Vehicle safety dynamic management method, cloud platform and storage medium

Abstract

The application relates to the technical field of automobile data security, and discloses a vehicle safety dynamic management method, a cloud platform and a storage medium. The method is applied to the cloud platform and comprises the following steps: obtaining context information of a vehicle; performing risk analysis on the vehicle context information to obtain risk assessment information; under the condition that a preset multi-target condition is met, determining a target operation combination for target data from multiple candidate safety operations based on the risk assessment information and the context information; and performing safety processing on the target data based on the target operation combination. The embodiment of the application can adaptively adjust a safety strategy, improve safety management efficiency and reduce resource consumption.

Description

Technical Field

[0001] This application relates to the field of automotive data security technology, and in particular to a method for dynamic management of vehicle safety, a cloud platform, and a storage medium. Background Technology

[0002] Current data security governance practices for intelligent connected vehicles face significant shortcomings, the core issue being the fundamental conflict between the static nature of security strategies and the dynamic nature of environmental risks. Existing technological systems primarily rely on pre-defined data classification labels and fixed-condition triggering mechanisms, with security decisions entirely based on offline-configured rule bases. This results in strategies failing to adaptively adjust to changes in the external environment or internal state during vehicle operation. For example, in low-risk scenarios, the system continuously executes redundant, high-intensity protection operations, such as excessive encryption or frequent desensitization, leading to idle onboard computing unit resources and reduced battery life. In high-risk scenarios, however, because policy update mechanisms lag behind the speed of threat evolution, the system struggles to promptly identify new attack patterns or environmental mutations, creating blind spots in security protection and significantly increasing the risk of data leakage. This rigidity stems from the lack of real-time correlation analysis capabilities for multi-source heterogeneous information such as vehicle operation data, environmental perception information, user interaction behavior, network traffic characteristics, and system logs. This inability to translate dynamic risk situations into precise operational instructions ultimately leads to inefficient security governance and unbalanced resource allocation. Summary of the Invention

[0003] The purpose of this application is to provide a vehicle safety dynamic management method, cloud platform, and storage medium that can adaptively adjust safety strategies, improve safety management efficiency, and reduce resource consumption.

[0004] This application provides a method for dynamic management of vehicle safety, applied to a cloud platform, including: Obtain the vehicle's context information; the context information is used to characterize the vehicle's current safety status features and is related to the target data to be processed; Risk assessment information is obtained by performing risk analysis on the vehicle context information; Under the condition of satisfying preset multi-objective conditions, based on the risk assessment information and the context information, a combination of target operations for the target data is determined from multiple candidate security operations; Based on the target operation combination, the target data is processed securely.

[0005] In some embodiments, the context information is generated by the vehicle-side agent through feature extraction of multi-source sampling data obtained from the vehicle. The multi-source sampling data includes vehicle operation data, environmental perception data, user interaction data, network traffic data, and / or system log data. The feature extraction includes using a lightweight feature extraction model to extract data content features, vehicle environmental state features, and / or external threat intelligence features from the multi-source sampling data.

[0006] In some embodiments, the risk analysis of the vehicle context information includes: A correlation analysis is performed on the multi-source heterogeneous features in the context information to obtain the correlation analysis results; the multi-source heterogeneous features include data content features, vehicle environment status features, and external threat intelligence features. Based on the correlation analysis results, safety assessment information and risk assessment information are generated; The risk assessment information is generated based on the safety assessment information and the risk assessment information.

[0007] In some embodiments, the correlation analysis of the multi-source heterogeneous features in the context information includes: The data content features, vehicle environment status features, and external threat intelligence features are subjected to modality-specific feature encoding to generate their respective corresponding feature vector representations; The feature vector representations are input into a preset cross-modal fusion network to model the interaction relationship between the feature vector representations based on an attention mechanism, thereby obtaining a cross-modal joint feature representation. The correlation analysis results are generated based on the cross-modal joint feature representation.

[0008] In some embodiments, determining a combination of target operations for the target data from multiple candidate security operations based on the risk assessment information and the context information, under the condition of satisfying preset multi-objective conditions, includes: Based on the risk assessment information, the context information, and the vehicle's computing resource information, the current state of the vehicle is determined; With the goal of maximizing the cumulative reward function, a target action sequence is determined from a preset candidate security operation action space based on the current state; the cumulative reward function is used to perform a weighted calculation on at least two of the security gain index, efficiency loss index, compliance index, and data utility index. Based on the target action sequence, the target operation combination is generated.

[0009] In some embodiments, the vehicle safety dynamic management method further includes: Obtain historical execution feedback data for the target operation combination; In a simulation environment, based on the historical execution feedback data, the parameters of the cumulative reward function and / or the target action sequence are deduced and verified to obtain the deduction and verification results; Based on the simulation and verification results, update the parameters of the cumulative reward function and / or determine the strategy for the target action sequence.

[0010] In some embodiments, the candidate security operation includes at least one of the following steps: Configure an encryption method for the target data based on the security situation characteristics; Reversible desensitization processing is performed on sensitive areas in the view of the target data; Embed traceability information into the target data; Based on the user information of the accessing user, the data tag information of the target data, and / or the environmental risk information of the current environment, a dynamic access authorization decision is made for the target data.

[0011] In some embodiments, the vehicle safety dynamic management method further includes: The interactive management interface is used to display security situation information, security policy status information, operation log information, data breach event tracing information, and / or platform health status information. In response to interactive operations on the interactive management interface, corresponding management operations are performed; the management operations include reviewing, publishing or rolling back the security situation information and the security policy status information, auditing and analyzing the operation log information, tracing and analyzing the source of the data leakage event, and monitoring and alerting the cloud platform for its health status.

[0012] This application embodiment also provides a cloud platform, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described vehicle safety dynamic management method.

[0013] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described vehicle safety dynamic management method.

[0014] The beneficial effects of this application are as follows: By performing risk analysis on vehicle context information, and under the condition of satisfying multiple objectives, a combination of target operations for the target data is determined from multiple candidate security operations based on risk assessment information and context information, thereby enabling secure processing of the target data. Thus, by performing risk analysis on the vehicle context information, dynamic and refined risk assessment is achieved, avoiding overprotection and resource waste in low-risk scenarios. Under the condition of satisfying preset multiple objectives, a combination of target operations for the target data is determined from multiple candidate security operations based on the risk assessment information and context information. This makes the generation of security policies no longer a fixed condition-action rule, but a dynamic decision based on real-time risk, vehicle resources, and multiple optimization objectives, effectively balancing the occupation of vehicle computing resources and user experience, improving security governance efficiency, and reducing resource consumption. Attached Figure Description

[0015] Figure 1 This is a diagram illustrating the application environment of the vehicle safety dynamic management method provided in the embodiments of this application.

