A method and system for adaptive protection of vehicle privacy data

Abstract

The application discloses a kind of self-adaptive protection method and system of vehicle privacy data, it is related to vehicle data processing technical field, the method includes: the classified multiple modal data stream obtained is determined, and the initial privacy level of each data is determined;Extract the context association information of multiple modal data stream, and context association information includes data generation timestamp, data use identification and user current operation intention;According to initial privacy level and context association information, dynamically adjust data privacy weight, generate the data processing strategy matching current scene;According to data processing strategy, the target data in multiple modal data stream is executed differential privacy processing.The application classifies and determines initial privacy level to vehicle-mounted multiple modal data stream, dynamically adjusts privacy weight in combination with scene, user, business related context information, executes differential privacy processing to target data, avoids the protection and demand disjunction problem caused by fixed protection strategy of prior art.

Description

Technical Field

[0001] This invention relates to the field of vehicle data processing technology, and specifically to an adaptive protection method and system for vehicle privacy data. Background Technology

[0002] With the rapid development of automotive intelligence and connectivity technologies, in-vehicle systems have become comprehensive platforms integrating data collection, transmission, and processing, leading to the increasingly widespread generation and application of multimodal data streams. Currently, vehicles can collect various types of data in real time, including images, audio, location trajectories, driving behavior, and vehicle status, through various hardware such as cameras, microphones, positioning modules, and bus systems. This data not only supports the realization of core functions such as in-vehicle navigation, maintenance, and entertainment, but also provides data support for the intelligent upgrade of vehicles. However, at the same time, in-vehicle data contains a large amount of privacy information involving users' personal identities, private itineraries, and driving habits, highlighting the risk of privacy leaks. How to ensure the safe protection of in-vehicle privacy data while guaranteeing the normal operation of intelligent vehicle functions has become a crucial issue that urgently needs to be addressed in the field of in-vehicle electronic technology, and a technological bottleneck that the industry must overcome in its development.

[0003] To protect in-vehicle privacy data, various approaches have emerged in existing technologies, with their core methods largely revolving around data classification and fixed protection strategies. Existing technologies typically begin by classifying in-vehicle data streams into private and non-private data. Then, based on preset privacy levels, a uniform protection method is applied to different types of private data. For example, all private data may be encrypted with the same strength, or specific types of private data may undergo fixed de-identification and interception operations. While some existing technologies introduce privacy level classifications, the protection strategies adopted after classification remain unchanged, failing to adjust to dynamic factors such as the actual operating scenarios of the vehicle and user intentions. Essentially, these remain static protection models.

[0004] While the aforementioned existing technologies can protect in-vehicle privacy data to a certain extent, they have significant technical shortcomings and are difficult to adapt to the complexity and dynamic needs of in-vehicle scenarios. The core problem lies in the use of fixed, static protection strategies. Regardless of whether privacy levels are defined, the corresponding protection methods in existing technologies are fixed and cannot be flexibly adjusted according to various dynamic changes during actual vehicle operation. This results in a serious disconnect between the protection strategy and actual application scenarios. This static protection mode either has excessively high protection strength, consuming too much in-vehicle computing resources and affecting the normal use of data for various core in-vehicle functions, or it has excessively low protection strength, failing to cope with the risk of privacy leakage in different scenarios and thus failing to meet the actual application needs of in-vehicle privacy data protection. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide an adaptive protection method and system for vehicle privacy data.

[0006] A first aspect of the present invention is to provide an adaptive protection method for vehicle privacy data, the method comprising: The acquired multimodal data streams are classified to determine the initial privacy level of each type of data; Extract the contextual association information of the multimodal data stream, which includes the data generation timestamp, data usage identifier, and the user's current operation intent; Based on the initial privacy level and the context association information, the data privacy weight is dynamically adjusted to generate a data processing strategy that matches the current scenario. According to the data processing strategy, differentiated privacy processing is performed on the target data in the multimodal data stream.

[0007] According to one aspect of the above technical solution, the step of classifying the acquired multimodal data streams and determining the initial privacy level of each type of data includes: A lightweight rule engine is used to perform initial filtering on the acquired multimodal data stream to identify and label explicit non-privacy data. The remaining implicit data after primary filtering is then fed into a lightweight deep learning model for semantic parsing and feature extraction to identify implicit privacy data.

[0008] According to one aspect of the above technical solution, the step of dynamically adjusting the data privacy weight based on the initial privacy level and the context association information to generate a data processing strategy matching the current scenario includes: A privacy weight calculation model is constructed, which includes a time decay factor, an intent matching factor, and a usage risk factor. The data generation timestamp of the target data is obtained, the time difference between the current time and the data generation timestamp is calculated, and a first weighting coefficient is calculated based on a preset time decay function, wherein the larger the time difference, the smaller the first weighting coefficient. The user's current operation intent is parsed to generate an intent identifier, and the semantic similarity between the intent identifier and the data usage identifier of the target data is calculated to obtain a second weight coefficient, wherein the higher the semantic similarity, the lower the second weight coefficient; The initial privacy level, the first weighting coefficient, and the second weighting coefficient are weighted and calculated to generate a dynamic comprehensive privacy score. The dynamic privacy comprehensive score is mapped to a preset policy matrix to generate a corresponding data processing policy, which includes data encryption, data anonymization, access interception, or plaintext transmission.

[0009] According to one aspect of the above technical solution, the step of generating a dynamic comprehensive privacy score by weighting the initial privacy level, the first weight coefficient, and the second weight coefficient is expressed as follows: ; in, For dynamic privacy comprehensive scoring, The target data is based on an initial privacy level baseline value determined using a privacy feature library. Time difference The time decay factor function, and satisfies That is, as the time difference increases, the value of this factor decreases monotonically. Based on semantic similarity The intention matching adjustment factor, where, That is, the vector of the user's current operational intent. Data usage identifier vector Cosine similarity between them This refers to environmental noise disturbance.

