A vehicle-mounted intelligent device and its skill constraint method, and a computer storage medium.
By encapsulating the capabilities of large language models into skill capsules and digitally signing them, combined with hierarchical desensitization and end-to-end recording, the problems of capability constraints, process tracking, and result auditing of large models in vehicle systems are solved, achieving data security, real-time performance, and traceability, while reducing expansion costs.
Patent Information
- Application Number
- CN202611438187.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-16
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies for deploying large language models in vehicle systems suffer from problems such as boundless capabilities, risks of unauthorized responses, privacy leaks, unpredictable inference latency, insufficient log storage, and high expansion costs, making it difficult to meet vehicle safety and real-time requirements.
By encapsulating each capability of the large model into digitally signed skill capsules, and activating and matching skill capsules when a task request is generated, secure task requests are output. Vehicle data sources are anonymized in a hierarchical manner, forming a full-link record and periodically signing and anchoring it, ensuring data security, real-time traceability and accountability.
It achieves unified intelligent processing of vehicle data under skill constraints, ensuring the safe operation of large models, effective output, and reduced after-sales costs, thus solving the shortcomings of large model capability constraints, process tracking, and result auditing.
Smart Images

Figure CN123067968A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of in-vehicle intelligent devices, and more particularly to an in-vehicle intelligent device and its skill constraint method, as well as a computer storage medium. Background Technology
[0002] As vehicle electronic and electrical architecture evolves from distributed to domain-centralized and even in-vehicle centralized architectures, the centralization of computing power in the central computing unit provides the architectural foundation for deploying large language models in vehicles. The industry is gradually exploring the application of large language models in in-vehicle scenarios such as fault diagnosis, intent understanding, and decision support. At the same time, standards and regulations such as vehicle functional safety (ISO 26262), expected functional safety (ISO 21448), cybersecurity (ISO / SAE 21434, UNECE R155), software upgrades (UNECE R156), and data security ("Several Provisions on the Management of Automotive Data Security") have put forward clear requirements for the controllability, traceability, and data compliance of in-vehicle systems.
[0003] However, directly deploying a general-purpose large language model on the vehicle side has the following technical drawbacks: First, the model's capabilities are boundless, capable of responding to arbitrary requests and lacking a constraint mechanism aligned with vehicle operation permissions, posing risks of unauthorized responses and accidental triggering. Second, the model can autonomously generate tool call requests, and existing solutions lack fine-grained control over the callable toolset, failing to prevent unauthorized operations. Third, the model's inherent illusion problem makes its output unverifiable, and directly using it for safety-critical decisions will introduce unacceptable risks. Fourth, model inference requires access to sensitive data such as location, driving behavior, and cockpit voice, and existing solutions lack differentiated filtering mechanisms based on data levels, posing privacy risks. Fifth, model inference latency is unpredictable, making it difficult to meet the deterministic real-time requirements of vehicle control. Sixth, the inference process lacks complete log storage and anti-tampering protection, making effective backtracking and responsibility determination impossible after anomalies occur. Seventh, adding or adjusting model capabilities requires retraining or full deployment, resulting in high expansion costs, long cycles, and difficulty in rollback.
[0004] In summary, existing technical solutions have significant shortcomings in terms of large model capability constraints, process tracking, and result auditing. There is an urgent need in this field for a technical solution that integrates large language models into vehicle systems in a constrained, traceable, and auditable manner while simultaneously supporting on-site preservation of intermittent faults. Summary of the Invention
[0005] The present invention aims to provide an in-vehicle intelligent device and its skill constraint method and computer storage medium, in order to solve the problems that the current in-vehicle intelligent devices have significant shortcomings in large model capability constraints, process tracking and result auditing.
[0006] To address the aforementioned technical problems, a first aspect of this invention provides a skill constraint method for in-vehicle intelligent devices, applied to in-vehicle intelligent devices, the method comprising: Collect vehicle data sources, and store them securely after classifying and de-identifying the vehicle data sources according to a preset data classification strategy. Each capability of the large model is encapsulated as a digitally signed skill capsule, and the skill capsule is activated and matched when a task request is generated to output a secure task request, wherein the task request carries a vehicle data source that has been hierarchically de-identified. The processed and secure task requests are routed to the corresponding channels for inference to obtain the inference results. The reasoning results are verified, and the reasoning results that pass the verification are output to the reasoning output. The entire chain of task requests, skill capsule matching, reasoning, and verification is recorded in the hash chain ledger and periodically signed and anchored.
[0007] Accordingly, a second aspect of the present invention provides an in-vehicle intelligent device, including a processor and a memory, wherein the memory stores a computer program, and the processor implements the in-vehicle intelligent device skill constraint method provided in the first aspect of the present invention when executing the computer program.
[0008] Accordingly, a third aspect of the present invention provides a computer storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, cause the processor to perform the in-vehicle intelligent device skill constraint method provided in the first aspect above.
[0009] Compared with existing technologies, this invention provides an in-vehicle intelligent device, its skill constraint method, and a computer storage medium. The skill constraint method collects vehicle data sources, performs hierarchical anonymization on the vehicle data sources according to a preset data hierarchical strategy, and then securely stores the data. Each capability of a large model is encapsulated as a digitally signed skill capsule, which is activated and matched when a task request is generated, outputting a secure task request. The task request carries the hierarchically anonymized vehicle data source. The processed secure task request is then routed to the corresponding channel for inference to obtain the inference result. The inference result is verified, and the verified inference result is output to the inference output. The entire chain of task request, skill capsule matching, inference, and verification is recorded in a hash chain ledger and periodically signed and anchored. This enables unified intelligent processing of vehicle data under skill constraints, ensuring the safe operation and effective output of the large model, guaranteeing data security, effectiveness, real-time processing, traceability, and accountability, and reducing after-sales costs. This addresses the significant shortcomings of current in-vehicle intelligent devices in terms of large model capability constraints, process tracking, and result auditing. Attached Figure Description
[0010] One or more embodiments are illustrated by way of example with reference numerals in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.