[0016] Figure 2 This is a flowchart of the vehicle safety dynamic management method provided in the embodiments of this application.

[0017] Figure 3 This is a flowchart of the specific method of step S102 provided in the embodiments of this application.

[0018] Figure 4 This is a flowchart of the specific method for step S104 provided in the embodiments of this application.

[0019] Figure 5 This is a schematic diagram of the hardware structure of the cloud platform provided in the embodiments of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0021] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and drawings are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0022] 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 application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application. Furthermore, the information, data, and signals involved in the embodiments of this application are all authorized by relevant parties or have been fully authorized by all parties, and the collection, use, and processing of related data comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0023] Existing intelligent connected vehicle data security governance technologies employ predefined data classification and grading labels and fixed condition-action security policies. Decision-making logic is based on a pre-built rule base, and these policies typically remain unchanged during operation. However, the generation and adjustment of these policies lag behind dynamic changes in risk, leading to overprotection in low-risk scenarios and wasted computing resources and energy. In high-risk scenarios, insufficient protection and delayed response create security blind spots. Specifically, the essence of this problem lies in the lack of dynamic perception and adaptation capabilities of security policies to the real-time security situation of vehicles. This results in a systematic disconnect between policy execution and actual risk levels, affecting the synergistic optimization of data processing efficiency and security capabilities. For example, when a vehicle is traveling at high speed, and environmental perception data detects malicious network scanning while system logs record abnormal login attempts, existing security policies, based on fixed rules, cannot dynamically upgrade the risk assessment, failing to promptly encrypt or control sensitive data. Furthermore, the vehicle may continue to process data at a conventional security level, exposing critical data to potential threats, while computing resources remain used for unnecessary low-level protection operations. As a preferred implementation, the multi-source sampling data in this scenario includes vehicle operation data, environmental perception data, and network traffic data. However, the pre-built rule base cannot correlate and analyze the changing trends of these heterogeneous features, thus failing to trigger targeted safety operations.

[0024] If these issues are not addressed, the continued disconnect between security strategies and the actual risk situation will expose vehicle data to a higher risk of unauthorized access, while the overall system efficiency will decrease due to inefficient resource allocation. Consequently, long-term operation may lead to weakened security incident response capabilities, accumulated data breach risks, and ultimately, negatively impact system reliability and compliance.

[0025] Based on this, embodiments of this application provide a vehicle safety dynamic management method, a cloud platform, and a storage medium. By performing risk analysis based on the vehicle's context information and dynamically determining the combination of safety operations, the safety strategy can be adaptively adjusted, thereby improving the efficiency of safety management and reducing resource consumption.

[0026] Figure 1This diagram illustrates the application environment of the vehicle safety dynamic management method provided in this embodiment. (See attached diagram.) Figure 1 This method is applied to a vehicle safety dynamic management system. The system includes a vehicle 101, a vehicle-side agent 102, and a cloud platform 103. The vehicle 101 and vehicle-side agent 102 are connected via a network, and the vehicle-side agent 102 and cloud platform 103 are also connected via a network. The vehicle-side agent 102 can be implemented using a standalone edge server or an edge server cluster composed of several edge servers. The vehicle-side agent 102 sends context information of the vehicle 101 to the cloud platform 103. This context information characterizes the current security posture of the vehicle 101 and is related to the target data to be processed. The cloud platform 103 obtains the context information of the vehicle 101, performs risk analysis on the vehicle context information to obtain risk assessment information, and, under the condition of satisfying preset multi-objective conditions, determines a target operation combination for the target data from multiple candidate security operations based on the risk assessment information and context information. Based on the target operation combination, it performs security processing on the target data.

[0027] See Figure 2 In one embodiment, a dynamic vehicle safety management method is provided, wherein the execution subject of the method is a cloud platform, including but not limited to steps S201 to S204.

[0028] Step S201: Obtain the vehicle's context information.

[0029] Contextual information is used to characterize the vehicle's current safety posture and is relevant to the target data to be processed. It can be understood that contextual information encompasses various pieces of information that collectively describe the vehicle's current state and environment, and is crucial for understanding the vehicle's safety posture and its relevance to specific data requiring safety attention.

[0030] Security posture characteristics are specific attributes or indicators extracted from contextual information, reflecting the vehicle's security status, potential vulnerabilities, and threats at a specific moment.

[0031] Target data refers to any data that is generated by the vehicle, transmitted to the vehicle, or processed inside the vehicle and requires security protection or processing, including sensor data, user data, operation logs, or communication data packets.

[0032] There are several ways to obtain vehicle context information. One approach is to directly output context information from a single sensor or system within the vehicle. For example, a vehicle's dashcam can record the vehicle's speed, location, and timestamps, which are then used directly as context information. Another approach is to obtain context information through pre-configured rules or manual input. For instance, a fixed context information template can be preset at the factory, or maintenance personnel can manually input the vehicle's current operating mode (such as "high-speed driving" or "parking and charging").

[0033] Step S202: Perform risk analysis on the vehicle context information to obtain risk assessment information.

[0034] Risk analysis refers to assessing collected contextual information and security posture characteristics to identify potential security risks, their likelihood of occurrence, and their potential impact.

[0035] Risk assessment information is the output of risk analysis, providing a quantitative or qualitative understanding of the safety risks associated with vehicles and their target data.

[0036] Risk analysis of the vehicle's context information can be performed using a pre-defined set of rules. For example, if the context information indicates that the vehicle's speed exceeds a certain threshold and it is located in a specific area, it can be directly classified as "high-risk." Another approach is to perform simple statistical analysis based on historical data. For instance, the frequency of a certain type of context information within a specific time period can be counted and matched with a pre-defined risk level.

[0037] Step S203: Under the condition of satisfying the preset multi-objective conditions, based on risk assessment information and context information, determine the target operation combination for the target data from multiple candidate security operations.

[0038] Multi-objective conditions refer to predefined standards or objectives that must be met to ensure safe operation. Multi-objective conditions typically involve balancing multiple factors such as safety effectiveness, operational efficiency, compliance requirements, and data utility.

[0039] Candidate security operations refer to a set of security actions or countermeasures that can be applied to target data or vehicles, such as encryption, anonymization, access control, or data embedding.

[0040] A target operation combination refers to one or more selected candidate security operations determined based on the current context and multi-target conditions to address identified risks in target data.

