based large model analyzes that its business intent is compliant, but believes that finding nearby shops only requires "rough positioning at the business district level." Therefore, the cloud issues a "suggest downgrade and parameter modification" instruction. This instruction can instruct the terminal to automatically replace ACCESS_FINE_LOCATION in the API request message of the operation request with ACCESS_COARSE_LOCATION before returning the data to the underlying layer for approval. This ensures that the AI ​​agent can successfully complete its intended work while strictly depriving unnecessary privacy from the underlying layer, implementing the principle of least privilege security.

[0133] In some embodiments of this application, the server can instruct the terminal to perform different permission management based on different risk assessment results (risk levels) of the large language model, thereby balancing security and the user experience of AI agent automation. Specifically, the server can instruct the terminal to perform a three-color intervention on operation requests according to the risk level assessment results: low-risk operation requests are green, indicating automatic permission; medium-risk operation requests are yellow, indicating permission after modifying execution parameters; and high-risk operation requests are red, indicating forced blocking and termination. In this way, while maintaining a malicious behavior detection rate of over 95%, unnecessary user disturbance can be reduced by 70%.

[0134] Thus, by constructing a cross-application operation chain graph through graph neural networks, the data flow and dependencies when AI agents call multiple applications or APIs can be explicitly captured from continuous operation logs, revealing combined risks that cannot be reflected by a single operation. Furthermore, by combining request features and multi-dimensional context, a large language model can perform semantic-level risk assessment, thereby significantly improving the ability to identify hidden anomalies, intention deviations, and data leakage paths in complex operation chains, and achieving a more comprehensive and accurate cloud risk assessment.

[0135] Step 203: The server sends the cloud-based risk assessment results to the terminal.

[0136] In the permission management method provided in this application embodiment, since the server can receive the risk assessment request sent by the terminal to assess the risk of the first operation request, and can perform a comprehensive risk assessment of the first operation request based on the continuous operation log, request feature data and multi-dimensional context features in the risk assessment request, and send the cloud risk assessment result obtained from the assessment to the terminal, it can utilize the abundant computing resources and global behavioral data in the cloud to achieve comprehensive risk assessment across applications and time domains, effectively identifying potential AI attacks in complex operation chains; at the same time, by entrusting complex high-risk assessment tasks to the cloud, terminal overhead can be reduced, and ultimately secure, efficient and intelligent permission management of AI agents can be achieved.

[0137] This application also provides a permission management method, such as... Figure 3 As shown, the permission management method may include the following steps:

[0138] Step 301: When the AI ​​agent triggers the first operation request, the terminal intercepts the first operation request and obtains the multi-dimensional context features corresponding to the terminal and the request feature data of the first operation request.

[0139] Step 302: Based on the multi-dimensional context features and the request feature data, the terminal determines the initial risk level corresponding to the first operation request using a lightweight risk assessment model.

[0140] In some embodiments of this application, after determining the initial risk level, if the initial risk level is low, then step 303 is executed; if the initial risk level is medium, then steps 304 and 305 are executed; if the initial risk level is high, then steps 306 to 312 or steps 306 to 313 are executed.

[0141] Step 303: The terminal allows the first operation request.

[0142] Step 304: The terminal displays a prompt message, which prompts the user to decide whether to grant the first operation request.

[0143] Step 305: The terminal performs permission management on the first operation request based on the user's input of the prompt information.

[0144] It is understandable that after receiving user input, the terminal can manage permissions for the first operation request according to the user's input instructions. For example, if the user allows it, the terminal allows the first operation request; otherwise, the terminal intercepts and terminates the first operation request.

[0145] Step 306: The terminal freezes the first operation request and sends a risk assessment request to the server.

[0146] Step 307: The server receives the first operation request.

[0147] The risk assessment request is used to request a risk assessment of the first operation request. The risk assessment request includes: the continuous operation log of the AI ​​agent within the first historical time period, the request feature data of the first operation request, and the multi-dimensional context features corresponding to the terminal.

[0148] Step 308: Based on the continuous operation logs, the server constructs a cross-application operation chain graph using a graph neural network model, with the applications or APIs called by the AI ​​agent in the past as nodes and the data flow direction as directed edges.