[0010] According to one aspect of the above technical solution, the step of performing differentiated privacy processing on target data in the multimodal data stream according to the data processing strategy includes: According to the security level identifier defined in the data processing strategy, the target data is routed to the corresponding processing channel, and the processing channel includes at least a high-security isolation channel and a normal data transmission channel; When it is determined that the target data has entered the high-security isolation channel, a feature localization algorithm that matches the target data type is invoked to lock the area to be processed containing sensitive information in the original data carrier; Based on the boundary coordinates of the region to be processed, an adaptive mask layer is generated and overlaid on the original data carrier to block the visual or semantic readability of sensitive information. The original data carrier covered with the adaptive mask layer is subjected to integrity verification and encapsulation to generate the final output data stream.

[0011] According to one aspect of the above technical solution, the step of generating an adaptive mask layer based on the boundary coordinates of the region to be processed, and overlaying the adaptive mask layer on the original data carrier to block the visual or semantic readability of sensitive information includes: Obtain the boundary coordinates of the area to be processed, and determine the coverage area based on the boundary coordinates; A base mask layer is generated based on the coverage area, and the edge areas of the base mask layer are smoothed to obtain an adaptive mask layer.

[0012] According to one aspect of the above technical solution, the step of obtaining the boundary coordinates of the area to be processed and determining the coverage area based on the boundary coordinates includes: The original data carrier is analyzed using a feature localization algorithm to extract geometric feature points of the target privacy object, and a closed contour line is generated based on the geometric feature points. The bounding rectangular region of the closed contour line is determined as the boundary coordinates of the region to be processed, and the minimum rectangular mask range covering the target privacy object is generated based on the boundary coordinates.

[0013] A second aspect of the present invention is to provide an adaptive protection system for vehicle privacy data, applied to the method described in the above-mentioned technical solution, the system comprising: The data classification module is used to classify the acquired multimodal data streams and determine the initial privacy level of each type of data. The information extraction module is used to extract the contextual information of the multimodal data stream, including the data generation timestamp, data usage identifier, and the user's current operation intent. The strategy generation module is used to dynamically adjust the data privacy weight based on the initial privacy level and the context association information, and generate a data processing strategy that matches the current scenario. The policy execution module is used to perform differentiated privacy processing on the target data in the multimodal data stream according to the data processing policy.

[0014] A third aspect of the present invention is to provide a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the steps of the method described in the above-described technical solution.

[0015] A fourth aspect of the present invention is to provide an electronic device, the electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the steps of the method described in the above technical solutions.

[0016] Compared with existing technologies, the adaptive protection method and system for vehicle privacy data shown in this invention have the following advantages: This invention classifies in-vehicle multimodal data streams and determines initial privacy levels. It then dynamically adjusts privacy weights based on contextual information related to the scenario, user, and business, generating a data processing strategy adapted to the current scenario. This allows for differentiated privacy processing of target data, avoiding the disconnect between protection and needs caused by fixed protection strategies in existing technologies. It does not excessively consume in-vehicle computing resources, effectively mitigating various privacy leakage risks and protecting user privacy while ensuring the normal use of data required for core functions such as in-vehicle navigation and maintenance. This improves the practicality and reliability of in-vehicle privacy protection, meeting the actual needs of the in-vehicle electronics field for privacy data protection and demonstrating significant industry application value. Attached Figure Description

[0017] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 A flowchart illustrating the adaptive protection method for vehicle privacy data provided in an embodiment of the present invention; Figure 2 This is a structural block diagram of an adaptive protection system for vehicle privacy data provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Several embodiments of the present invention are shown in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that the disclosure of the present invention will be more thorough and complete.

[0019] It should be noted that when a component is said to be "fixed to" another component, it can be directly on the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0021] Example 1 Please see Figure 1The first embodiment of the present invention provides an adaptive protection method for vehicle privacy data, the method comprising steps S10-S40: Step S10: Classify the acquired multimodal data streams and determine the initial privacy level of each type of data; Step S20: Extract the context association information of the multimodal data stream. The context association information includes the data generation timestamp, data usage identifier, and the user's current operation intent. Step S30: Based on the initial privacy level and the context association information, dynamically adjust the data privacy weight to generate a data processing strategy that matches the current scenario; Step S40: Perform differentiated privacy processing on the target data in the multimodal data stream according to the data processing strategy.

[0022] In practical implementation, the in-vehicle privacy protection system, also known as the privacy protection function, is activated first. The system synchronously acquires multimodal data streams generated during vehicle operation through a preset data source interface. These data streams originate from various in-vehicle hardware modules and software applications, specifically including: image and video data collected by in-vehicle cameras; in-vehicle voice and environmental audio data collected by in-vehicle microphones; vehicle status and driving behavior data transmitted via CAN bus and LIN bus; real-time positioning coordinates and driving trajectory data collected by GPS / BeiDou positioning modules; user operation data generated by the vehicle-machine interface module; and related data generated by in-vehicle entertainment, navigation, maintenance, and other applications. Because these data are diverse in form and have significantly different privacy attributes, they need to be categorized first to clarify the initial privacy baseline.

[0023] Specifically, the system has a built-in preset data classification rule base and privacy feature base. The classification rule base includes data type division standards, such as text, image, audio, location, behavior, and vehicle status. The privacy feature base includes privacy attribute identifiers corresponding to various types of data, such as faces, ID numbers, and mobile phone numbers related to personal identity; home addresses, company addresses, and private itineraries related to private scenarios; and driving habits and in-vehicle conversations related to sensitive behaviors. When classifying data, the system automatically categorizes the acquired multimodal data streams according to the classification rule base, matches each type of data with the privacy feature base, and determines the corresponding initial privacy level for each type of data based on whether the data contains sensitive information, the type of sensitive information, and the risk of leakage.

[0024] Initial privacy levels are typically categorized into three levels: high, medium, and low. For example, in-vehicle facial images, real-time location trajectories, and in-vehicle voice conversations are classified as high-sensitivity; driving behavior data and vehicle fault codes are classified as medium-sensitivity; and external ambient temperature and vehicle speed on public roads are classified as low-sensitivity.

[0025] After classifying the multimodal data streams and determining the initial privacy level, the contextual information of the data is extracted. The purpose is to obtain contextual and real-time information that can affect data privacy risks, so that subsequent privacy protection strategies can be adapted to the current actual scenario.