[0041] Determining the target combination of operations for the target data from multiple candidate security operations can be based on a fixed security policy. For example, if the risk assessment information indicates "medium risk," then a pre-defined encryption operation is executed regardless of the context information. Alternatively, the selection of candidate security operations can employ a greedy algorithm, choosing the single operation that appears optimal at the moment, until a simple condition is met. For instance, operations that minimize risk are prioritized, regardless of their impact on efficiency.

[0042] Step S204: Based on the target operation combination, perform secure processing on the target data.

[0043] Security processing refers to the actual execution of a predetermined combination of target operations on the target data, applying selected security measures to protect or manage the data.

[0044] Security processing of target data can be achieved by directly applying a defined combination of target operations to the target data. For example, if the target operation combination is "encryption," then a preset encryption algorithm can be used to encrypt the target data. Another approach is to use batch processing to collect multiple sets of target data and then uniformly apply the target operation combination. For example, all collected log data can be anonymized periodically.

[0045] The following example will provide a more detailed explanation of the above technical solution: Consider a scenario where a smart connected car, vehicle A, is driving on a city road. This vehicle is continuously collecting a large amount of sensor data (e.g., radar data, camera data), navigation data, and user driving behavior data, which are considered as target data to be processed.

[0046] First, the cloud platform acquires the vehicle's contextual information. This contextual information might include: the vehicle is currently on an urban road, traveling at 40 km / h, the weather is clear, the vehicle's system load is moderate, and an autonomous driving assistance function is being activated. This information is used to characterize the vehicle's current safety posture. For example, an urban road environment might imply high traffic density and potential collision risks, while the activation of autonomous driving assistance functions might involve a reliance on high-precision map data.

[0047] Next, the cloud platform performs risk analysis on the vehicle's contextual information to obtain risk assessment information. For example, based on the vehicle's current speed, location, environmental conditions, and the activation status of autonomous driving assistance functions, it may assess a "moderate risk of data breach" and a "low risk of system tampering." This risk assessment information reflects the likelihood and potential impact of different types of security threats in the current context.

[0048] Subsequently, under the premise of meeting preset multi-objective conditions, such as ensuring data security while minimizing the consumption of vehicle computing resources and the impact on user experience, the cloud platform, based on the risk assessment information and the context information, determines a combination of target operations for the target data from multiple candidate security operations. For example, for the aforementioned "moderate data leakage risk," a combination of target operations might be determined from candidate operations such as "data encryption," "sensitive area desensitization," and "data watermark embedding," combined with the vehicle's current computing resource information and tolerance for efficiency loss. This could be, for example, "lightweight encryption of navigation data" and "reversible desensitization of facial regions in camera data." This combination aims to balance security requirements with system performance.

[0049] Finally, the cloud platform performs security processing on the target data based on this combination of operations. Specifically, the vehicle's navigation data is subject to lightweight encryption to prevent unauthorized access; simultaneously, facial information contained in the image data captured by the camera is reversibly anonymized to protect user privacy, while retaining necessary information for subsequent analysis. In this way, the vehicle's data receives timely and appropriate protection in a dynamically changing operating environment, avoiding the problems of over-protection or under-protection that may result from rigid security strategies.

[0050] Based on the examples above, this method demonstrates a significant technological contribution in addressing the problems of rigid strategies and inability to dynamically respond to changes in the environment and risks in existing technologies. Existing technologies typically employ predefined, fixed rules; for example, applying the same encryption strength to all data regardless of the vehicle's operating state. In contrast, this method, by acquiring the vehicle's contextual information, can perceive the vehicle's operating environment, security posture characteristics, and the relevance of the target data to be processed in real time. For instance, when driving on urban roads, it can identify risks associated with high traffic density, while existing solutions may fail to distinguish such subtle environmental changes.

[0051] Furthermore, this method performs risk analysis on the vehicle's contextual information to obtain risk assessment information, achieving dynamic and refined risk evaluation. In the example above, it can distinguish between "moderate data leakage risk" and "low system tampering risk" and make decisions based on these assessment results. This contrasts with the static risk judgment based on coarse-grained classification labels in existing technologies, avoiding overprotection and resource waste in low-risk scenarios.

[0052] Crucially, this method, under the premise of satisfying preset multi-objective conditions, determines the combination of target operations for the target data from multiple candidate security operations based on the risk assessment information and the context information. This makes the generation of security policies no longer a fixed condition-action rule, but a dynamic decision based on real-time risk, vehicle resources, and multiple optimization objectives (e.g., security gain, efficiency loss, compliance). For example, in the example, it can intelligently select a combination of "lightweight encryption" and "reversible desensitization" instead of uniformly adopting the highest strength encryption, thereby effectively balancing the occupation of vehicle computing resources and user experience while ensuring security. Existing solutions often struggle to achieve this multi-objective balance, causing policy adjustments to lag behind changes in risk.

[0053] Ultimately, based on this combination of target operations, the target data is processed securely, ensuring the precise implementation of security measures. This dynamic and adaptive governance mechanism enables vehicle data security strategies to respond in real time to dynamic changes in the environment and risks, effectively compensating for the shortcomings of existing technologies where strategy generation and adjustment lag behind dynamic changes in risks, and significantly improving the flexibility, efficiency, and effectiveness of intelligent connected vehicle data security governance.

[0054] In some embodiments, context information is generated by the vehicle-side agent through feature extraction of multi-source sampled data obtained from the vehicle. Multi-source sampled data includes vehicle operation data, environmental perception data, user interaction data, network traffic data, and / or system log data. Feature extraction includes using a lightweight feature extraction model to extract data content features, vehicle environmental state features, and / or external threat intelligence features from the multi-source sampled data.

[0055] Multi-source sampling data refers to various types of data collected from the vehicle's internal and external environments. This data can comprehensively reflect the vehicle's operating status, environmental conditions, user behavior, and system health.

[0056] Environmental perception data can include images, point clouds, distance information, etc., acquired from sensors such as cameras, radar, and lidar.

[0057] User interaction data includes the driver's voice commands, touch operations, driving habits, etc.

[0058] Network traffic data can include communication data from the vehicle's internal bus (such as CAN, in-vehicle Ethernet) and communication data between the vehicle and external networks (such as V2X, 5G).

[0059] System log data includes operating system logs, application error logs, and security event logs.

[0060] Feature extraction refers to identifying, selecting, and transforming representative and discriminative information from raw, multi-source sampled data. Its purpose is to transform the raw data into a simpler, more easily analyzed form, thereby reducing data dimensionality and highlighting key information. For example, statistical analysis methods can be used to extract statistical features such as the mean, variance, and peak value of the data, or signal processing techniques can be used to extract time-domain and frequency-domain features.