[0149] Step 309: Based on the cross-application operation chain diagram, request feature data, and multi-dimensional context features, the server performs in-depth analysis and cross-application operation chain detection using a large language model to obtain the cloud risk assessment result corresponding to the first operation request.

[0150] Step 310: The server sends the cloud-based risk assessment results to the terminal.

[0151] Step 311: The terminal receives the cloud-based risk assessment results sent by the server.

[0152] In some embodiments of this application, when the cloud risk assessment result indicates that the first operation request is low risk or a medium risk request that can be modified and approved, the terminal continues to execute step 311 below. When the cloud risk assessment result indicates that the first operation request is a high risk request, the terminal continues to execute step 312 or step 313 below.

[0153] Step 312: The terminal allows the first operation request or allows the first operation request after modifying the execution parameter level of the first operation request.

[0154] Step 313: The terminal forcibly terminates the first operation request.

[0155] In this way, a lightweight verification model is deployed on the device side to handle high-frequency, low-risk requests, while only complex, high-risk operations are handled by the cloud. This enables effective, real-time, and intelligent permission management of the autonomous AI agent on mobile devices with limited device resources.

[0156] It should be noted that the above-described method embodiments, or the various possible implementations of the method embodiments, can be executed individually, or, provided there are no contradictions, they can be combined with each other. The specific implementation can be determined according to actual usage requirements, and this application embodiment does not impose any restrictions on this.

[0157] The permission management method provided in this application can be executed by a permission management device. This application uses the execution of the permission management method by a permission management device as an example to illustrate the permission management device provided in this application.

[0158] This application provides an access control device, such as... Figure 4 As shown, the access control device 400 may include: a processing module 401; the processing module 401 is used for:

[0159] When the AI ​​agent triggers the first operation request, the first operation request is intercepted, and the multi-dimensional context features corresponding to the terminal and the request feature data of the first operation request are obtained.

[0160] Based on the multidimensional context features and the request feature data, the initial risk level corresponding to the first operation request is determined by a preset model;

[0161] Based on the permission control policy corresponding to the initial risk level, the first operation request is subject to permission management.

[0162] The multidimensional context features include at least two of the following: spatiotemporal context features, user context features, device context features, and social context features.

[0163] In some possible embodiments, the access control policy corresponding to the initial risk level mentioned above includes any of the following:

[0164] Allow low-risk operation requests;

[0165] The user decides whether to approve a medium-risk operation request.

[0166] For high-risk operation requests, execute the permission management policy corresponding to the cloud risk assessment results determined by the server.

[0167] In some possible embodiments, the above-described processing module is specifically used to perform feature extraction processing on the execution context information corresponding to the first operation request to obtain the request feature data;

[0168] The request feature data includes at least one of the following: agent identity information, execution parameters, and task intent information.

[0169] In some possible embodiments, the access control device further includes a sending module and a receiving module. Specifically, the aforementioned processing module is used to freeze the first operation request when the initial risk level is high.

[0170] The sending module is used to send a risk assessment request to the server. The risk assessment request is used to request a risk assessment of the first operation request. The risk assessment request includes the historical API call logs of the AI ​​agent within a first historical time period, the request feature data, and the multidimensional context features.

[0171] A receiving module is used to receive cloud-based risk assessment results sent by the server;

[0172] The aforementioned processing module is specifically used to execute the permission control policy corresponding to the cloud risk assessment result on the first operation request.

[0173] In some possible embodiments, the access control policy corresponding to the above cloud risk assessment results includes any of the following:

[0174] Allow low-risk operation requests;

[0175] Forcefully terminate high-risk operation requests and output security warning information;

[0176] Operation request parameters for medium-risk level operations are downgraded before being allowed.

[0177] In some possible embodiments, the above-described processing module is further configured to, after executing the first permission control policy corresponding to the initial risk level on the first operation request, update the model parameters of the preset model by taking the user's confirmation result of the first operation request as a feedback sample when the initial risk level is a medium risk level.