[0026] The extracted contextual information includes three parts: data generation timestamp, which is the precise system time when each piece of data was collected and generated. This timestamp is used to determine the timeliness of the data; for example, recent location data poses a higher privacy risk than location data from six months ago. Data usage identifier, which is a usage tag carried when the data is called and transmitted. It is pre-set by the application calling the data to clarify the purpose of the data use. For example, the usage identifier for navigation applications calling location data is "navigation and positioning," and the usage identifier for maintenance applications calling vehicle fault data is "fault diagnosis." The user's current operation intent is determined by analyzing the user's recent operation behavior, such as continuously clicking on navigation applications, issuing voice commands to "query home address and route," activating the in-car camera for a call, vehicle operating status, and application service call records. This allows for a comprehensive judgment of the user's current true operation intent. For example, if a user turns on navigation and enters their home address, their operation intent can be determined as "navigation home." In this case, the use of location data is consistent with the user's intention, and the privacy risk is reduced.

[0027] Then, the privacy weight is dynamically adjusted. Specifically, the initial privacy level is used as the base value, and combined with the extracted contextual information, the privacy weight of each type of data is dynamically adjusted in real time through a preset weight calculation model.

[0028] The privacy weight is used to quantify the actual privacy risk of data in the current scenario. The higher the weight, the higher the privacy risk, and the stronger the protection required. During the adjustment process, factors such as time decay (i.e., the older the data, the lower the privacy risk and the lower the weight), intent matching (the closer the user's action intent matches the data's purpose, the lower the privacy risk and the lower the weight), and scenario risk (e.g., data from private areas has a higher privacy risk than data from public areas, and therefore a higher weight) are comprehensively considered to ensure that the adjusted privacy weight truly reflects the actual data risk at the current moment and in the current scenario. For example, a highly sensitive piece of location data, if generated 10 minutes ago, identified as "navigation location," and the user's intent is "home navigation," will have its privacy weight appropriately reduced; if the data was generated 10 minutes ago, identified as "ad push," and the user's intent is "turn off ads," its privacy weight will be significantly increased.

[0029] Based on the adjusted privacy weights, a data processing strategy perfectly suited to the current scenario is generated by matching the strategy from a pre-defined strategy library. It should be noted that the strategy library pre-stores processing methods, security levels, transmission channels, storage rules, and other information corresponding to different privacy weight ranges. For example, the processing strategy for high privacy weight data is "high-strength end-to-end encryption + local isolated storage + strict access permission verification"; the processing strategy for medium privacy weight data is "partial de-identification + lightweight encryption + time-limited access"; and the processing strategy for low privacy weight data is "anonymization + plaintext transmission + regular storage".

[0030] Finally, according to the generated data processing strategy, differentiated privacy processing is performed on the target data in the multimodal data stream, i.e., the sensitive data that requires privacy protection. Furthermore, the core of differentiated processing is "protection on demand," avoiding over-protection that renders the data unusable, while also preventing insufficient protection that leads to privacy leaks. For example, highly sensitive in-vehicle facial images are masked and encrypted end-to-end to ensure that facial information cannot be recognized and is not cracked during transmission; moderately sensitive driving behavior data is partially anonymized to hide specific driving trajectory details while retaining statistical data for maintenance analysis; and low-sensitivity vehicle ambient temperature data is anonymized and transmitted directly in plaintext to ensure the normal use of applications such as navigation and weather.

[0031] It should also be noted that the entire above process is completed at the vehicle edge node, ensuring the real-time nature of data processing, avoiding the impact of cloud transmission delays on the normal use of the vehicle scenario, and reducing the transmission of raw sensitive data to the outside world, thereby reducing the risk of leakage.

[0032] Compared with existing technologies, the adaptive protection method for vehicle privacy data shown in this embodiment has the following advantages: This embodiment classifies in-vehicle multimodal data streams and determines initial privacy levels. It then dynamically adjusts privacy weights based on contextual information related to the scenario, user, and business, generating a data processing strategy adapted to the current scenario. This allows for differentiated privacy processing of target data, avoiding the disconnect between protection and needs caused by fixed protection strategies in existing technologies. It does not excessively consume in-vehicle computing resources, effectively mitigating various privacy leakage risks and protecting user privacy while ensuring the normal use of data required for core functions such as in-vehicle navigation and maintenance. This enhances the practicality and reliability of in-vehicle privacy protection, meeting the actual needs of the in-vehicle electronics field for privacy data protection and demonstrating significant industry application value.

[0033] Example 2 The second embodiment of the present invention also provides an adaptive protection method for vehicle privacy data, which is basically similar to the method shown in the first embodiment, except that: In this embodiment, the step of classifying the acquired multimodal data streams and determining the initial privacy level of each type of data includes: A lightweight rule engine is used to perform initial filtering on the acquired multimodal data stream to identify and label explicit non-privacy data. The remaining implicit data after primary filtering is then fed into a lightweight deep learning model for semantic parsing and feature extraction to identify implicit privacy data.

[0034] It should be noted that classifying multimodal data streams involves two progressive stages: primary filtering and deep identification. By proceeding from a coarse to a precise approach, both identification efficiency and accuracy are ensured, avoiding the omission of private data or misclassification of non-privacy data.

[0035] In the initial filtering stage, the lightweight rule engine can complete the preliminary screening of multimodal data streams within milliseconds, without the need for complex semantic analysis or feature extraction. It achieves rapid filtering solely through pre-defined rule matching. Furthermore, the lightweight rule engine has a built-in explicit non-privacy data judgment rule library. This library is pre-configured by technical personnel based on the attributes of vehicle data and includes information such as the characteristics, fields, formats, and data types of various explicit non-privacy data. Specifically, this includes: common vehicle operating parameters (such as vehicle speed, RPM, engine temperature, and fuel level), environmental data (such as outside temperature, humidity, and light intensity), publicly available service data (such as public road names and navigation routes), and device status data without personal association (such as in-vehicle screen brightness and audio volume).

[0036] During the initial filtering process, the lightweight rule engine reads each piece of data in the multimodal data stream one by one, matches the data's features and fields with the judgment rules in the rule base, and if the data fully meets the judgment criteria for explicit non-privacy data, it is immediately identified and marked as "non-privacy data" and classified into a low privacy level or no privacy level, directly excluding it from the subsequent deep recognition process.