[0011] Figure 1 This is a structural schematic diagram of an in-vehicle intelligent device provided by the present invention; Figure 2 This is a flowchart illustrating a method for constraining the skills of an in-vehicle intelligent device provided by the present invention; Figure 3 This is a schematic diagram of the structure of a processor in an in-vehicle intelligent device provided by the present invention; Figure 4 This is a structural schematic diagram of a skill constraint plane module in an in-vehicle intelligent device provided by the present invention; Figure 5 This is a schematic diagram illustrating the execution of the five ironclad rules of the guard submodule in a vehicle-mounted intelligent device provided by the present invention. Figure 6 This is a schematic diagram of the workflow of the inference plane module in an in-vehicle intelligent device provided by the present invention; Figure 7 This is another flowchart illustrating a method for constraining the skills of an in-vehicle intelligent device provided by the present invention; Figure 8 This is a schematic diagram of the execution flow of the fault analysis skill capsule in the skill constraint method for in-vehicle intelligent devices provided by the present invention. Detailed Implementation
[0012] To facilitate understanding of the present invention, a more detailed description is provided below with reference to the accompanying drawings and specific embodiments. It should be noted that when an element is described as being "fixed to" another element, it can be directly on the other element, or one or more intermediate elements may exist between them. When an element is described as being "connected" to another element, it can be directly connected to the other element, or one or more intermediate elements may exist between them. The terms "upper," "lower," "inner," "outer," "bottom," etc., used in this specification indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. Furthermore, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance.
[0013] Unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. The term "and / or" as used in this specification includes any and all combinations of one or more of the associated listed items.
[0014] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.
[0015] Please see Figure 1 , Figure 1 This is a schematic diagram of the structure of an in-vehicle intelligent device 900 provided in an embodiment of the present invention. In terms of hardware implementation, the in-vehicle intelligent device 900 can be a TBOX (Telematics-BOX).
[0016] The in-vehicle intelligent device 900 includes: one or more processors 901, a memory 902, and one or more computer programs stored in the memory 902 and executable on the processor 901. The memory 902 and the processor 901 are coupled together via a bus system 903. The one or more computer programs are executed by the processor 901 to perform the corresponding functions in the methods described in the following method embodiments.
[0017] The processor 901 can be a Central Processing Unit (CPU), a Network Processor (NP), an Electronic Control Unit (ECU), a hardware chip, or any combination thereof. The aforementioned hardware chip can be an Application Specific Integrated Circuit (ASIC), a Programmable Logic Device (PLD), or a combination thereof. The aforementioned PLD can be a Complex Programmable Logic Device (CPLD), a Field-Programmable Gate Array (FPGA), a Generic Array Logic (GAL), or any combination thereof.
[0018] Memory 902 is used to store program code, etc. Memory 902 may include volatile memory (VM), such as random access memory (RAM); memory 902 may also include non-volatile memory (NVM), such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid-state drive (SSD); memory 902 may also include combinations of the above types of memory.
[0019] The memory 902 can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the program instructions / modules corresponding to the methods in the embodiments of the present invention. The processor 901 executes various functional applications and data processing of the method and apparatus by running the non-volatile software programs, instructions, and modules stored in the memory 902, that is, it implements the functions of the various modules or units of the method and apparatus provided in the following method embodiments.
[0020] The memory 902 may include a program storage area and a data storage area. The program storage area may store the operating system and application programs required for at least one function. The data storage area may store data created based on the use of the device.
[0021] The one or more modules are stored in the memory 902. When executed by the one or more processors 901, they perform the methods in any of the following method embodiments. For example, they perform the method steps described in the following method embodiments to implement the functions of the modules described in the following device embodiments.
[0022] Please see Figure 2 This invention provides a skill constraint method for in-vehicle intelligent devices based on a large model, applicable to in-vehicle intelligent devices. The method includes: S1. Collect vehicle data sources, and store them securely after classifying and de-identifying the vehicle data sources according to a preset data classification strategy. S2. Encapsulate each capability of the large model into a digitally signed skill capsule, and activate and match the skill capsule when a task request is generated to output a secure task request, wherein the task request carries a vehicle data source that has been graded and desensitized. S3. The processed secure task requests are routed to the corresponding channels for inference to obtain the inference results; S4. Verify the reasoning results and output the valid reasoning results to the reasoning output. S5. Record the entire chain of task requests, skill capsule matching, reasoning, and verification into the hash chain ledger, and periodically sign and anchor it.
[0023] This embodiment provides a method for skill constraint of in-vehicle intelligent devices based on a large model. It collects vehicle data sources, performs hierarchical anonymization on the vehicle data sources according to a preset data hierarchical strategy, and then securely stores the data. Each capability of the large model is encapsulated as a digitally signed skill capsule, which is activated and matched when a task request is generated, outputting a secure task request. The task request carries the hierarchically anonymized vehicle data source. The processed secure task request is then routed to the corresponding channel for inference to obtain the inference result. The inference result is verified, and the verified inference result is output to the inference output. The entire chain formed by task request, skill capsule matching, inference, and verification is recorded in a hash chain ledger and periodically signed and anchored. This method enables unified intelligent processing of vehicle data under skill constraints, ensuring the safe operation and effective output of the large model, guaranteeing data security, effectiveness, real-time performance, traceability, and accountability, and reducing after-sales costs. This addresses the significant shortcomings of current in-vehicle intelligent devices in terms of large model capability constraints, process tracking, and result auditing.
[0024] In one embodiment, in step S1, vehicle data sources are collected, and the vehicle data sources are classified and de-identified according to a preset data classification strategy before being securely stored.
[0025] Specifically, please refer to the following: Figure 3 The processor 901 is equipped with a data plane module 10, which is used to collect vehicle data sources, perform hierarchical desensitization on the vehicle data sources according to a preset data hierarchical strategy, and then store them securely.
[0026] It is understandable that the process of collecting vehicle data sources in step S1, and then safely storing the vehicle data sources after performing hierarchical desensitization according to the preset data hierarchical strategy, is also the specific processing process of the data plane module 10.
[0027] In step S1, vehicle data sources are collected, and after being classified and anonymized according to a preset data classification strategy, they are securely stored. This specifically includes: data collection and adaptation, data classification, an anonymization engine, and secure storage. The specific process is as follows: S11. Data Acquisition and Adaptation: The data plane module collects vehicle data sources, which include at least one of the following: vehicle static data, bus signals, operation logs, diagnostic data, cockpit inputs, and environmental data.
[0028] S12. Data Classification: The data plane module classifies vehicle data sources according to a preset data classification strategy. This strategy includes four levels: Level 0 (public data), Level 1 (sensitive in-vehicle data), Level 2 (privacy data), and Level 3 (control data). Level 0 public data includes basic vehicle attributes, announcement and certification information, and non-personalized static data. The disclosure or release of this type of data will not cause any harm to vehicle owner privacy, vehicle safety, or public interest. For example, Level 0 public data may include: vehicle announcement parameters such as brand, model name, curb weight, body dimensions, wheelbase, and driving range (CLTC cycle); basic software information such as the vehicle's operating system version number, map engine version number, and publicly available API interface documentation; and static environmental information such as the current administrative division (province / city level only) and traffic congestion index without facial features.