[0178] In the permission management method provided in this application embodiment, when the first operation request of the AI ​​agent is intercepted, the risk level corresponding to the first operation request can be dynamically determined by a preset model based on the multi-dimensional context features of the terminal and the request data features of the first operation request. Therefore, compared with the static pop-up permission mechanism where the user decides whether to allow it, the permission management method provided in this application embodiment can improve the dynamism, accuracy and intelligence of permission management.

[0179] This application provides an access control device, such as... Figure 5 As shown, the permission management device 500 may include: a receiving module 501, a processing module 502, and a sending module 503.

[0180] The receiving module 501 is used to receive a risk assessment request sent by the terminal. The risk assessment request is used to request a risk assessment of the first operation request. The risk assessment request includes: the continuous operation log of the AI ​​agent in the first historical time period, the request feature data of the first operation request, and the multi-dimensional context features corresponding to the terminal.

[0181] The processing module 502 is used to evaluate the cloud risk assessment result corresponding to the first operation request based on the continuous operation log, the request feature data and the multi-dimensional context features in the risk assessment request received by the receiving module 501.

[0182] The sending module 503 is used to send the cloud risk assessment result obtained by the processing module 502 to the terminal.

[0183] The multidimensional context features include at least two of the following: spatiotemporal context features, user context features, device context features, and social context features.

[0184] In some possible embodiments, the processing module described above is specifically used for:

[0185] Based on the continuous operation log, a cross-application operation chain graph is constructed using a graph neural network model, with the applications or APIs called by the AI ​​agent in the past as nodes and the data flow direction as directed edges.

[0186] Based on the cross-application operation chain diagram, request feature data, and multi-dimensional context features, the cloud risk assessment result corresponding to the first operation request is determined through a large language model.

[0187] In the permission management device provided in this application embodiment, since the server can receive the risk assessment request sent by the terminal to assess the risk of the first operation request, and can perform a comprehensive risk assessment of the first operation request based on the continuous operation log, request feature data and multi-dimensional context features in the risk assessment request, and send the cloud risk assessment result obtained from the assessment to the terminal, it can utilize the abundant computing resources and global behavioral data in the cloud to achieve comprehensive risk assessment across applications and time domains, effectively identifying potential AI attacks in complex operation chains; at the same time, by entrusting complex high-risk assessment tasks to the cloud, terminal overhead can be reduced, and ultimately secure, efficient and intelligent permission management of AI agents can be achieved.

[0188] In the embodiments of this application Figure 4 The access control device shown can be an electronic device or a component within an electronic device, such as an integrated circuit or a chip. The electronic device can be a terminal or other devices besides a terminal. For example, the electronic device can be a mobile phone, tablet computer, laptop computer, PDA, in-vehicle electronic device, mobile internet device (MID), augmented reality (AR) / virtual reality (VR) device, robot, wearable device, ultra-mobile personal computer (UMPC), netbook, or personal digital assistant (PDA), etc. The access control device shown in embodiment 5 of this application can be a server, network attached storage (NAS), personal computer (PC), television (TV), ATM, or self-service machine, etc., and this application embodiment does not specifically limit the device.

[0189] The permission management device in this application embodiment can be a device with an operating system. This operating system can be Android, iOS, or other possible operating systems; this application embodiment does not specifically limit the specific operating system used.

[0190] The embodiments provided in this application, Figure 4 The access control device shown can achieve Figures 1 to 3 The various processes implemented by the terminal in the method embodiment shown will not be described again here to avoid repetition.

[0191] The application provides the following: Figure 5 The access control device shown can achieve Figures 1 to 3 The various processes implemented by the server in the method embodiment shown will not be described again here to avoid repetition.

[0192] Optionally, such as Figure 6 As shown, this application embodiment also provides a communication device 600, including a processor 601 and a memory 602. The memory 602 stores a program or instructions that can run on the processor 601. When the program or instructions are executed by the processor 601, they implement the various steps of the above-described terminal-side method embodiment or the various steps of the above-described server-side method embodiment, and can achieve the same technical effect. To avoid repetition, they will not be described again here.