[0037] For example, if the outside temperature collected by the vehicle's onboard sensor is 25°C, this data does not involve any personal privacy information and matches the rule in the rule base that "ambient temperature data is non-private data." Therefore, it is marked as non-private data and does not require further processing. This initial filtering can quickly identify most of the non-private data, significantly reducing the amount of data required for subsequent deep recognition, lowering the computational load on the vehicle, and improving the efficiency of the entire initial privacy level determination process.

[0038] Then, the deep recognition stage begins. After the initial filtering, the remaining data consists of implicit data that cannot be directly determined by simple rules. The characteristics of this type of data are that privacy information is deeply hidden, and it does not have obvious non-privacy features. It may contain semantically ambiguous and feature-indistinct privacy information, which cannot be identified by rule matching. For example, in-car images containing blurry faces, voice commands containing home addresses, location trajectory fragments containing private trips, and ID number fragments hidden in ordinary text.

[0039] For this type of implicit data, it is input into a pre-trained and optimized lightweight deep learning model. Different recognition logics and algorithms are used for different types of implicit data: For text-based implicit data such as in-vehicle messages and speech-to-text, the model uses natural language processing technology for semantic parsing and sensitive field extraction to identify sensitive information such as ID card numbers, mobile phone numbers, and home addresses. For image and video-based implicit data such as footage captured by in-vehicle cameras, the model uses object detection algorithms for feature extraction and object recognition to accurately detect private targets such as faces, license plates, and private items in the footage. For audio-based implicit data such as in-vehicle voice conversations, the model uses speech recognition and semantic understanding technology to extract sensitive content such as family information, personal preferences, and private conversations. For location and behavior-based implicit data such as driving trajectories, the model uses trajectory analysis algorithms to identify sensitive trajectory points such as home addresses, company addresses, and private locations.

[0040] After the model identification is completed, detailed identification results are output, marking the type and sensitivity of privacy information contained in each piece of implicit data, and then these implicit privacy data are classified into the corresponding initial privacy level according to their sensitivity.

[0041] For example, images containing clear human faces are classified as high sensitivity, images containing blurry human faces are classified as medium sensitivity, and audio containing ordinary speech but no sensitive content is classified as low sensitivity.

[0042] Ultimately, by combining the aforementioned primary filtering with deep recognition, comprehensive and accurate identification of all privacy data in the multimodal data stream was achieved, providing an accurate and reliable basis for determining the initial privacy level, while also taking into account recognition efficiency and the resource constraints of the in-vehicle environment.

[0043] In this embodiment, the step of dynamically adjusting the data privacy weight based on the initial privacy level and the context association information to generate a data processing strategy matching the current scenario includes: A privacy weight calculation model is constructed, which includes a time decay factor, an intent matching factor, and a usage risk factor. The data generation timestamp of the target data is obtained, the time difference between the current time and the data generation timestamp is calculated, and a first weighting coefficient is calculated based on a preset time decay function, wherein the larger the time difference, the smaller the first weighting coefficient. The user's current operation intent is parsed to generate an intent identifier, and the semantic similarity between the intent identifier and the data usage identifier of the target data is calculated to obtain a second weight coefficient, wherein the higher the semantic similarity, the lower the second weight coefficient; The initial privacy level, the first weighting coefficient, and the second weighting coefficient are weighted and calculated to generate a dynamic comprehensive privacy score. The dynamic privacy comprehensive score is mapped to a preset policy matrix to generate a corresponding data processing policy, which includes data encryption, data anonymization, access interception, or plaintext transmission.

[0044] Specifically, based on the actual needs of in-vehicle privacy protection and combined with the influencing factors of data privacy risks, a privacy weight calculation model is constructed. This model includes three core influencing factors, which measure the actual privacy risks of data from three dimensions: data timeliness, user intent rationality, and data usage purpose security, to ensure the comprehensiveness and accuracy of the weight calculation.

[0045] The core influencing factors mentioned above include: a time decay factor, which characterizes the impact of data timeliness on privacy risks, i.e., the longer the data was generated, the lower the privacy sensitivity and the smaller the privacy risk; an intent matching factor, which characterizes the impact of the degree of matching between the user's current operational intent and the data's purpose on privacy risks, i.e., the higher the matching degree, the more the data usage conforms to the user's intention, and the smaller the privacy risk; and a purpose risk factor, which characterizes the impact of the security risks of the data's purpose itself on privacy risks, i.e., the safer and more compliant the purpose, the smaller the privacy risk.

[0046] Then, the first weighting coefficient, which is also the coefficient corresponding to the time decay factor, is calculated. First, the data generation timestamp of the target data is obtained, which is the precise system time when the data was collected and generated. This time is stored in the attribute field of the data. The difference between the current system time and the data generation timestamp is calculated to obtain the length of time the data has been generated, which is also the time difference. For example, if the current time is 14:30 on May 10, 2026, and the data generation timestamp is 14:20 on May 10, 2026, then the time difference is 10 minutes. Then, the calculated time difference is input into the preset time decay function to calculate the first weighting coefficient.

[0047] The time decay function employs a monotonically decreasing function, such as an exponential decay function, pre-configured by technical personnel based on the time-sensitive characteristics of the privacy data, adhering to the rule that "the larger the time difference, the smaller the first weight coefficient." For example, when the time difference is 10 minutes, the first weight coefficient is 0.9; when the time difference is 1 hour, the first weight coefficient is 0.7; when the time difference is 24 hours, the first weight coefficient is 0.3; and when the time difference exceeds 7 days, the first weight coefficient is 0.1. This achieves a gradual reduction in the privacy weight of historical data, preventing expired data from continuously occupying high-security resources while ensuring that recently sensitive data receives strong protection.