[0029] Level 1 (L1) in-vehicle sensitive data consists of fundamental signals reflecting the vehicle's real-time mechanical status, control unit parameters, and the surrounding physical environment. This type of data does not directly identify a specific individual, but if maliciously exploited, it can reveal defects in the vehicle's control logic or, in combination with other data, indirectly pinpoint specific targets. For example, L1 in-vehicle sensitive data may include: powertrain and chassis operating conditions, such as engine speed, motor torque percentage, battery cell voltage / temperature, tire pressure monitoring system (TPMS) values, and yaw rate; driver operation signals (non-identity-related), such as steering wheel grip force, accelerator / brake pedal travel percentage, and gear position; and raw environmental perception data, such as millimeter-wave radar point clouds before AI recognition, raw LiDAR data, rain sensor values, and light intensity values.
[0030] Level 2 privacy data refers to data that, alone or in combination with other information, can identify a specific natural person's identity, whereabouts, biometric characteristics, or financial status. This type of data is strictly governed by the Personal Information Protection Law and falls under the category of sensitive personal information. For example, Level 2 privacy data may include: biometric identification, such as facial feature values, pupil gaze area, voiceprint features, and fingerprint unlock data collected by a DMS (Driver Monitoring System) camera; precise whereabouts tracking, such as the vehicle's historical latitude and longitude trajectory over seven consecutive days, precise coordinates of the user's permanent residence (home / work), and specific house numbers from navigation search history; in-vehicle private life data, such as images of passengers captured by in-vehicle cabin cameras and audio content of in-vehicle conversations recorded by microphones; and personal digital assets, such as third-party accounts logged into the vehicle's infotainment system (WeChat / Music), contact synchronization records, and payment transaction records.
[0031] Level 3 control data refers to instruction-level data that directly intervenes in the vehicle's core actuators, such as longitudinal (acceleration / deceleration), lateral (steering), and high-voltage power-on / off. This type of data carries extremely high functional safety (ISO 26262) and cybersecurity (R155) risks. For example, Level 3 control data may include: remote control commands, such as remote unlocking / locking, remote air conditioning activation, remote vehicle location with flashing lights and horn, and remote summoning (parking in / out) commands initiated by an app; chassis dynamic control, such as target values for steer-by-wire angles, wheel-end braking torque distribution requests, real-time output torque requests for drive motors, and air suspension height adjustment commands; firmware upgrade commands, such as binary executable files and start-up flags written to the core ECU (Electronic Control Unit) during OTA (Over-The-Air) updates; and high-voltage safety commands, such as high-voltage relay disconnect / close commands issued by the BMS (Battery Management System) and airbag deployment commands triggered by collision signals.
[0032] S13. Data Masking: For vehicle data sources that have already undergone data classification, masking is performed according to their corresponding masking strategies. Specifically, this includes: Level 0 public data itself does not involve personal information or security risks and can be published without anonymization. Therefore, the necessity for anonymizing Level 0 public data is extremely low. Anonymization strategies for Level 0 public data include: no anonymization required, but technical measures to verify data integrity and prevent malicious tampering can be adopted. Specifically, no anonymization processing is required, and it can be directly provided or publicly released; however, data integrity must be verified to prevent malicious tampering (e.g., through digital signatures or verification codes); if Level 0 public data is stored or displayed together with Level 2 privacy data and / or Level 3 control data, it must be isolated through data classification and tagging to avoid indirect exposure of sensitive information due to correlation analysis.
[0033] While Level 1 in-vehicle sensitive data does not directly target individuals, it may indirectly expose vehicle control logic or, combined with other data, locate specific targets. Therefore, it needs to be anonymized before being transmitted outside the vehicle. Thus, the necessity for anonymizing Level 1 in-vehicle sensitive data is moderate. Anonymization strategies for Level 1 in-vehicle sensitive data include: anonymization before transmission outside the vehicle, which can be achieved through techniques such as accuracy reduction, feature extraction, and / or aggregated statistics.
[0034] For example, for powertrain / chassis operating data (engine speed, motor torque, battery voltage), anonymization can be achieved by downgrading accuracy and aggregating statistics. For instance, precise values can be replaced with ranges (such as "torque 65-75%)" or statistical values (mean, quantiles), or aggregated by time window (such as replacing millisecond-level sampling with per-minute averages).
[0035] For raw sensor data (radar point clouds, lidar arrays), feature-level desensitization can be performed. For example, detailed information such as reflection intensity can be removed from the point cloud, retaining only feature-level data such as obstacle outlines; or the density of the original point cloud can be reduced by downsampling.
[0036] For environmental sensing data (rainfall, light intensity values), direct transmission or simple generalization can be used for anonymization. This type of data does not involve personal information and can be used directly; if it is necessary to protect the privacy of the driving environment, the precise values can be generalized into levels (such as "heavy rain" or "strong light").
[0037] Level 2 privacy data falls under the jurisdiction of the Personal Information Protection Law as sensitive personal information and must undergo strict anonymization. Therefore, the anonymization of Level 2 privacy data is extremely necessary. Anonymization strategies for Level 2 privacy data include: in-vehicle processing with strict anonymization, employing techniques such as obfuscation, differential privacy, pseudonymization, and / or local encryption. Specifically, this involves local encrypted storage within the vehicle, with the key stored in a secure chip (SE); data should generally not be transmitted back to the cloud; if transmission is absolutely necessary (e.g., in emergency situations), individual user authorization (not bundled authorization) is required; one-click deletion and data anonymization capabilities should be provided; and the consent period for processing sensitive personal information should not be set to "permanent" or "always allowed."
[0038] For example, regarding the privacy of biometric data (face, voiceprint, fingerprint): For face images / videos, local contouring / blurring / mosaic can be used for desensitization. For instance, the detected face area can be blurred, mosaicked, or contoured to eliminate identity information while retaining the semantics of the scene. For voiceprints / voices, local speech recognition combined with deletion of the original audio can be used for desensitization. For instance, a local speech recognition model can be used in the vehicle to extract semantic commands, and the original audio stream can be deleted immediately without retaining the speaker's characteristics. For fingerprint features, feature value hashing combined with deletion of the original data can be used for desensitization. For instance, only the irreversible hashed feature value is stored, and the original fingerprint template is encrypted and stored in the vehicle's security chip without being uploaded to the cloud.
[0039] For privacy-sensitive location tracking data (GPS coordinates, historical paths): spatial fuzzification can be used for anonymization, such as replacing precise coordinates with a circular area with a radius of 50–200 meters, or mapping them to a preset "Point of Interest Grid" (POI Grid), to prevent track traceability; and / or, precision downgrading can be used for anonymization, such as reducing latitude and longitude precision from meters to hundreds of meters or kilometers; and / or, differential privacy noise can be used for anonymization, such as adding controllable noise to the track data to make individual tracks indistinguishable while maintaining the availability of statistical features; and / or, time generalization can be used for anonymization, such as generalizing precise timestamps to time periods (e.g., "morning" or "afternoon"), thus destroying the traceability of the time dimension.