[0061] Lightweight feature extraction models refer to models that consume fewer computational resources and operate more efficiently, enabling them to extract features effectively even with limited computing power and storage space. Examples include deep learning models with fewer parameters (such as MobileNet and SqueezeNet), models built on traditional machine learning algorithms like decision trees and support vector machines, and expert systems based on rule matching and pattern recognition.

[0062] Data content features refer to the characteristics extracted from the semantics or structure of the data itself. These features can reveal the specific meaning, sensitivity, or potential risks of the data. For example, they can identify whether a data packet contains sensitive data such as personally identifiable information or vehicle location information, or analyze the type and content structure of the data transmission protocol.

[0063] Vehicle environmental state characteristics refer to the features that describe the physical and operational environment in which the vehicle is currently located. These characteristics are crucial for assessing the external risks faced by the vehicle. For example, they may include the vehicle's geographical location, current weather conditions, road type (such as highways or urban roads), traffic density, information on surrounding obstacles, and the vehicle's current driving mode (such as autonomous driving or manual driving).

[0064] External threat intelligence features refer to information related to known or potential external security threats. These features help identify and defend against attacks originating from outside. For example, they may include identification information such as known malicious IP addresses, malware signatures, exploit patterns, and types of network attacks (e.g., DDoS attacks, phishing attacks), as well as threat warnings from security intelligence platforms.

[0065] This application deploys a vehicle-side agent and utilizes a lightweight feature extraction model to process multi-source sampled data locally. This enables real-time and comprehensive acquisition of multi-dimensional information, including vehicle operation, environment, user interaction, network traffic, and system logs, from which key data content features, vehicle environmental status features, and external threat intelligence features are extracted. Specifically, the vehicle-side agent acts as a local data aggregation and preliminary processing center within the vehicle, responsible for real-time collection of multi-source sampled data generated by the vehicle's internal and external environments. This multi-source sampled data covers multiple dimensions, including vehicle operation data, environmental perception data, user interaction data, network traffic data, and system log data, ensuring comprehensive perception of the vehicle's status and environment. To accommodate the limited computing resources on the vehicle side, the vehicle-side agent uses a lightweight feature extraction model to efficiently process this raw multi-source sampled data. This model can quickly extract key data content features, vehicle environmental status features, and external threat intelligence features from massive amounts of data. By performing feature extraction locally on the vehicle, the amount of raw data that needs to be transmitted to the cloud platform is significantly reduced, alleviating network bandwidth pressure, and improving the real-time generation of contextual information. These extracted and refined features collectively constitute the vehicle's contextual information, which is then transmitted to the cloud platform. Based on this high-quality, multi-dimensional contextual information, the cloud platform performs risk analysis, enabling a more accurate assessment of the vehicle's current safety posture and providing a solid foundation for determining the appropriate combination of safety actions targeting the data. This collaborative mechanism between the vehicle and the cloud allows the cloud platform to gain more insightful safety situational awareness, thereby improving the accuracy and response efficiency of the entire dynamic vehicle safety management approach.

[0066] The following example illustrates this. Assume a connected intelligent vehicle is in motion. Its vehicle-side agent continuously collects various types of data. Specifically, it collects vehicle operation data, such as vehicle speed and engine RPM obtained via the CAN bus; environmental perception data, such as road images obtained through onboard cameras and distances to obstacles ahead obtained through radar; user interaction data, such as commands issued by the driver through a voice assistant; network traffic data, such as data packets communicating between the vehicle and V2X infrastructure; and system log data, such as abnormal logs generated by the onboard operating system. The vehicle-side agent integrates a lightweight feature extraction model, such as a hybrid model based on rules and statistical analysis. This model processes this data in real time: extracting data content features from network traffic data, such as identifying whether data packets contain the vehicle's precise location information; extracting vehicle environmental status features from environmental perception data, such as determining whether the vehicle is currently in adverse weather conditions like rain or fog, or whether it is passing through a traffic congestion area; and extracting external threat intelligence features from system logs and network traffic, such as detecting abnormal port scanning behavior or communication attempts with known malicious IP addresses. The extracted data content features, vehicle environmental status features, and external threat intelligence features are then integrated to form the current vehicle's contextual information and sent to the cloud platform.

[0067] See Figure 3 In one embodiment, the method for risk analysis of vehicle context information includes, but is not limited to, steps S301 to S303.

[0068] Step S301: Perform association analysis on the multi-source heterogeneous features in the context information to obtain the association analysis results.

[0069] It is understandable that multi-source heterogeneous features are core elements constituting contextual information, characterizing the vehicle's security posture from different dimensions. Multi-source heterogeneous features include data content features, vehicle environmental state features, and external threat intelligence features.

[0070] This association analysis can employ statistical methods, such as correlation analysis and regression analysis, to quantify the statistical association strength between different features; or utilize machine learning models, such as decision trees, association rule mining, and graph neural networks, to discover complex nonlinear relationships and potential patterns between features.

[0071] The results of correlation analysis are structured or unstructured outputs formed after analyzing multi-source heterogeneous features. They reflect the interactions and potential risk patterns among these features. These results can be presented as risk event chains, attack graphs, risk scoring matrices, or feature weight lists, or as joint feature vectors after dimensionality reduction or fusion, providing a more comprehensive characterization of the vehicle's current safety posture.

[0072] Step S302: Based on the correlation analysis results, generate safety assessment information and risk assessment information.

[0073] Safety evaluation information refers to the assessment of a vehicle's current safety protection capabilities, safety configuration, compliance, etc., reflecting the vehicle's own "immunity".

[0074] Risk assessment information refers to the evaluation of potential threats, vulnerabilities, attack probabilities, and the potential losses they may cause, reflecting the "disease" risks faced by vehicles.

[0075] Step S303: Generate risk assessment information based on safety assessment information and risk assessment information.

[0076] This application's solution employs correlation analysis of multi-source heterogeneous features within vehicle context information. First, it integrates and performs pattern recognition on dispersed, multi-type data content features, vehicle environmental state features, and external threat intelligence features, thereby obtaining correlation analysis results that reflect the interactions between these features. This correlation analysis provides a comprehensive data foundation for subsequent risk assessment. Based on these results, the cloud platform can further generate security evaluation information to assess the vehicle's own security capabilities and compliance, while simultaneously generating risk evaluation information to identify potential threats and vulnerabilities. Finally, by comprehensively considering both security and risk evaluation information, a comprehensive and accurate risk assessment can be generated. This hierarchical and progressive analysis method effectively solves the challenge of directly processing complex, multi-source heterogeneous data for risk assessment, ensuring the accuracy and comprehensiveness of the risk assessment and providing a solid basis for subsequent security decisions.