[0048] Then, the second weighting coefficient, which corresponds to the intent matching factor, is calculated. First, the user's current operational intent is analyzed. Specifically, this is done by collecting recent user actions (such as touch operation records, voice commands, and application call records), vehicle operating status (such as driving location and driving mode), and application service request information. After comprehensive analysis, standardized intent identifiers are generated. For example, if a user continuously clicks on a navigation application, enters their home address, and issues the voice command "Navigate home," the intent identifier is "Home navigation." If a user opens a vehicle maintenance application and clicks "Fault detection," the intent identifier is "Vehicle maintenance." Subsequently, the data usage identifier inherent in the target data is extracted. This identifier is pre-set by the application calling the data and is used to clarify the purpose of data use. For example, the usage identifier for a navigation application calling location data is "Navigation positioning," for an advertising application calling location data is "Ad push," and for a maintenance application calling vehicle fault data is "Fault diagnosis."

[0049] Furthermore, the semantic similarity between the intent identifier and the data purpose identifier is calculated using a cosine similarity algorithm. This involves converting both the intent identifier and the data purpose identifier into vector form and calculating the cosine of the angle between the two vectors. A cosine value closer to 1 indicates higher semantic similarity, while a value closer to 0 indicates lower semantic similarity. For example, if the intent identifier is "home navigation" and the data purpose identifier is "navigation and positioning," their semantic similarity is 0.9, indicating a high degree of overlap between the user's intent and the data purpose. Conversely, if the intent identifier is "home navigation" and the data purpose identifier is "ad push," their semantic similarity is 0.1, indicating a complete mismatch between the user's intent and the data purpose. Finally, a second weight coefficient is generated based on the semantic similarity, following the rule that "the higher the semantic similarity, the smaller the second weight coefficient." For example, when the semantic similarity is 0.9, the second weight coefficient is 0.1; when the semantic similarity is 0.5, the second weight coefficient is 0.5; and when the semantic similarity is 0.1, the second weight coefficient is 0.9. This allows for dynamic adjustment of the privacy weight based on the user's intent matching, ensuring that privacy risks are reduced when data usage aligns with user intent and increased when it does not.

[0050] Then, a weighted calculation is performed to generate a dynamic privacy comprehensive score. This is done by obtaining the initial privacy level benchmark value of the target data, and then performing a comprehensive calculation on the initial privacy level benchmark value, the first weight coefficient, and the second weight coefficient according to a preset weighting formula to generate a dynamic privacy comprehensive score.

[0051] Then, the mapping strategy matrix generates data processing strategies. By pre-establishing a strategy matrix, which uses a dynamic privacy comprehensive score as the core index, the score is divided into different risk level ranges. Each risk level range corresponds to a preset data processing strategy. The strategy content includes data processing methods (encryption, de-identification, access interception, plaintext transmission, etc.), transmission channels (high-security isolation channels, ordinary transmission channels), access permissions (administrator permissions, ordinary permissions, temporary permissions), storage methods (local storage, cloud encrypted storage, temporary storage), etc.

[0052] The above process maps the calculated dynamic privacy comprehensive score to the policy matrix, finds the corresponding risk level range, and automatically calls the preset data processing policy corresponding to that range as the basis for subsequent privacy processing, ensuring that the data processing policy is accurately matched with the privacy risk of the current data.

[0053] Specifically, the step of generating a dynamic comprehensive privacy score by weighting the initial privacy level, the first weight coefficient, and the second weight coefficient is expressed as follows: ; in, For dynamic privacy comprehensive scoring, The target data is based on an initial privacy level baseline value determined using a privacy feature library. Time difference The time decay factor function, and satisfies That is, as the time difference increases, the value of this factor decreases monotonically. Based on semantic similarity The intention matching adjustment factor, where, That is, the vector of the user's current operational intent. Data usage identifier vector Cosine similarity between them This refers to environmental noise disturbance.

[0054] Example 3 The third embodiment of the present invention also provides an adaptive protection method for vehicle privacy data. The method shown in this embodiment is basically similar to the method shown in the first embodiment, except that: In this embodiment, the step of performing differentiated privacy processing on the target data in the multimodal data stream according to the data processing strategy includes: According to the security level identifier defined in the data processing strategy, the target data is routed to the corresponding processing channel, and the processing channel includes at least a high-security isolation channel and a normal data transmission channel; When it is determined that the target data has entered the high-security isolation channel, a feature localization algorithm that matches the target data type is invoked to lock the area to be processed containing sensitive information in the original data carrier; Based on the boundary coordinates of the region to be processed, an adaptive mask layer is generated and overlaid on the original data carrier to block the visual or semantic readability of sensitive information. The original data carrier covered with the adaptive mask layer is subjected to integrity verification and encapsulation to generate the final output data stream.

[0055] Specifically, the generated data processing strategy is first read, and the security level identifiers defined in the strategy are extracted. These security level identifiers are determined based on a dynamic privacy comprehensive score and correspond one-to-one with the risk levels in the strategy matrix. For example, low risk corresponds to security level identifier "L", medium risk corresponds to security level identifier "M", and high risk corresponds to security level identifier "H". These security level identifiers indicate the required protection strength and transmission path for the target data. Based on the security level identifiers, the target data is routed to the corresponding processing channel. The processing channels are pre-configured data processing paths with different security levels, including at least two types: high-security isolation channels and ordinary data transmission channels. These two types of channels are independent of each other to avoid cross-data leakage.

[0056] The high-security isolation channel employs independent hardware resources, encryption protocols, and storage space, specifically designed for processing high-privacy-risk data. Data transmission within this channel utilizes end-to-end high-strength encryption, and storage is local and isolated, not sharing resources with other channels. Strict access control checks are also implemented to prevent unauthorized access, tampering, or leakage of data. In contrast, the ordinary data transmission channel uses conventional transmission methods and storage resources to process low- and medium-risk data. Data transmission within this channel can employ lightweight encryption or plaintext transmission, and storage uses conventional methods, balancing processing efficiency and basic security while improving data transmission and processing speed. For example, highly sensitive in-vehicle facial images and real-time location trajectories are routed to the high-security isolation channel, while moderately sensitive driving behavior data and low-sensitivity vehicle environment data are routed to the ordinary data transmission channel.

[0057] Then, the feature localization of sensitive areas is performed. When it is determined that the target data has entered the high-security isolation channel, it means that the data is high privacy risk data and requires high-intensity and precise privacy protection. At this time, the feature localization algorithm that matches the data type of the target will be called to accurately locate the area to be processed containing sensitive information in the original data carrier, avoiding indiscriminate processing of the entire data segment. While reducing the amount of computation, the non-sensitive parts of the original data are preserved to the maximum extent, thereby improving the usability of the data.