[0040] For privacy data related to private life within the vehicle (cabin images, recordings): anonymization can be achieved by processing the data within the vehicle and not collecting it by default. For example, cabin data should not be collected by default and should be processed within the vehicle and not transmitted outside the vehicle. If it is necessary to provide the data outside the vehicle, the user's separate consent must be obtained. And / or, anonymization can be achieved by using local encrypted storage. For example, if temporary storage is required, it must be encrypted using a security chip (SE) at the vehicle end, with the key stored in the hardware security module.
[0041] For privacy data of personal digital assets (accounts, contacts, payment records): de-identification / pseudo-identification can be used for anonymization, such as replacing account IDs, device IDs, etc. with random tokens or pseudonyms, so that the data cannot be associated with individuals without additional information; and / or, HMAC salted hashing can be used for anonymization, such as using HMAC-SHA256 salted hashing to generate irreversible pseudo-identifiers for device identifiers such as MAC addresses, and decoupling them from user accounts.
[0042] Level 3 control data involves vehicle longitudinal / lateral control commands. Desensitization would compromise the executability of these commands; its core focus is on preventing tampering, replay attacks, and hijacking. Therefore, desensitization of Level 3 control data is extremely necessary and not applicable. Desensitization strategies for Level 3 control data include: not using desensitization, but employing digital signatures, end-to-end encryption, and / or physical isolation as protective measures. Specifically, this involves the highest level of security protection; command transmission must use HSM encryption and digital signatures to prevent replay attacks and man-in-the-middle attacks; and adding independent physical verification at the execution end to ensure that the driver has the highest priority when there is a conflict between the "command" and the "driver's intent."
[0043] For example, for remote control commands (remote unlocking, remote air conditioning, remote summoning), digital signatures and anti-replay measures can be used for desensitization. For instance, each command can be digitally signed using an HSM (Hardware Security Module), and the receiving end can execute it after verifying the signature. Timestamps and nonces can be added to prevent replay attacks.
[0044] For chassis dynamic control (steering angle, braking torque, drive torque), end-to-end encryption and independent verification can be used for desensitization. For example, command transmission uses commercial cryptography for end-to-end encryption; independent physical verification is added to the execution end (such as detecting the pedal status when braking) to ensure that the driver has the highest priority.
[0045] For OTA firmware upgrades (ECU executable files), de-identification processing can be performed using encrypted transmission and integrity verification. For example, the firmware package can be transmitted using SM4 / AES encryption, and the receiving end can verify the digital signature and hash value to ensure integrity and trustworthiness of the source.
[0046] S14. Secure storage: Securely store the vehicle data source after data anonymization.
[0047] In this embodiment, the data plane module encompasses a four-level data classification system: L0 public, L1 in-vehicle sensitive, L2 privacy, and L3 control, along with differentiated anonymization and control strategies for each level. This achieves hierarchical protection of in-vehicle data throughout its entire lifecycle, real-time anonymization processing before transmission outside the vehicle, a dynamic balance between privacy protection and data utility, minimal external transmission of sensitive data, and a traceable closed-loop data flow in collaboration with the audit plane. Specifically: I. Implement differentiated hierarchical management of vehicle data throughout its entire lifecycle. This involves establishing a four-level data classification system: L0 open, L1 in-vehicle sensitive, L2 privacy, and L3 control. Differentiated collection, storage, transmission, and processing strategies are applied to each level of data. Specifically: L0 data can flow freely; L1 data must be anonymized before transmission outside the vehicle; L2 data is encrypted and stored in the vehicle and is not transmitted outside the vehicle by default; and L3 data uses hardware-level encryption and physical isolation. This ensures that the level of data control matches its security risk level, avoiding either overly lenient or insufficient protection.
[0048] II. Building Real-Time Desensitization and Anonymization Capabilities Before Data Transmission from Outside the Vehicle. By deploying a real-time desensitization pipeline before data leaves the vehicle, accuracy downgrading and feature extraction are performed on raw L1-level sensor data. For L2-level biometric features and precise trajectories, blurring, differential privacy noise addition, or pseudonym replacement are performed to ensure that data uploaded to the cloud no longer possesses the ability to directly or indirectly identify individuals. Sensitive information such as faces and license plates in video and image data is anonymized at the vehicle end before being output.
[0049] Third, achieve a dynamic balance between privacy protection and data utility. Through the refined design of a graded de-identification strategy, while removing personally identifiable information, the analytical value and usability of the data are preserved to the maximum extent, avoiding the loss of data utility due to excessive de-identification.
[0050] Fourth, achieve localized processing and minimize the external transmission of sensitive data. By forcibly limiting operations such as L2-level biometric recognition, cockpit image and audio processing to the vehicle, the raw data does not leave the vehicle by default; at the same time, following the "minimum necessary" principle, only necessary fields that have been de-identified are uploaded, fundamentally reducing the privacy leakage risks and compliance costs caused by data transmission.
[0051] V. Form a traceable closed loop for data flow in collaboration with the auditing platform. By binding TraceIDs to each stage of data collection, de-identification, and transmission, and pushing audit events to the auditing platform, the hierarchical label, de-identification operation type, external transmission target, and timestamp of each piece of data are recorded to ensure that the entire data flow is traceable, the operation behavior is auditable, and the compliance status is verifiable.
[0052] In one embodiment, in step S2, each capability of the large model is encapsulated as a digitally signed skill capsule, and the skill capsule is activated and matched when a task request is generated, wherein the task request carries a vehicle data source that has been hierarchically de-identified.
[0053] Specifically, please refer to Figure 3 The processor 901 is equipped with a skill constraint plane module 20, which encapsulates each capability of the large model into a digitally signed skill capsule and activates and matches the skill capsule when a task request is generated, wherein the task request carries a vehicle data source that has been hierarchically de-identified.
[0054] In this invention, a large model is introduced into the processor 901, and its operation is constrained by a skill capsule constraint mechanism. For example, the large model is a large language model, such as: Gemini series large models (e.g., Gemini 3.1 Pro, Gemini 3.5 Flash), ChatGPT series large models (e.g., GPT-5.4, GPT-5 Turbo), Qwen series large models, GLM series large models (e.g., GLM-5.2), Claude series large models (e.g., Claude Opus 4.6, Sonnet 4.6), and DeepSeek series large models (e.g., DeepSeek-V3, DeepSeek-R1).
[0055] The Skill Constraint Plane Module 20 is located before the Reasoning Plane Module 30. Its logic is: without a valid skill capsule, no reasoning will be triggered, ensuring that the large model can only operate within the predefined, authorized, and controllable capability boundaries.