[0193] Figure 7 A schematic diagram of the hardware structure of a terminal to implement an embodiment of this application.

[0194] The terminal 1500 includes, but is not limited to, the following components: radio frequency unit 1501, network module 1502, audio output unit 1503, input unit 1504, sensor 1505, display unit 1506, user input unit 1507, interface unit 1508, memory 1509, and processor 1510.

[0195] Those skilled in the art will understand that the terminal 1500 may also include a power supply (such as a battery) for supplying power to various components. The power supply may be logically connected to the processor 1510 through a power management system, thereby enabling functions such as managing charging, discharging, and power consumption through the power management system. Figure 7 The terminal structure shown does not constitute a limitation on the terminal. The terminal may include more or fewer components than shown, or combine certain components, or have different component arrangements, which will not be elaborated here.

[0196] The processor 1510 is used for:

[0197] When the AI ​​agent triggers the first operation request, the first operation request is intercepted, and the multi-dimensional context features corresponding to the terminal and the request feature data of the first operation request are obtained.

[0198] Based on the multidimensional context features and the request feature data, the initial risk level corresponding to the first operation request is determined by a preset model;

[0199] Based on the access control policy corresponding to the initial risk level, the first operation request is subject to access control.

[0200] In some possible embodiments, the above-mentioned multidimensional context features include at least two of the following: spatiotemporal context features, user context features, device context features, and social context features.

[0201] In some possible embodiments, the access control policy corresponding to the initial risk level mentioned above includes any of the following:

[0202] Allow low-risk operation requests;

[0203] The user decides whether to approve a medium-risk operation request.

[0204] For high-risk operation requests, execute the permission management policy corresponding to the cloud risk assessment results determined by the server.

[0205] In some possible embodiments, the processor 1510 is specifically used to perform feature extraction processing on the execution context information corresponding to the first operation request to obtain the request feature data;

[0206] The request feature data includes at least one of the following: agent identity information, execution parameters, and task intent information.

[0207] In some possible embodiments, the processor 1510 is specifically configured to freeze the first operation request if the initial risk level is high.

[0208] Radio frequency unit 1501 is used to send a risk assessment request to the server. The risk assessment request is used to request a risk assessment of the first operation request. The risk assessment request includes the historical API call logs of the AI ​​agent within a first historical time period, the request feature data and the multidimensional context features.

[0209] Radio frequency unit 1501 is used to receive cloud risk assessment results sent by the server;

[0210] The aforementioned processor 1510 is specifically used to execute the permission control policy corresponding to the cloud risk assessment result on the first operation request.

[0211] In some possible embodiments, the access control policy corresponding to the above cloud risk assessment results includes any of the following:

[0212] Allow low-risk operation requests;

[0213] Forcefully terminate high-risk operation requests and output security warning information;

[0214] Operation request parameters for medium-risk level operations are downgraded before being allowed.

[0215] In some possible embodiments, the processor 1510 is further configured to, after executing the first permission control policy corresponding to the initial risk level on the first operation request, update the model parameters of the preset model by taking the user's confirmation result of the first operation request as a feedback sample when the initial risk level is a medium risk level.

[0216] In the terminal provided in this application embodiment, when the first operation request of the AI ​​agent is intercepted, the risk level corresponding to the first operation request can be dynamically determined by a preset model based on the multi-dimensional context features of the terminal and the request data features of the first operation request. Therefore, compared with the static pop-up permission mechanism where the user decides whether to allow it, the permission management method provided in this application embodiment can improve the dynamism, accuracy and intelligence of permission management.

[0217] It should be understood that, in this embodiment, the input unit 1504 may include a graphics processing unit (GPU) 15041 and a microphone 15042. The GPU 15041 processes image data of still images or videos obtained by an image capture device (such as a camera) in video capture mode or image capture mode. The display unit 1506 may include a display panel 15061, which may be configured in the form of a liquid crystal display, an organic light-emitting diode, or the like. The user input unit 1507 includes at least one of a touch panel 15071 and other input devices 15072. The touch panel 15071 is also called a touch screen. The touch panel 15071 may include a touch detection device and a touch controller. Other input devices 15072 may include, but are not limited to, physical keyboards, function keys (such as volume control buttons, power buttons, etc.), trackballs, mice, and joysticks, which will not be described in detail here.