[0058] The selection of feature localization algorithms is determined based on the type of target data, with different localization logics employed for different data types: For image and video data, object detection algorithms, such as the SSD algorithm, are used to accurately locate sensitive targets such as faces, license plates, and private items by extracting pixel and contour features from the image, and outputting the boundary coordinates of the area to be processed; for audio data, speech segmentation and sensitive feature extraction algorithms are used to locate speech segments containing sensitive content and determine the time interval and audio range to be processed; for text data, sensitive field localization algorithms are used to locate text regions containing sensitive fields such as ID card numbers, mobile phone numbers, and home addresses, and determine the range of characters to be processed; for location and behavior data, trajectory point localization algorithms are used to locate regions containing sensitive trajectory points such as home addresses, company addresses, and private locations, and determine the trajectory segments to be processed.

[0059] For example, for a face image inside a car, the feature localization algorithm accurately detects the outline of the face and outputs the coordinates of the upper left and lower right corners of the face region, which is then used as the region to be processed. For audio data containing sensitive speech, the feature localization algorithm locates the start and end times of the sensitive speech segment, which is then used as the region to be processed.

[0060] Then, an adaptive mask layer is generated and overlaid. After locking the area to be processed, an adaptive mask layer that precisely matches the area to be processed is generated based on the boundary coordinates of the area. The purpose is to block the visual or semantic readability of sensitive information through mask overlay, thereby achieving privacy protection, while ensuring that the mask is naturally integrated with the original data carrier without destroying the overall structure and usability of the data.

[0061] Specifically, first, based on the boundary coordinates of the area to be processed, determine the size, shape, and position of the mask layer to ensure that the mask layer can completely cover the area to be processed without omitting any sensitive information, while not exceeding the range of the area to be processed to avoid obscuring non-sensitive parts.

[0062] After generating the base mask layer, it is precisely overlaid on the area to be processed in the original data carrier to mask sensitive information. During the overlay process, it is necessary to ensure the alignment accuracy between the mask layer and the original data carrier to avoid misalignment or displacement that could lead to the leakage of sensitive information. At the same time, edge smoothing is used to make the mask edges blend naturally with the original data, improving the visual rationality of the data and preventing data anomalies or cracking due to abrupt mask edges.

[0063] Next, data integrity verification is performed on the original data carrier covered with an adaptive mask layer. The purpose is to ensure that the privacy processing does not damage the structure, format and non-sensitive parts of the original data, and to ensure that the processed data can be transmitted, stored and used normally, avoiding data failure due to improper processing.

[0064] Specifically, data integrity verification includes: data format verification, checking whether the processed data format is consistent with the original data and whether it meets the requirements of vehicle communication protocols and applications; data length verification, checking whether the processed data length is consistent with the original data; data structure verification, checking whether the processed data is missing, tampered with, or damaged to ensure data integrity; and data availability verification, checking whether the non-sensitive parts of the processed data can be read and used normally.

[0065] For example, the process verifies whether non-sensitive areas of image data are clearly visible, whether non-sensitive segments of audio data can be played normally, and whether non-sensitive content of text data can be read normally. If the verification passes, the process proceeds to the next step of the encapsulation process; if the verification fails, the mask overlay step will be automatically re-executed until the verification passes.

[0066] Finally, the data is encapsulated and output. In accordance with the requirements of the vehicle communication protocol and data processing strategy, the processed raw data carrier is encapsulated to generate the final output data stream.

[0067] During the encapsulation process, necessary attribute tags are added to the data, including data processing identifiers, security level identifiers, processing timestamps, and data purpose identifiers, to facilitate identification and verification during subsequent data transmission, storage, and access. Simultaneously, the encapsulated data undergoes final encryption to ensure data security during transmission and storage.

[0068] After encapsulation, the output data stream is sent to the corresponding transmission channel for transmission or storage according to the data processing strategy requirements. For example, the data stream from the high-security isolation channel is transmitted to a local isolated storage module or a designated high-security cloud node, while the data stream from the ordinary data transmission channel is transmitted to a regular storage module or the corresponding application, thus completing the final differentiated privacy processing flow.

[0069] In this embodiment, the step of generating an adaptive mask layer based on the boundary coordinates of the region to be processed, and then overlaying the adaptive mask layer on the original data carrier to block the visual or semantic readability of sensitive information includes: Obtain the boundary coordinates of the area to be processed, and determine the coverage area based on the boundary coordinates; A base mask layer is generated based on the coverage area, and the edge areas of the base mask layer are smoothed to obtain an adaptive mask layer.

[0070] Specifically, the boundary coordinates are first obtained and the coverage area is determined. From the output of the feature localization algorithm, the precise boundary coordinates of the area to be processed are obtained. The specific form of the boundary coordinates is determined according to the data type: for image and video data, the boundary coordinates are the geometric coordinates of the area to be processed; for text data, the boundary coordinates are the start and end character positions of the sensitive field; for audio data, the boundary coordinates are the start and end times of the sensitive speech segment; for localization and behavior data, the boundary coordinates are the start and end coordinate points of the sensitive trajectory segment.

[0071] After obtaining the boundary coordinates, the coverage area of ​​the mask layer is determined based on the boundary coordinates. The determination of the coverage area follows the principle of "minimum effective coverage", that is, ensuring that the mask layer can completely cover all sensitive information in the area to be processed, while minimizing the coverage area as much as possible to avoid obscuring non-sensitive parts of the original data, and to retain the effective information and usability of the data to the maximum extent.

[0072] For example, for a face region in an image, the boundary coordinates are the rectangular coordinates (x1, y1) to (x2, y2), and the coverage area determined by the system is this rectangular region, ensuring that the face region is completely covered without obscuring the non-sensitive background around the face; for the sensitive field "182XXXX0126" in text, the boundary coordinates are the Xth character to the Yth character, and the coverage area is this character interval, ensuring that the sensitive phone number is completely covered without obscuring other non-sensitive text content; for sensitive audio segments, such as a start time of 00:05 and an end time of 00:15, the coverage area is this time interval, ensuring that the sensitive audio is completely covered without affecting the playback of other non-sensitive audio segments.