[0056] It is understandable that the process of encapsulating each capability of the large model into a digitally signed skill capsule in step S2, and activating and matching the skill capsule when a task request is generated, is also the specific processing process of the skill constraint plane module 20.
[0057] Please refer to the following: Figure 4 The skill constraint plane module 20 includes a skill repository submodule 21, a capsule loading submodule 22, a skill arbitrator 23, and a runtime guard submodule 24.
[0058] In step S2, each capability of the large model is encapsulated as a digitally signed skill capsule, and the skill capsule is activated and matched when a task request is generated, and a secure task request is output. The task request carries a vehicle data source that has been anonymized and graded, specifically including steps S21-S24.
[0059] S21, the Skill Repository Submodule 21 encapsulates each capability of the large model into a digitally signed skill capsule and stores it, i.e., signed capsule storage.
[0060] The Skill Repository submodule 21 is responsible for encapsulating each capability of the large model into a skill capsule with its own digital signature and storing it persistently. The Skill Repository submodule 21 includes a skill repository. The specific implementation process includes: S211, Capability Abstraction and Capsule Structure Definition.
[0061] The Skill Repository submodule 21 first standardizes and decomposes the large model capabilities that can be opened on the vehicle side. Each capability of the large model is abstracted into an independent skill capsule. The internal structure of the skill capsule includes: a unique capability identifier (Skill ID), a capability semantic description (for matching), a trigger condition template, capsule prompt words, whitelist tools (e.g., calling the navigation API, querying vehicle status), the minimum required data permission level (L0~L3), and an output contract template (including output format schema, numerical security boundaries, and prohibited sensitive words).
[0062] S212, Digital Signature Encapsulation.
[0063] Before each skill capsule is released to the skill repository, the vehicle manufacturer's security center digitally signs the entire contents of the skill capsule using a private signing key. The signature value serves as an unforgeable tag for the skill capsule and is appended to the end of the skill capsule package. The signing process also includes the skill capsule version number and timestamp, ensuring that if an older version is exploited, the signature can be invalidated through a version upgrade.
[0064] S213, Hierarchical storage and index construction.
[0065] Skill Repository Submodule 21 stores skill capsules in the vehicle-side secure storage area according to their security level, specifically as follows: System Core Area: This system core area is read-only and is used to store high-risk skill capsules (corresponding to L3 control) involving chassis control and safety early warning. The signature key is verified by the HSM firmware. User Function Area: The user function area is updatable and is used to store common skill capsules such as cockpit interaction, entertainment and navigation. It can be incrementally updated via OTA.
[0066] To accelerate subsequent matching, the skill repository submodule 21 also maintains a lightweight inverted index (based on skill semantic tags and trigger condition keywords) to pre-associate skill capsules with common scenario descriptions.
[0067] S214, Incremental synchronization and anti-tampering protection.
[0068] The Skill Repository submodule 21 periodically connects to a trusted cloud source to retrieve newly added or updated skill capsule packages and writes them into the skill repository. Before a new skill capsule is written into the skill repository, its source legitimacy must be verified through signature verification.
[0069] The metadata of the skill repository itself (including the index) is periodically hashed and reported to the audit plane module for anchoring, ensuring that the overall integrity of the skill repository can be traced.
[0070] S22, Capsule Loading Submodule 22 activates the skill capsule, i.e., signature verification and dependency verification, when a task request (vehicle event / user intent) is generated.
[0071] The capsule loading submodule 22 includes a capsule loader, which is responsible for activating the skill capsules to be used from the stored state to the runnable state when a task request is generated. The specific implementation process includes: S221, Triggered loading and preloading strategies.
[0072] When the system starts or a user logs in, the capsule loader does not load all skill capsules in batches. Instead, it uses a preloading strategy: basic safety skill capsules (such as emergency braking assist) are loaded first as resident capsules; the remaining skill capsules are only loaded into memory when a task request arrives.
[0073] When a task request arrives, the capsule loader locates candidate skill capsules from the skill repository based on the intent tag carried in the task request.
[0074] S222, Double verification before loading into memory (signature verification and dependency verification).
[0075] The capsule loader reads the complete binary package of the candidate skill capsule from storage and first verifies the digital signature of the skill capsule using the automaker's root certificate public key pre-installed in the security chip. If the signature is invalid or the certificate chain is untrusted, the skill capsule is marked as illegitimate, loading is refused, and an audit alert is triggered.
[0076] After the signature verification is successful, the capsule loader enters the dependency verification phase, which includes: Tool dependency verification: Check if each API in the whitelist tools declared in the skill capsule exists in the current system version and is compatible with the interface; Hardware dependency verification: Check whether the computing power threshold required for the skill capsule to run meets the current vehicle system load conditions; Data dependency verification: Check whether the minimum data permission level required by the skill capsule (such as L2 facial recognition) has been authorized by the user in the current driving scenario.
[0077] S223, Activation State Context Assembly.
[0078] After all verifications pass, the capsule loader creates an independent runtime sandbox context for the skill capsule in memory, loads the skill capsule's capsule hint, whitelist tool, data permission tag, and output contract template into the context, marks the skill capsule as active, and registers it in the skill arbitrator's candidate pool. Skill capsules that fail to load or time out will not enter the candidate pool.
[0079] When a task request is generated, S23 and Skill Arbitrator 23 match the task request with activated skill capsules. If the match fails, the task request is rejected and transferred to the audit plane module for auditing. If the match is successful, the task request is transferred to the runtime guard submodule for further processing.
[0080] Skill Arbitrator 23 is the decision-making center, responsible for matching activated skill capsules to task requests, thus achieving task-skill matching. Specifically, it includes: S231, Standardization of task request semantics.
[0081] When a vehicle event or user intent arrives, the skill arbitrator 23 first transforms it into a standardized task description tuple, which consists of an action domain, an object domain, and a constraint domain.
[0082] For example, if the user intent is "I'm a little cold", the skill arbitrator will transform the user intent into a standardized task description tuple, such as: {Action: "Adjust", Object: "Temperature", Constraint: "Raise by 2 degrees"}.
[0083] S232, Multi-dimensional matching algorithm.
[0084] The skill arbitrator 23 performs multi-dimensional matching between standardized tasks and the ability semantic descriptions and trigger condition templates of all activated skill capsules, including at least one of the following: semantic similarity matching, which calculates the cosine similarity between the task description and the ability semantic description of the skill capsule through embedded vectors; and contextual condition matching, which compares whether the vehicle state (e.g., vehicle speed, gear, geographical location) triggered by the task meets the preset pre-trigger conditions of the skill capsule.
[0085] The matching process uses a threshold system, including: when the highest matching score is lower than a preset threshold, the matching is considered unsuccessful; if multiple skill capsule scores are higher than the preset threshold, the skill capsule with the highest score is selected.