[0077] The following is a concrete example illustrating this. Graph Neural Networks (GNNs) are used to perform correlation analysis on multi-source heterogeneous features. For instance, data content features (such as data sensitivity), vehicle environmental status features (such as vehicle location and network connection type), and external threat intelligence features (such as known vulnerability IDs) are used as nodes or node attributes in a graph. The GNN learns the complex relationships between nodes, thus obtaining a correlation analysis result containing potential risk propagation paths and correlation patterns. Based on the correlation analysis result output by the GNN, a pre-trained classification model (such as Support Vector Machine (SVM) or Random Forest) can be used to generate security assessment information (e.g., the vehicle's current security level is "medium") and risk assessment information (e.g., there is a "high" risk of data leakage). The security assessment information can be generated based on the vehicle's security configuration, patch update status, etc., while the risk assessment information can be generated based on detected abnormal behavior patterns, external threat intelligence matching degree, etc. Finally, a weighted fusion module integrates the security assessment information and the risk assessment information. For example, if the safety assessment information shows that the vehicle has good safety features, but the risk assessment information shows that there is a high risk of external attack, the final risk assessment information may be judged as "medium to high risk", with a specific risk description and recommendations.

[0078] In some embodiments, correlation analysis is performed on multi-source heterogeneous features in context information, including: performing modality-specific feature encoding on data content features, vehicle environment state features, and external threat intelligence features to generate corresponding feature vector representations; inputting the feature vector representations into a preset cross-modal fusion network to model the interaction relationship between the feature vector representations based on an attention mechanism to obtain a cross-modal joint feature representation; and generating correlation analysis results based on the cross-modal joint feature representation.

[0079] Modality-specific feature encoding refers to the process of extracting features from different types of data (i.e., different modalities, such as text, images, and numerical data) using specialized algorithms or models. Its purpose is to transform raw, heterogeneous data into a unified, computable numerical representation, i.e., feature vector representation, while preserving the inherent semantic information of each modality. For example, for text data, recurrent neural networks (RNNs), convolutional neural networks (CNNs), or Transformer encoders can be used for encoding; for image or video data, deep convolutional neural networks (DCNNs) can be used for feature extraction; and for structured numerical data, multilayer perceptrons (MLPs) or simple linear transformations can be used for encoding.

[0080] Feature vector representation refers to the abstraction and compression of raw data into a fixed-length or variable-length numerical vector after modality-specific feature encoding. This vector can effectively capture the key information and semantic features of the original data and serve as the input to subsequent cross-modal fusion networks. These vectors can be high-dimensional dense vectors or sparse vectors after dimensionality reduction, depending on the specific application scenario and computing resources.

[0081] Cross-modal fusion networks are neural network architectures specifically designed to process and integrate features from different data modalities. These networks aim to learn and understand the interrelationships and dependencies between different modalities, thereby generating a more informative joint representation. The network can employ various architectures; for example, it can be a multi-branch network where each branch processes features from one modality and fuses them in intermediate layers; or it can be a unified Transformer architecture that achieves inter-modal interaction through self-attention or cross-attention mechanisms.

[0082] Modeling the interactions between feature vector representations involves using cross-modal fusion networks and attention mechanisms to deeply analyze and understand how features from different modalities influence, complement, or constrain each other. This modeling process aims to reveal the complex patterns of correlation hidden behind heterogeneous data, rather than simply piecing together features. For example, the network can learn that under specific vehicle environmental conditions, the weight of a certain data content feature on risk assessment increases significantly, or that there is a causal relationship between external threat intelligence features and specific user interaction data.

[0083] Cross-modal joint feature representation refers to a unified, high-dimensional feature vector formed by integrating the feature information and interaction relationships of all modalities after processing by a cross-modal fusion network. This joint feature representation is more comprehensive and richer than any single-modal feature representation because it includes the semantic information of all modalities and the deep correlations between them. This representation forms the basis for subsequent generation of correlation analysis results, providing a more comprehensive perspective for risk assessment.

[0084] This application's solution transforms raw, heterogeneous data into unified, semantically rich feature vector representations by performing modality-specific feature encoding on data content features, vehicle environmental state features, and external threat intelligence features. This step ensures that data from different modalities can be effectively compared and fused in subsequent processing. These feature vector representations are then input into a pre-defined cross-modal fusion network. The core of this fusion network lies in utilizing an attention mechanism to deeply model the interaction relationships between the various feature vector representations. Through the attention mechanism, the network can dynamically identify and focus on the most relevant parts between different modal features, thereby capturing their complex, non-linear interactions. For example, when vehicle environmental state features show anomalies, the attention mechanism may enhance its focus on specific keywords in the data content features, or when external threat intelligence features indicate potential attacks, it will pay more attention to abnormal patterns in vehicle operation data. This refined interaction modeling enables the fusion network to generate a comprehensive cross-modal joint feature representation, which not only includes the semantic information of each modality itself, but more importantly, it integrates the deep correlations between modalities. Ultimately, based on this highly integrated cross-modal joint feature representation, more accurate and comprehensive correlation analysis results can be generated. In this way, the proposed solution can overcome the limitations of traditional methods in handling heterogeneous data association, significantly improve the depth and breadth of association analysis, and thus provide a more solid foundation for the subsequent generation of security assessment information and risk assessment information, thereby improving the accuracy and reliability of the overall risk assessment.

[0085] The following is a concrete example. Suppose a cloud platform receives contextual information about a vehicle, including: data content features (e.g., abnormal API call records in the vehicle's internal system logs), vehicle environmental state features (e.g., geofence information of the vehicle's current location, sensor data such as speed, acceleration, and braking frequency), and external threat intelligence features (e.g., an IP address communicating with the vehicle being flagged as malicious). To perform correlation analysis on these multi-source heterogeneous features, they are first encoded using modality-specific features. Specifically, for text data in the system logs, a pre-trained Transformer encoder can be used for feature extraction to generate text feature vectors; for vehicle sensor data, a Multilayer Perceptron (MLP) can be used to convert it into numerical feature vectors; and for external threat intelligence, it can be encoded as binary or embedded vectors. Next, these generated feature vector representations are input into a cross-modal fusion network based on a Transformer architecture. Within this network, multi-head self-attention and cross-attention mechanisms are used to interactively model the feature vectors from different modalities. For example, the cross-attention mechanism allows text feature vectors to "query" sensor feature vectors and threat intelligence feature vectors to identify whether there is a correlation between them. In this way, the network learns that when a malicious IP address communicates with a vehicle, if this is accompanied by specific abnormal API calls and the vehicle's acceleration behavior in unauthorized areas, then there is a strong correlation between these features. Ultimately, the fusion network outputs a cross-modal joint feature representation, which contains comprehensive information from all modalities and their interaction patterns. This joint feature representation is then fed into a binary classifier to generate association analysis results, such as outputting a risk association score between 0 and 1, or directly classifying it as a "high-risk association" or "normal association."