[0073] Furthermore, during the process of determining the coverage area, the boundary coordinates will be verified and corrected to avoid the boundary coordinates shifting due to minor errors in the feature localization algorithm, which would lead to inaccurate coverage.

[0074] Then, a base mask layer is generated and a smooth transition is applied. Based on the defined coverage area, a base mask layer is generated first. The size, shape, and coverage area of ​​the base mask layer are completely consistent, and the fill method is determined according to the data type and privacy protection requirements. For example, the base mask layer for image and video data can use solid color blocks, blur fill, or texture fill; while the base mask layer for text data can use the same fill color as the text background color, or use character replacement; and the base mask layer for audio data can use random noise fill, or use a method to attenuate the volume until it is inaudible.

[0075] The above process involves smoothing the edge areas of the base mask layer to ensure a natural connection between the mask edges and the original data carrier, eliminating obvious boundary marks.

[0076] After a smooth transition process, the final adaptive mask layer is obtained. This layer can completely cover the sensitive information in the area to be processed, blocking the visual or semantic readability of the sensitive information, while also blending naturally with the original data carrier without affecting the overall integrity and usability of the data. Finally, the adaptive mask layer is precisely applied to the corresponding positions on the original data carrier according to the boundary coordinates, completing the masking of the sensitive information.

[0077] In this embodiment, the steps of obtaining the boundary coordinates of the area to be processed and determining the coverage area based on the boundary coordinates include: The original data carrier is analyzed using a feature localization algorithm to extract geometric feature points of the target privacy object, and a closed contour line is generated based on the geometric feature points. The bounding rectangular region of the closed contour line is determined as the boundary coordinates of the region to be processed, and the minimum rectangular mask range covering the target privacy object is generated based on the boundary coordinates.

[0078] Specifically, extracting the geometric feature points of the target privacy object involves first performing a deep analysis of the original data carrier using a feature localization algorithm that matches the target data type, thereby accurately extracting the geometric feature points of the target privacy object from the original data. These geometric feature points are key points that characterize the outline and location of the privacy object.

[0079] The extraction methods for geometric feature points differ depending on the type of raw data carrier: For image and video data, feature point detection algorithms such as SIFT are used to extract key contour points of privacy objects. For example, key contour points of the eyes, nose, mouth, and cheeks are extracted from face images; key contour points of the four corners and edges are extracted from license plate images; and key contour points of items are extracted from images of private items. For text data, character feature extraction algorithms are used to extract the position coordinates of the start, end, and middle key characters of sensitive fields as geometric feature points. For audio data, audio feature extraction algorithms are used to extract the start, end, and middle time points of sensitive speech segments as geometric feature points. For location and behavior data, trajectory feature extraction algorithms are used to extract the start, end, and middle key trajectory points of sensitive trajectory segments as geometric feature points.

[0080] During the extraction of geometric feature points, the extracted feature points will also be denoised and filtered to remove invalid feature points caused by data interference and algorithm errors, and retain valid feature points that can accurately represent the contours of privacy objects, so as to ensure the accuracy and integrity of feature points.

[0081] Then, a closed contour line is generated by fitting. Through curve fitting algorithm, discrete geometric feature points are fitted and calculated to generate a closed contour line that can completely wrap the entire target privacy object. This ensures that all parts of the target privacy object are contained inside the closed contour line without missing any sensitive areas. At the same time, the contour line fits the actual contour of the privacy object as closely as possible, avoiding the contour line being too large or too small.

[0082] After the closed contour line is generated, it will be verified to check whether the contour line completely wraps the target privacy object and whether there are any problems such as breaks or offsets. If there are any problems, the geometric feature points will be automatically re-extracted and fitted until a complete and accurate closed contour line is generated.

[0083] Then, determine the boundary coordinates and coverage of the area to be processed, perform geometric calculations on the generated closed contour line, and find the smallest bounding rectangle that can completely enclose the closed contour line. The smallest bounding rectangle is a rectangular area that can completely enclose the target privacy object with the smallest area and the smallest range. Its characteristics are that the four sides of the rectangle are tangent to the leftmost, rightmost, topmost, and bottommost points of the closed contour line, respectively, ensuring that all parts of the target privacy object are contained inside the rectangle. At the same time, the rectangular area is as small as possible to avoid occluding non-sensitive parts of the original data.

[0084] The calculation also includes determining the boundary coordinates of the minimum bounding rectangle. These boundary coordinates are typically based on the coordinate system of the original data carrier. For example, in an image data coordinate system, the origin is at the top left corner (0,0), the positive x-axis points to the right, and the positive y-axis points downwards. The boundary coordinates of the minimum bounding rectangle are the top left corner (x_min, y_min) and the bottom right corner (x_max, y_max), where x_min is the minimum x-coordinate of the closed contour, y_min is the minimum y-coordinate, x_max is the maximum x-coordinate, and y_max is the maximum y-coordinate. These boundary coordinates define the boundary coordinates of the region to be processed, thus clarifying the location and size of the region.

[0085] Finally, based on the determined boundary coordinates, a minimum rectangular mask range covering the target privacy object is generated. This range is completely consistent with the minimum bounding rectangle, ensuring that the mask layer can completely cover the target privacy object with the minimum range. While achieving privacy protection, the effective information of the original data is preserved to the maximum extent, improving the usability and compliance of the processed data.

[0086] Example 4 A fourth embodiment of the present invention provides an adaptive protection system for vehicle privacy data, applied to the method described in any of the above embodiments, the system comprising: Data classification module 10 is used to classify the acquired multimodal data streams and determine the initial privacy level of each type of data; The information extraction module 20 is used to extract the contextual association information of the multimodal data stream, including the data generation timestamp, data usage identifier, and the user's current operation intent. The strategy generation module 30 is used to dynamically adjust the data privacy weight based on the initial privacy level and the context association information, and generate a data processing strategy that matches the current scenario. The strategy execution module 40 is used to perform differentiated privacy processing on the target data in the multimodal data stream according to the data processing strategy.