[0086] S233, Matching failure handling link.
[0087] If a match fails, the skill arbitrator 23 immediately terminates the subsequent reasoning process for the task request. Simultaneously, the skill arbitrator 23 encapsulates the failed match record (including task description, candidate capsule list, and reason for failure) into an audit payload, pushes it to the audit plane module 50 via an internal interface, and binds it to a newly generated TraceID for record-keeping. This is used for subsequent investigations to determine if there are any missing skills in the skill repository or misunderstandings of intent.
[0088] S234, Match successful, transfer.
[0089] If the match is successful, the skill arbitrator 23 binds the task request to the memory pointer of the activated skill capsule to form a "task-skill binding pair", and together with the vehicle data source in the task, it is handed over to the runtime guard submodule 24 for the next stage of clamping processing.
[0090] S24, the runtime guard submodule 24 processes the successfully matched task requests based on five preset ironclad rules, and outputs the safe task requests to the inference plane module 30. The five ironclad rules include: no operation without skills, whitelist tools cannot be overridden, data is accessed at different levels, output must meet the contract, and full traceability.
[0091] Please see Figure 5 The runtime guard submodule 24 is the final execution checkpoint of the skill constraint plane module 20. It receives task-skill binding pairs, performs forced assembly and filtering of the inference pre-environment according to the five ironclad rules, and then delivers the safe task request to the inference plane module 30.
[0092] Ironclad Rule 1: No skill required, no action (R1).
[0093] The "no skill, no execution" rule can also be described as "no match, no rejection." The runtime guard submodule 24 forcibly checks at the processing entry point whether the incoming context carries a valid, activated skill capsule. If the skill capsule pointer is null or the skill capsule has been marked as invalid, the runtime guard submodule 24 directly intercepts the task, without sending any request to the inference engine of the inference plane module 30. This interception action synchronously triggers the audit plane module 50 to record a "no skill interception" event.
[0094] Rule #2: Whitelist tools cannot be bypassed (R2).
[0095] When constructing the runtime environment of the inference engine, the runtime guard submodule 24 extracts whitelisted tools (such as a list of allowed API calls) from the skill capsule. The runtime guard submodule 24 attaches a tool call proxy outside the inference engine. This proxy intercepts all external tool call requests issued during inference, checks whether the tool call is a whitelisted tool in the current skill capsule, and only allows it if the target tool name exists in the current skill capsule's whitelist; otherwise, the tool call is silently blocked, and a standard error indicating that the tool is unavailable is returned to the inference engine, preventing large models from making phantom calls to unauthorized system interfaces.
[0096] Rule Three: Access data by level (R3).
[0097] Runtime guard submodule 24 reads the minimum permission level of data required for the skill from the skill capsule. When processing the vehicle data source carried by the task, runtime guard submodule 24 performs field-by-field filtering, specifically including: If the data level is less than or equal to the capsule permission level, the vehicle data source is allowed to be input as is into the inference engine of the inference plane module 30. If the data level is greater than the capsule permission level (e.g., the capsule only authorizes access to L1 vehicle speed, but the task attempts to pass in L2 precise coordinates), the runtime guard submodule 24 will automatically desensitize and replace the field (e.g., blur the precise coordinates to the city area level) or directly remove the field to ensure that the inference engine can only access input data that does not exceed the skill capsule's permissions.
[0098] Rule Four: Output must satisfy the contract (R4).
[0099] After the inference engine outputs its inference results, runtime guard submodule 24 does not release it immediately. Instead, it first loads the output contract template from the skill capsule. Runtime guard submodule 24 then performs a forced match between the inference results and the contract, specifically including: Format validation: Checks whether the output conforms to the agreed-upon JSON / structured format; Content boundary verification: whether the control command value is within the physical safety range preset in the contract (e.g., temperature adjustment does not exceed ±5℃). If the verification fails, the runtime guard submodule 24 discards the inference result, triggers a safety fallback response (such as returning a general failure message), and records the contract violation to the audit plane module 50.
[0100] Rule Five: Leave a complete record (R5).
[0101] During the entire lifecycle of task processing (pre-filtering, inference, and post-validation), the runtime guard submodule 24 sends key audit events to the audit plane module 50 in a synchronous interception manner: the hash of the injected prompt word digest, the sequence of allowed tool calls, the list of filtered data fields, and the output contract validation results.
[0102] All audit events share the TraceID of the current task, ensuring that the skill execution forms a complete closed loop in the audit chain, meeting the compliance requirements of traceability and non-repudiation.
[0103] In this invention, the skill constraint plane module 20 includes four sub-modules: a skill repository sub-module, a capsule loading sub-module, a skill arbitrator, and a runtime guard sub-module. When executing a complete task, its flow sequence is as follows: 1. When a task request arrives, the skill arbitrator performs semantic matching; 2. If a match fails, reject the application and forward it for audit and evidence storage. 3. Upon successful matching, the capsule loader in the capsule loading submodule performs signature verification and dependency verification, and activates the skill capsule; 4. Bind the task and capsule, and transfer it to the runtime guard submodule; 5. During runtime, the guard submodule executes Iron Rule 2 (tool filtering) and Iron Rule 3 (data filtering) in sequence, and then injects capsule prompt words into the inference engine; 6. The inference engine generates output inference results, and the guard submodule executes Iron Rule Four (output contract verification) during runtime. 7. Upon successful verification, the final inference result is released to the execution layer or security arbitration access control. The entire process follows Ironclad Rule Five, pushing audit snapshots of each stage to the audit plane module.
[0104] In summary, this skill constraint plane module 20, through a skill repository, capsule loading, skill arbitration, runtime guards, and five ironclad rules, digitally signs and encapsulates the capabilities of large models, performs multi-dimensional matching and mandatory guarding at runtime, ensuring that only authorized skills can access, only whitelisted tools can call, only compliant data can be entered into the model, and only secure outputs can be distributed. The entire process is bound to TraceID for on-chain traceability, thus constructing a deterministic security boundary and traceable anti-tampering mechanism covering the entire "access—execution—output—audit" chain for generative AI on the vehicle side. Specifically: I. Achieve manageable encapsulation and least privilege execution of large model capabilities. By encapsulating each large model capability into a digitally signed capsule with an embedded tool whitelist and data permission tags, and combining this with runtime guards to enforce filtering of API calls and data fields, we ensure that large model inference only calls system resources within preset authorization boundaries, fundamentally preventing unauthorized access and data abuse.
[0105] Second, ensure deterministic access to the inference entry point and rapid isolation of invalid tasks. A skill arbitrator performs multi-dimensional semantic and contextual matching on task requests. If a match fails, the request is immediately rejected and transferred for auditing and evidence storage. This prevents invalid, unknown, or malicious task requests from entering the inference engine, thereby reducing unnecessary consumption of vehicle-side computing resources at the source and improving the overall system efficiency.