[0218] The memory 1509 can be used to store software programs and various data. The memory 1509 may primarily include a first storage area for storing programs or instructions and a second storage area for storing data. The first storage area may store the operating system, application programs or instructions required for at least one function (such as sound playback, image playback, etc.). Furthermore, the memory 1509 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct memory bus RAM (DRRAM). The memory 1509 in this embodiment includes, but is not limited to, these and any other suitable types of memory.

[0219] Processor 1510 may include one or more processing units; optionally, processor 1510 integrates an application processor and a modem processor, wherein the application processor mainly handles operations involving the operating system, user interface, and applications, and the modem processor mainly handles wireless communication signals, such as a baseband processor. It is understood that the aforementioned modem processor may also not be integrated into processor 1510.

[0220] This application also provides a readable storage medium storing a program or instructions. When the program or instructions are executed by a processor, they implement the various processes of the above-described permission processing method embodiments and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0221] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0222] This application embodiment also provides a chip, which includes a processor and a communication interface. The communication interface is coupled to the processor. The processor is used to run programs or instructions to implement the various processes of the above-described permission processing method embodiments and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0223] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0224] This application provides a computer program product that is stored in a storage medium and executed by at least one processor to implement the various processes of the above-described permission processing method embodiments, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0225] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0226] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a computer software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0227] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for managing access permissions, characterized in that, The method includes: When the AI ​​agent triggers the first operation request, the first operation request is intercepted, and the multi-dimensional context features corresponding to the terminal and the request feature data of the first operation request are obtained. Based on the multidimensional context features and the request feature data, the initial risk level corresponding to the first operation request is determined by a preset model; Based on the permission control policy corresponding to the initial risk level, the first operation request is subject to permission management. The multidimensional context features include at least two of the following: spatiotemporal context features, user context features, device context features, and social context features.

2. The method according to claim 1, characterized in that, The access control policy corresponding to the initial risk level includes any one of the following: Allow low-risk operation requests; The user decides whether to approve a medium-risk operation request. For high-risk operation requests, execute the permission management policy corresponding to the cloud risk assessment results determined by the server.

3. The method according to claim 1, characterized in that, The step of obtaining the request feature data of the first operation request includes: The execution context information corresponding to the first operation request is subjected to feature extraction processing to obtain the request feature data; The request feature data includes at least one of the following: agent identity information, execution parameters, and task intent information.

4. The method according to claim 2, characterized in that, The execution of the first access control policy corresponding to the initial risk level on the first operation request includes: If the initial risk level is high risk, the first operation request is frozen and a risk assessment request is sent to the server. The risk assessment request is used to request a risk assessment of the first operation request. The risk assessment request includes the AI ​​agent's historical application interface API call logs within the first historical time period, the request feature data, and the multidimensional context features. Receive the cloud-based risk assessment results sent by the server; The permission control policy corresponding to the cloud risk assessment result is executed on the first operation request.

5. The method according to claim 4, characterized in that, The access control policy corresponding to the cloud-based risk assessment results includes any one of the following: Allow low-risk operation requests; Forcefully terminate high-risk operation requests and output security warning information; Operation request parameters for medium-risk level operations are downgraded before being allowed.

6. The method according to claim 2, characterized in that, After executing the first access control policy corresponding to the initial risk level on the first operation request, the method further includes: If the initial risk level is medium risk, the user's confirmation result of the first operation request is used as a feedback sample to update the model parameters of the preset model.

7. A method for managing access permissions, characterized in that, The method includes: A risk assessment request is sent by the receiving terminal. The risk assessment request is used to request a risk assessment of the first operation request. The risk assessment request includes: the continuous operation log of the AI ​​agent within a first historical time period, the request feature data of the first operation request, and the multi-dimensional context features corresponding to the terminal. Based on the continuous operation log, the request feature data, and the multidimensional context features, evaluate the cloud risk assessment result corresponding to the first operation request; Send the cloud-based risk assessment results to the terminal; The multidimensional context features include at least two of the following: spatiotemporal context features, user context features, device context features, and social context features.