[0086] See Figure 4 In one embodiment, the method for determining a combination of target operations for target data from a plurality of candidate security operations includes, but is not limited to, steps S401 to S403.

[0087] Step S401: Determine the current state of the vehicle based on risk assessment information, context information, and vehicle computing resource information.

[0088] Step S402: With maximizing the cumulative reward function as the optimization objective, the target action sequence is determined from the preset candidate safe operation action space based on the current state.

[0089] The cumulative reward function is used to weight at least two of the following metrics: security gain, efficiency loss, compliance, and data utility.

[0090] Step S403: Generate target operation combination based on target action sequence.

[0091] When determining the current state of a vehicle, this state is crucial information that comprehensively reflects the vehicle's operating environment, safety risks, and its own capabilities. It encompasses not only the safety threats faced by the vehicle (provided by risk assessment information) and its operating environment (characterized by contextual information), but also incorporates the vehicle's own processing capacity limitations (reflected by computing resource information), making the description of the vehicle's state more comprehensive and realistic. For example, a state fusion module can structurally integrate risk assessment information (such as risk level and threat type), contextual information (such as vehicle location, speed, sensor data, and network connectivity status), and computing resource information (such as CPU utilization, memory usage, battery level, and network bandwidth) to form a multi-dimensional vector or state descriptor. Alternatively, machine learning models, such as neural networks, can be used as input to train the model and output a potential representation that characterizes the vehicle's current comprehensive state.

[0092] In the process of determining the target action sequence from a pre-defined candidate security operation action space based on the current state, with the goal of maximizing the cumulative reward function, the cumulative reward function defines the "goodness" or "badness" of different security operation combinations. By maximizing this function, the optimal action sequence in the long run can be selected. This function balances multiple key indicators through weighted operations, avoiding the bias that may arise from optimizing a single objective. These indicators include: security gain indicators (such as the degree of risk reduction, the decrease in attack success rate, or the number of vulnerability fixes); efficiency loss indicators (such as the increase in data processing latency, network bandwidth usage, or computational resource consumption); compliance indicators (such as the degree of compliance with regulatory requirements, privacy protection level, or data usage restrictions); and data utility indicators (such as data availability, data accuracy, or data integrity). For example, a reinforcement learning framework can be used, where the current state of the vehicle serves as the observation of the reinforcement learning agent, the candidate security operation action space serves as the available actions, and the cumulative reward function serves as the reward signal for the reinforcement learning environment. Through training, the reinforcement learning agent can learn a policy that, given the current state, can select the action sequence that maximizes the future cumulative reward. Alternatively, multi-objective optimization algorithms, such as genetic algorithms or particle swarm optimization algorithms, can be used. These algorithms use the cumulative reward function as the objective function and, given the current state and the space of candidate safe operation actions, search for the action sequence that maximizes the value of the function.

[0093] When generating a target operation combination based on a target action sequence, the target action sequence is a series of security measures to be taken within a certain period or in response to a specific event. Transforming it into a target operation combination means concretizing the abstract action sequence into an executable set of security operation instructions. For example, a mapping module can be used to map each action in the action sequence (e.g., "encryption," "de-identification," "authorization") to a predefined specific security operation (e.g., "use AES-256 encryption algorithm," "blur sensitive areas of the image," "grant temporary read-only access") in a one-to-one or one-to-many manner, thus forming a complete operation combination. Alternatively, a rule engine can be used to dynamically generate a series of specific security operation instructions based on the action type and order in the action sequence, combined with current context information; these instructions together constitute the target operation combination.

[0094] This application's solution first integrates risk assessment information, contextual information, and the vehicle's own computational resource information to construct a more comprehensive and accurate description of the vehicle's current operating status. Based on this, instead of simply selecting safe operations, it introduces a decision-making mechanism with the optimization objective of maximizing the cumulative reward function. This cumulative reward function cleverly weights multiple key indicators, such as safety gain, efficiency loss, compliance, and data utility, enabling a comprehensive consideration of various factors during decision-making. By determining an optimal target action sequence from a pre-defined space of candidate safe operation actions based on the current state, it ensures that the selected safety measures not only effectively address the current risk but also fully consider the vehicle's resource constraints and other important non-safety objectives. Finally, this optimized target action sequence is transformed into a specific, executable combination of target operations, thereby achieving the safe processing of target data. This mechanism transforms safety governance decision-making from a static, single-objective orientation to a dynamic, multi-objective optimization orientation, significantly improving the intelligence and adaptability of safety governance.

[0095] The following is a concrete example to illustrate this. Suppose a connected intelligent vehicle is performing autonomous driving, and its cloud platform needs to perform secure processing on the sensor data uploaded by the vehicle. First, the current state of the vehicle is determined based on risk assessment information (e.g., the detection of a network attack risk against the autonomous driving system in the area where the vehicle is located), contextual information (e.g., the vehicle is currently driving on a highway, the amount of sensor data is large, and the network connection stability is generally poor), and the vehicle's computing resource information (e.g., the CPU utilization of the vehicle's onboard computing unit has reached 75%, and memory usage is high). This current state is then fed into the decision-making module as a comprehensive input. The decision-making module optimizes by maximizing the cumulative reward function, which may set the weight of "security gain indicators" (such as reducing the risk of attack) to 0.4, "efficiency loss indicators" (such as data processing latency) to 0.3, "data utility indicators" (such as maintaining data integrity) to 0.2, and "compliance indicators" (such as privacy protection) to 0.1. Based on this, a target action sequence is determined from the preset candidate security operation action space (e.g., including "lightweight encryption of data", "partial desensitization of data", "limiting data upload rate", "enabling intrusion detection system" etc.).

[0096] In some embodiments, the vehicle safety dynamic management method further includes: acquiring historical execution feedback data of the target operation combination; in a simulation environment, performing deduction and verification on the parameters of the cumulative reward function and / or the target action sequence based on the historical execution feedback data to obtain deduction and verification results; and updating the parameters of the cumulative reward function and / or the strategy for determining the target action sequence based on the deduction and verification results.