[0087] Compared with existing technologies, the adaptive protection system for vehicle privacy data shown in this embodiment has the following advantages: This embodiment classifies in-vehicle multimodal data streams and determines initial privacy levels. It then dynamically adjusts privacy weights based on contextual information related to the scenario, user, and business, generating a data processing strategy adapted to the current scenario. This allows for differentiated privacy processing of target data, avoiding the disconnect between protection and needs caused by fixed protection strategies in existing technologies. It does not excessively consume in-vehicle computing resources, effectively mitigating various privacy leakage risks and protecting user privacy while ensuring the normal use of data required for core functions such as in-vehicle navigation and maintenance. This enhances the practicality and reliability of in-vehicle privacy protection, meeting the actual needs of the in-vehicle electronics field for privacy data protection and demonstrating significant industry application value.

[0088] Example 5 A fifth embodiment of the present invention provides a computer-readable storage medium storing computer instructions, which, when executed by a processor, implement the steps of the method described in any of the above embodiments.

[0089] Example 6 A sixth embodiment of the present invention provides an electronic device, the electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the steps of the method described in any of the above embodiments.

[0090] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0091] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.

Claims

1. An adaptive protection method for vehicle privacy data, characterized in that, The method includes: The acquired multimodal data streams are classified to determine the initial privacy level of each type of data; Extract the contextual association information of the multimodal data stream, which includes the data generation timestamp, data usage identifier, and the user's current operation intent; Based on the initial privacy level and the context association information, the data privacy weight is dynamically adjusted to generate a data processing strategy that matches the current scenario. According to the data processing strategy, differentiated privacy processing is performed on the target data in the multimodal data stream.

2. The adaptive protection method for vehicle privacy data according to claim 1, characterized in that, The steps for classifying the acquired multimodal data streams and determining the initial privacy level for each data type include: A lightweight rule engine is used to perform initial filtering on the acquired multimodal data stream to identify and label explicit non-privacy data. The remaining implicit data after primary filtering is then fed into a lightweight deep learning model for semantic parsing and feature extraction to identify implicit privacy data.

3. The adaptive protection method for vehicle privacy data according to claim 1, characterized in that, The steps of dynamically adjusting data privacy weights and generating a data processing strategy that matches the current scenario based on the initial privacy level and the context association information include: A privacy weight calculation model is constructed, which includes a time decay factor, an intent matching factor, and a usage risk factor. The data generation timestamp of the target data is obtained, the time difference between the current time and the data generation timestamp is calculated, and a first weighting coefficient is calculated based on a preset time decay function, wherein the larger the time difference, the smaller the first weighting coefficient. The user's current operation intent is parsed to generate an intent identifier, and the semantic similarity between the intent identifier and the data usage identifier of the target data is calculated to obtain a second weight coefficient, wherein the higher the semantic similarity, the lower the second weight coefficient; The initial privacy level, the first weighting coefficient, and the second weighting coefficient are weighted and calculated to generate a dynamic comprehensive privacy score. The dynamic privacy comprehensive score is mapped to a preset policy matrix to generate a corresponding data processing policy, which includes data encryption, data anonymization, access interception, or plaintext transmission.

4. The adaptive protection method for vehicle privacy data according to claim 3, characterized in that, The step of generating a dynamic comprehensive privacy score by weighting the initial privacy level, the first weighting coefficient, and the second weighting coefficient is expressed as follows: ； in, For dynamic privacy comprehensive scoring, The target data is based on an initial privacy level baseline value determined using a privacy feature library. Time difference The time decay factor function, and satisfies That is, as the time difference increases, the value of this factor decreases monotonically. Based on semantic similarity The intention matching adjustment factor, where, That is, the vector of the user's current operational intent. Data usage identifier vector Cosine similarity between them This refers to environmental noise disturbance.

5. The adaptive protection method for vehicle privacy data according to claim 1, characterized in that, The steps of performing differentiated privacy processing on target data in the multimodal data stream according to the data processing strategy include: According to the security level identifier defined in the data processing strategy, the target data is routed to the corresponding processing channel, and the processing channel includes at least a high-security isolation channel and a normal data transmission channel; When it is determined that the target data has entered the high-security isolation channel, a feature localization algorithm that matches the target data type is invoked to lock the area to be processed containing sensitive information in the original data carrier; Based on the boundary coordinates of the region to be processed, an adaptive mask layer is generated and overlaid on the original data carrier to block the visual or semantic readability of sensitive information. The original data carrier covered with the adaptive mask layer is subjected to integrity verification and encapsulation to generate the final output data stream.

6. The adaptive protection method for vehicle privacy data according to claim 5, characterized in that, The steps of generating an adaptive mask layer based on the boundary coordinates of the region to be processed, and overlaying the adaptive mask layer on top of the original data carrier to block the visual or semantic readability of sensitive information, include: Obtain the boundary coordinates of the area to be processed, and determine the coverage area based on the boundary coordinates; A base mask layer is generated based on the coverage area, and the edge areas of the base mask layer are smoothed to obtain an adaptive mask layer.

7. The adaptive protection method for vehicle privacy data according to claim 6, characterized in that, The steps of obtaining the boundary coordinates of the region to be processed and determining the coverage area based on the boundary coordinates include: The original data carrier is analyzed using a feature localization algorithm to extract geometric feature points of the target privacy object, and a closed contour line is generated based on the geometric feature points. The bounding rectangular region of the closed contour line is determined as the boundary coordinates of the region to be processed, and the minimum rectangular mask range covering the target privacy object is generated based on the boundary coordinates.

8. An adaptive protection system for vehicle privacy data, characterized in that, The system, applicable to the method of any one of claims 1-7, comprises: The data classification module is used to classify the acquired multimodal data streams and determine the initial privacy level of each type of data. The information extraction module is used to extract the contextual information of the multimodal data stream, including the data generation timestamp, data usage identifier, and the user's current operation intent. The strategy generation module is used to dynamically adjust the data privacy weight based on the initial privacy level and the context association information, and generate a data processing strategy that matches the current scenario. The policy execution module is used to perform differentiated privacy processing on the target data in the multimodal data stream according to the data processing policy.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1-7.

10. An electronic device, characterized in that, The electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the steps of the method according to any one of claims 1-7.

Images