[0106] Third, ensure consistency between generative output and vehicle physical safety boundaries. By using runtime guards to force verification of whether the inference output conforms to the output contract template preset by the skill capsule (including format specifications and numerical safety range), combined with data-level filtering and desensitization processing on the input side (L0~L3), the generation of instructions that exceed the physical limits of vehicle actuators or violate regulatory requirements by large models is fundamentally prevented, thus constraining the output of nondeterministic generative AI within a deterministic safety envelope.
[0107] IV. Constructing a trustworthy loading and integrity assurance system for skills. By performing digital signature verification and triple dependency checks on tools, hardware, and data before activation through the capsule loading submodule, it is ensured that only skill capsules with legitimate sources, no tampering, version compatibility, and that meet the current system resource conditions can enter the running state, thus eliminating operational risks caused by defects in the skills themselves or supply chain attacks from the source.
[0108] V. Forming an audit-friendly, end-to-end closed-loop evidence chain. Through a runtime guard-enforced end-to-end traceability mechanism, skill matching conclusions, data filtering lists, tool call sequences, and output verification results are bound to the end-to-end TraceID and uniformly pushed to the audit plane for hash chain storage, providing complete and tamper-proof factual evidence for fault backtracking, security incident tracing, and compliance auditing.
[0109] In one embodiment, in step S3, the processed secure task request is routed to the corresponding channel for inference to obtain the inference result.
[0110] Specifically, please refer to Figure 3 The processor 901 is equipped with an inference plane module 30, which is used to divert the processed secure task requests to the corresponding channels for inference, obtain the inference results, and output them to the secure arbitration plane module 40.
[0111] It is understandable that the process of diverting the processed secure task request to the corresponding channel for inference and obtaining the inference result in step S3 is also the specific processing process of the inference plane module 30.
[0112] Please refer to the following: Figure 6 The inference plane module 30 includes a hard real-time channel and an intelligent channel. The hard real-time channel consists of a rule engine and a distillation mini-model, with a deterministic latency of ≤10ms. The intelligent channel consists of an edge-side quantization mini-model, with a deterministic latency of ≤3s, and is interruptible and degrades upon timeout.
[0113] Smaller models are suitable for rapid, reactive decision-making, while larger models are suitable for slower but more informative analysis.
[0114] The inference plane module 30 includes a real-time classifier. The real-time classifier routes the processed and safe task requests to the hard real-time channel or the intelligent channel for processing. High-priority tasks can be preempted, and tasks automatically degrade upon timeout.
[0115] Specifically, the real-time classifier uses the following traffic splitting criteria to split traffic based on the technical characteristics of the hard real-time channel (≤10ms deterministic latency) and the intelligent channel (≤3s deterministic latency, interruptible, timeout-degradable): Tasks that can be completed using the rule engine or distillation mini-models are prioritized for use in the hard real-time channel; tasks involving driving safety are mandatory to use the hard real-time channel and cannot be preempted. Specifically: task requests involving driving safety and chassis control, and / or requiring millisecond-level response and deterministic latency, are routed to the hard real-time channel; task requests involving complex semantic understanding, inference planning, multi-turn dialogue, and / or allowing interruption or timeout degradation are routed to the intelligent channel; and a fallback rule is established for automatically degrading to the hard real-time channel when a task in the intelligent channel times out, i.e., when a task in the intelligent channel times out (e.g., >3s), it automatically switches to the hard real-time channel's distillation mini-model fallback solution or the rule engine for rapid response; the response quality after degradation may decrease (e.g., changing from the optimal path to a safe and feasible path), but the functionality is guaranteed to be available.
[0116] Specifically, the hard real-time channel consists of a rule engine and a distillation mini-model. Applicable conditions include: ≤10ms deterministic latency, high priority preemption, and automatic degradation upon timeout.
[0117] For example, for vehicle safety and control (highest priority), such as a task request for AEB emergency braking, specifically collision risk assessment and braking triggering, this task is safety-critical, with an end-to-end latency of ≤10ms, and should be assigned to the hard real-time channel. Another example is a task request for high-voltage safety, specifically BMS relay on / off and insulation monitoring; this task is also high-voltage safety, requiring hardware-level real-time response, and should be assigned to the hard real-time channel.
[0118] For perception and warning tasks, such as a forward collision warning (FCW) task that is based on rules to determine the distance to the vehicle in front, the task requires extremely high timeliness of the warning. The rule engine can quickly determine this, so it should be assigned to the hard real-time channel.
[0119] For vehicle status monitoring tasks, such as motor / engine operating conditions, specifically monitoring abnormal speed, temperature, and torque, this is a deterministic monitoring task and should be assigned to the hard real-time channel.
[0120] For basic execution instructions, such as task requests for receiving and verifying remote control instructions, specifically for real-time verification and execution scheduling of instructions like remote unlocking and remote air conditioning, the task requires the instruction receiving end to respond in milliseconds and to schedule execution immediately after verification. This should be allocated to a hard real-time channel.
[0121] The intelligent channel is composed of a large-scale quantization model on the edge. Applicable conditions include: ≤3s deterministic latency, interruptibility, timeout degradation, and handling of complex reasoning and semantic understanding tasks.
[0122] For example, for high-level autonomous driving decision-making tasks, such as end-to-end driving decisions, specifically multimodal fusion of VLA models and driving strategy generation, this task requires large-scale model inference. The latency of the VLA backbone network is incompatible with real-time control, and this task should be assigned to the intelligent lane. Another example is a complex scenario path planning task, specifically decisions regarding unprotected left turns, roundabout passages, and mixed pedestrian / vehicle intersections. This task requires comprehensive inference from a large model and allows for sub-second latency; this task should also be assigned to the intelligent lane.
[0123] For cockpit interaction and user service tasks, such as multimodal interactions like gesture recognition, facial expression understanding, and cabin occupant status analysis, which involve complex multimodal fusion and require large-scale edge-side model processing, these tasks should be assigned to the intelligent channel. Conversely, for personalized recommendation tasks like music recommendations and route preference learning, which are low-frequency, high-value tasks requiring no real-time response, these should also be assigned to the intelligent channel.
[0124] For analysis and planning tasks (low priority, interruptible), such as a driving style analysis task (specifically, user driving behavior modeling and habit learning), this task is an offline / background task with the lowest priority and should be assigned to the intelligent channel. Another example is an OTA upgrade planning task (specifically, manual upgrade package download strategy and installation timing decision), which is non-real-time and can be processed entirely asynchronously; this task should also be assigned to the intelligent channel.