[0097] The proposed solution acquires historical execution feedback data of the target operation combination and verifies the parameters of the cumulative reward function and / or the strategy for determining the target action sequence in a simulation environment. This allows the system to learn from actual operational experience, identify, and correct shortcomings in the strategy. This iterative optimization process enables the security governance system to learn from historical experience and continuously improve its decision-making capabilities, thereby ensuring that it can always select the optimal or near-optimal combination of security operations under constantly changing security situations, effectively balancing multiple objectives such as security and efficiency.

[0098] The following is a concrete example to illustrate this. When implementing dynamic vehicle safety governance methods, the cloud platform can first acquire historical execution feedback data. For example, when a combination of target operations (such as encrypting specific data and restricting access) is executed on an actual vehicle, information such as the encryption strength, the impact on data transmission latency, the success rate of user access, and whether any data breach attempts have occurred is recorded. This constitutes the historical execution feedback data. Subsequently, this historical execution feedback data is imported into a simulation environment. This simulation environment can be a microservice architecture built on a Kubernetes cluster, in which a simulated vehicle agent, a data flow simulator, a threat simulator, and a security policy enforcement module are deployed. In the simulation environment, this historical data can be used to run a series of simulation experiments to extrapolate and verify the parameters of the cumulative reward function. For example, if the weight of the "efficiency loss indicator" in the cumulative reward function is set improperly, it may lead to an overemphasis on efficiency in security decisions, sacrificing necessary security. By simulating the effects of security decisions under different weight configurations based on historical feedback data in the simulation environment, this potential bias can be discovered, and the weight of the efficiency loss indicator can be adjusted accordingly to achieve a more reasonable balance between security and efficiency.

[0099] In some embodiments, the candidate security operation includes at least one of the following steps: Configure encryption methods for target data based on security situation characteristics; Perform reversible desensitization processing on sensitive areas in the view of the target data; Embed traceability information into target data; Based on the user information of the accessing user, the data tag information of the target data, and / or the environmental risk information of the current environment, dynamic access authorization decisions are made for the target data.

[0100] This application's solution incorporates various specific security operations into the candidate scope, enabling the cloud platform to flexibly select and configure encryption methods for target data based on risk assessment and contextual information during dynamic vehicle safety management. This provides appropriate data confidentiality protection at different risk levels. Simultaneously, for views involving privacy or sensitive information, the platform can reversibly anonymize sensitive areas, protecting data privacy while also considering data usability and compliance requirements in specific scenarios. To enhance data traceability and accountability, the platform can also embed traceability information into the target data, ensuring that any data manipulation can be recorded and audited. Furthermore, regarding data access control, this application introduces a dynamic access authorization decision mechanism. This mechanism comprehensively considers the user information of the accessing user, the data tag information of the target data, and the environmental risk information of the current environment, thereby achieving fine-grained, adaptive access control and effectively preventing unauthorized access and data misuse. By incorporating these diverse, dynamically selectable and configurable security operations into the candidate scope, the governance approach of this application can more comprehensively and accurately address various security threats to vehicle data, providing integrated security protection from multiple dimensions such as data confidentiality, privacy protection, integrity traceability and access control, significantly improving the depth and breadth of vehicle data security governance.

[0101] In some embodiments, the vehicle safety dynamic management method further includes: displaying an interactive management interface; and performing corresponding management operations in response to interactive operations on the interactive management interface.

[0102] The interactive management interface is used to display security posture information, security policy status information, operation log information, data breach incident tracing information, and / or platform health status information. The interactive management interface can be a graphical user interface (GUI) or a command-line interface (CLI), designed to provide administrators with an intuitive and convenient entry point to view system status and perform management operations. This interface can be implemented based on web technology and accessed through a browser, with its backend interacting with the cloud platform; alternatively, the interface can be a standalone desktop application installed on the management terminal, communicating with the cloud platform through a pre-defined application programming interface (API). The core function of the interactive management interface is to present the complex internal operating status and security events to the user in an easily understandable way and to receive user commands.

[0103] Management operations include reviewing, publishing, or rolling back security posture and security policy status information; auditing and analyzing operation logs; tracing the source of data breaches; and monitoring and alerting the cloud platform's health status. Interactive operations refer to actions such as clicking, inputting, and selecting on the interactive management interface. The cloud platform recognizes these actions and triggers predefined management operations. For example, it can capture user interface interactions through an event listener mechanism and translate these interactions into requests to backend services; or it can receive requests from the frontend interface through an API gateway and route them to the corresponding microservices in the cloud platform for processing.

[0104] This application's solution presents the operational status of the vehicle safety dynamic governance method (including obtaining contextual information, risk analysis, determining target operation combinations, and safety processing) and the execution of safety policies, as well as potential safety events, to management personnel in a visual manner through an interactive management interface. When management personnel operate through the interactive management interface, they can respond and execute corresponding management operations in a timely manner, such as reviewing, publishing, or rolling back safety policies, thereby directly influencing or adjusting the "preset multi-objective conditions" or "candidate safety operations" in the automated governance process. Simultaneously, by auditing and analyzing operation log information and tracing the source of data breach incidents, management personnel can supervise and verify the decisions and results of automated safety processing, ensuring its compliance and effectiveness. Furthermore, health status monitoring and alerts are provided for the cloud platform to ensure the stability and reliability of the infrastructure supporting the operation of the safety governance method. This human-machine collaborative mechanism transforms the automated safety governance process from a "black box" into a transparent, controllable, and auditable process, improving the management efficiency and security of vehicle safety dynamic governance.

[0105] This application also provides a cloud platform. See appendix. Figure 5 , Figure 5 This is a schematic diagram of the main structure of a cloud platform according to an embodiment of this application. Figure 5 As shown, the cloud platform in this embodiment mainly includes a processor 501 and a memory 502. The memory 502 can be configured to store a program for executing the vehicle safety dynamic management method of the above-described method embodiments. The processor 501 can be configured to execute the program in the memory 502, which includes, but is not limited to, a program for executing the vehicle safety dynamic management method of the above-described method embodiments. For ease of explanation, only the parts related to the embodiments of this application are shown. For specific technical details not disclosed, please refer to the method section of the embodiments of this application.