8. The method according to claim 7, characterized in that, The step of evaluating the cloud risk assessment result corresponding to the first operation request based on the continuous operation log, the request feature data, and the multi-dimensional context features includes: Based on the continuous operation log, a cross-application operation chain graph is constructed using a graph neural network model, with the application or application interface API historically called by the AI ​​agent as nodes and the data flow direction as directed edges. Based on the cross-application operation chain diagram, request feature data, and multi-dimensional context features, the cloud risk assessment result corresponding to the first operation request is determined through a large language model.

9. A permission management device, characterized in that, The device includes: a processing module; Processing module, used for: When the AI ​​agent triggers the first operation request, the first operation request is intercepted, and the multi-dimensional context features corresponding to the terminal and the request feature data of the first operation request are obtained. Based on the multidimensional context features and the request feature data, the initial risk level corresponding to the first operation request is determined by a preset model; Based on the permission control policy corresponding to the initial risk level, the first operation request is subject to permission management. The multidimensional context features include at least two of the following: spatiotemporal context features, user context features, device context features, and social context features.

10. The apparatus according to claim 9, characterized in that, The access control policy corresponding to the initial risk level includes any one of the following: Allow low-risk operation requests; The user decides whether to approve a medium-risk operation request. For high-risk operation requests, execute the permission management policy corresponding to the cloud risk assessment results determined by the server.

11. The apparatus according to claim 9, characterized in that, The processing module is specifically used to perform feature extraction processing on the execution context information corresponding to the first operation request to obtain the request feature data. The request feature data includes at least one of the following: agent identity information, execution parameters, and task intent information.

12. The apparatus according to claim 11, characterized in that, The device further includes a transmitting module and a receiving module; The processing module is specifically used to freeze the first operation request when the initial risk level is high risk level. The sending module is used to send a risk assessment request to the server. The risk assessment request is used to request a risk assessment of the first operation request. The risk assessment request includes the historical application programming interface (API) call logs of the AI ​​agent within a first historical time period, the request feature data, and the multidimensional context features. The receiving module is used to receive the cloud-based risk assessment results sent by the server; The processing module is specifically used to execute the permission control policy corresponding to the cloud risk assessment result on the first operation request.

13. The apparatus according to claim 12, characterized in that, The access control policy corresponding to the cloud-based risk assessment results includes any one of the following: Allow low-risk operation requests; Forcefully terminate high-risk operation requests and output security warning information; Operation request parameters for medium-risk level operations are downgraded before being allowed.

14. The apparatus according to claim 10, characterized in that, The processing module is further configured to, after executing the first permission control policy corresponding to the initial risk level on the first operation request, and if the initial risk level is medium risk level, use the user's confirmation result of the first operation request as a feedback sample to update the model parameters of the preset model.

15. An access control device, characterized in that, The device includes: a receiving module, a processing module, and a transmitting module; The receiving module is used to receive a risk assessment request sent by the terminal. The risk assessment request is used to request a risk assessment of the first operation request. The risk assessment request includes: the continuous operation log of the AI ​​agent within a first historical time period, the request feature data of the first operation request, and the multi-dimensional context features corresponding to the terminal. The processing module is used to evaluate the cloud risk assessment result corresponding to the first operation request based on the continuous operation log, the request feature data and the multi-dimensional context features in the risk assessment request received by the receiving module. The sending module is used to send the cloud risk assessment result obtained by the processing module to the terminal; The multidimensional context features include at least two of the following: spatiotemporal context features, user context features, device context features, and social context features.

16. The apparatus according to claim 15, characterized in that, The processing module is specifically used for: Based on the continuous operation log, a cross-application operation chain graph is constructed using a graph neural network model, with the application or application interface API historically called by the AI ​​agent as nodes and the data flow direction as directed edges. Based on the cross-application operation chain diagram, request feature data, and multi-dimensional context features, the cloud risk assessment result corresponding to the first operation request is determined through a large language model.

Images