[0106] In some possible implementations of this application, the cloud platform may include multiple processors 501 and multiple memories 502. The program executing the vehicle safety dynamic management method of the above-described method embodiments can be divided into multiple subroutines. Each subroutine can be loaded and run by a processor 501 to execute different steps of the vehicle safety dynamic management method of the above-described method embodiments. Specifically, each subroutine can be stored in a different memory 502, and each processor 501 can be configured to execute programs in one or more memories 502 to jointly implement the vehicle safety dynamic management method of the above-described method embodiments. That is, each processor 501 executes different steps of the vehicle safety dynamic management method of the above-described method embodiments to jointly implement the vehicle safety dynamic management method of the above-described method embodiments.

[0107] The aforementioned multiple processors 501 can be processors deployed on the same device. For example, the aforementioned cloud platform can be a high-performance device composed of multiple processors, and the aforementioned multiple processors 501 can be processors configured on that high-performance device. Alternatively, the aforementioned multiple processors 501 can also be processors deployed on different devices. For example, the aforementioned cloud platform can be a server cluster, and the aforementioned multiple processors 501 can be processors on different servers within the server cluster.

[0108] This application also provides a computer-readable storage medium. In one embodiment of the computer-readable storage medium according to this application, the computer-readable storage medium can be configured to store a program that performs the vehicle safety dynamic management method of the above-described method embodiments. This program can be loaded and run by a processor to implement the above-described vehicle safety dynamic management method. For ease of explanation, only the parts related to the embodiments of this application are shown; for specific technical details not disclosed, please refer to the method section of the embodiments of this application. The computer-readable storage medium can be a memory formed by various electronic devices. Optionally, in the embodiments of this application, the computer-readable storage medium is a non-transitory computer-readable storage medium.

[0109] The vehicle safety dynamic governance method, cloud platform, and storage medium provided in this application analyze the risk based on the vehicle's context information. Under multi-objective conditions, based on risk assessment information and context information, a combination of target operations for the target data is determined from multiple candidate safety operations, thereby performing security processing on the target data. Thus, by analyzing the vehicle's context information, dynamic and refined risk assessment is achieved, avoiding overprotection and resource waste in low-risk scenarios. Under preset multi-objective conditions, based on the risk assessment information and context information, a combination of target operations for the target data is determined from multiple candidate safety operations. This means that the generation of security policies is no longer a fixed condition-action rule, but rather a dynamic decision based on real-time risk, vehicle resources, and multiple optimization objectives. This effectively balances the occupation of vehicle computing resources and user experience, improving security governance efficiency and reducing resource consumption.

[0110] Exemplary embodiments of this disclosure have been specifically shown and described above. It should be understood that this disclosure is not limited to the detailed structures, arrangements, or implementations described herein; rather, this disclosure is intended to cover various modifications and equivalent arrangements contained within the spirit and scope of the appended claims.

Claims

1. A dynamic vehicle safety management method, applied to a cloud platform, characterized in that, include: Obtain the vehicle's context information; The context information is used to characterize the current safety status of the vehicle and is related to the target data to be processed; Risk assessment information is obtained by performing risk analysis on the vehicle context information; Under the condition of satisfying preset multi-objective conditions, based on the risk assessment information and the context information, a combination of target operations for the target data is determined from multiple candidate security operations; Based on the target operation combination, the target data is processed securely.

2. The vehicle safety dynamic management method according to claim 1, characterized in that, The context information is generated by the vehicle-side agent through feature extraction of multi-source sampling data obtained from the vehicle. The multi-source sampling data includes vehicle operation data, environmental perception data, user interaction data, network traffic data, and / or system log data. The feature extraction includes using a lightweight feature extraction model to extract data content features, vehicle environmental status features, and / or external threat intelligence features from the multi-source sampling data.

3. The vehicle safety dynamic management method according to claim 1, characterized in that, The risk analysis of the vehicle context information includes: A correlation analysis is performed on the multi-source heterogeneous features in the context information to obtain the correlation analysis results; the multi-source heterogeneous features include data content features, vehicle environment status features, and external threat intelligence features. Based on the correlation analysis results, safety assessment information and risk assessment information are generated; The risk assessment information is generated based on the safety assessment information and the risk assessment information.

4. The vehicle safety dynamic management method according to claim 3, characterized in that, The association analysis of the multi-source heterogeneous features in the context information includes: The data content features, vehicle environment status features, and external threat intelligence features are subjected to modality-specific feature encoding to generate their respective corresponding feature vector representations; The feature vector representations are input into a preset cross-modal fusion network to model the interaction relationship between the feature vector representations based on an attention mechanism, thereby obtaining a cross-modal joint feature representation. The correlation analysis results are generated based on the cross-modal joint feature representation.

5. The vehicle safety dynamic management method according to claim 1, characterized in that, Under the condition of satisfying preset multi-objectives, based on the risk assessment information and the context information, determining the target operation combination for the target data from multiple candidate security operations includes: Based on the risk assessment information, the context information, and the vehicle's computing resource information, the current state of the vehicle is determined; With the goal of maximizing the cumulative reward function, a target action sequence is determined from a preset candidate security operation action space based on the current state; the cumulative reward function is used to perform a weighted calculation on at least two of the security gain index, efficiency loss index, compliance index, and data utility index. Based on the target action sequence, the target operation combination is generated.

6. The vehicle safety dynamic management method according to claim 5, characterized in that, Also includes: Obtain historical execution feedback data for the target operation combination; In a simulation environment, based on the historical execution feedback data, the parameters of the cumulative reward function and / or the target action sequence are deduced and verified to obtain the deduction and verification results; Based on the simulation and verification results, update the parameters of the cumulative reward function and / or determine the strategy for the target action sequence.

7. The vehicle safety dynamic management method according to claim 1, characterized in that, The candidate security operation includes at least one of the following steps: Configure an encryption method for the target data based on the security situation characteristics; Reversible desensitization processing is performed on sensitive areas in the view of the target data; Embed traceability information into the target data; Based on the user information of the accessing user, the data tag information of the target data, and / or the environmental risk information of the current environment, a dynamic access authorization decision is made for the target data.

8. The vehicle safety dynamic management method according to claim 1, characterized in that, Also includes: Display the interactive management interface; The interactive management interface is used to display security situation information, security policy status information, operation log information, data breach incident tracing information, and / or platform health status information; In response to interactive operations on the interactive management interface, corresponding management operations are performed; the management operations include reviewing, publishing or rolling back the security situation information and the security policy status information, auditing and analyzing the operation log information, tracing and analyzing the source of the data leakage event, and monitoring and alerting the cloud platform for its health status.

9. A cloud platform, characterized in that, The cloud platform includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the vehicle safety dynamic management method according to any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the vehicle safety dynamic management method as described in any one of claims 1 to 8.

Images