[0125] For complex inference tasks requiring anonymization, such as a task requesting anonymized group trajectory analysis, specifically traffic flow analysis using L2 trajectories with differential privacy noise enhancement, this type of task, being anonymized and not a real-time statistical analysis, should be assigned to the intelligent channel. Another example is a task requesting anonymized cabin behavior analysis, specifically user preference learning (not involving original biometrics). This task's data does not leave the device; it involves processing on a large edge model and should also be assigned to the intelligent channel.
[0126] In this embodiment, the inference plane module 30 employs a dual-channel architecture—a hard real-time channel (rule engine + distilled small model) and an intelligent channel (edge-side quantization large model)—combined with high-priority preemptive scheduling and automatic timeout degradation mechanisms. This achieves a balance between deterministic millisecond-level response for safety-critical tasks and flexible inference for non-safety complex tasks in vehicle scenarios, while maintaining optimal efficiency and accuracy with limited computing power. Furthermore, it collaborates with the audit plane to form a traceable closed loop for inference decision-making. Specifically: I. Achieving deterministic hierarchical assurance of task response latency. By diverting task requests to a hard real-time channel (≤10ms deterministic latency) and an intelligent channel (≤3s interruptible latency) based on safety criticality and inference complexity, millisecond-level deterministic response assurance is provided for safety-critical tasks such as AEB emergency braking and ESC vehicle stability control. At the same time, ample inference window is provided for non-safety-critical complex semantic understanding tasks, balancing real-time performance and generalization capability within a single inference plane.
[0127] II. Construct a preemptible scheduling and resource isolation mechanism for high-priority tasks. Through a hard real-time channel, support preemptive scheduling of low-priority tasks by high-priority tasks, ensuring that emergency control commands (such as collision warnings and high-voltage safety) can obtain computing resources instantly in emergency scenarios, unaffected by the blocking of other low-priority inference tasks within the channel, thereby guaranteeing the timeliness of critical control links is not disrupted.
[0128] 3. Implement timeout tolerance and security degradation for intelligent channel inference tasks. Through the intelligent channel timeout automatic degradation mechanism (switching to the distillation small model or rule engine fallback solution of the hard real-time channel when >3s), it is ensured that the system can still return a safe response with basic availability when the large model has insufficient computing power, inference is blocked, or abnormally freezes, thus avoiding vehicle function suspension or timeout loss of control due to the uncertainty delay of the large model.
[0129] IV. Achieving the optimal balance between inference efficiency and accuracy under limited on-board computing power. By deploying high-frequency rule-based tasks and lightweight distillation models on the hard real-time channel, and deploying complex multimodal inference tasks on the edge quantization model, the computational load is offloaded under the condition of limited on-board computing power resources, avoiding the dual dilemma of excessive latency and insufficient accuracy faced by a single model architecture in extreme scenarios.
[0130] V. Form an observable closed loop in collaboration with the audit plane. By recording the channel identifier (hard real-time / intelligent), actual latency, model version, and degradation trigger event during task execution in each channel of the inference plane, and binding the above inference metadata to TraceID, the audit plane can achieve traceability of channel ownership for each inference decision, reproducibility of timeout degradation, and quantifiable analysis of performance bottlenecks.
[0131] In one embodiment, in step S4, the inference result is verified, and the inference result that passes the verification is output to the inference output.
[0132] Specifically, please refer to the following: Figure 3 The processor 901 is equipped with a security arbitration plane module 40, which is used to verify the inference results and output the verified inference results to the inference output.
[0133] It is understandable that the process of verifying the inference result in step S4 and outputting the verified inference result to the inference output is also the specific processing process of the security arbitration plane module 40.
[0134] Please see Figure 6 The safety arbitration plane module 40 includes a safety arbitration access control system. This system performs contract, evidence, and confidence level verification and control isolation on the inference results. After the inference results pass the verification by the safety arbitration access control system, they are output to the cockpit display / voice and / or diagnostic reports, cloud-based anonymized summaries, or controlled execution suggestions. Specifically, after the inference results pass the verification by the safety arbitration access control system, if the inference results involve suggestion-type content, they are output to the cockpit display / voice and / or diagnostic reports; if the inference results involve control-type content, only suggestions are generated, and the original vehicle control system makes decisions and executes based on the generated suggestions. The large model never directly drives vehicle control.
[0135] Specifically, the security arbitration access control system performs contract, evidence, and confidence level verification and control isolation on the reasoning results.
[0136] The security arbitration access control is deployed between the inference plane module 30 and the execution gateway. All inference results from the hard real-time channel or intelligent channel must pass through four checkpoints of the security arbitration access control (contract, evidence, confidence verification, and control isolation) before being transformed into vehicle control commands or user-visible feedback. The security arbitration access control employs a veto mechanism; if any checkpoint fails, the inference result is rejected or a fallback solution is triggered. Simultaneously, the arbitration conclusion (including the reason for rejection) is sent back to the audit plane module 50 for evidence storage. Specifically, this includes: S41, First checkpoint: Contract verification.
[0137] The purpose of contract verification is to ensure that the reasoning result is within the scope of the authorized contract. Contract verification involves obtaining the operational semantics of the reasoning result, matching it with the currently effective user permission contract, scenario constraint contract, and legal mandatory contract, and evaluating it in real time. If the operation is not authorized or any precondition is not met, it is judged as a contract violation and the reasoning result is rejected.
[0138] The specific process includes: S411, Contractual Standard Loading.
[0139] When the security arbitration access control is activated, it loads the currently active set of dynamic contracts from the security storage area. This set of contracts includes: User permission agreement: The level of functional permissions that the current driving account has (such as whether remote vehicle control is allowed, and whether driving mode can be modified); Scenario-constrained contracts: Action whitelist defined based on conditions such as geographic location fences, time windows, and vehicle status (whether in P gear); Legally binding contracts: The enforcement boundaries that are forcibly defined by local traffic regulations (such as maximum speed limits, prohibition of changing lanes over solid lines, etc.).
[0140] S412, Matching the reasoning result with the contract.
[0141] The security arbitration access control parsing and reasoning results are used to extract key operational semantics (e.g., changing lanes to the left, activating sentry mode, raising the suspension).
[0142] Using the semantics of the operation as the primary key, the corresponding authorization policy is retrieved from the contract set. If the operation is not explicitly defined in any contract, it is determined to be an "illegal operation" and is directly rejected.
[0143] S413, Real-time evaluation of conditional expressions.
[0144] The contract includes dynamic conditions (e.g., only available when the vehicle speed is below 30 km / h). The security arbitration access control system has a built-in lightweight rule engine that extracts the current vehicle CAN signal and environmental context in real time to perform Boolean evaluation on the condition expressions in the contract.
[0145] Contract validation passes only if all prerequisites are met; otherwise, it is marked as a "scenario violation" and the specific unmet conditions are recorded for use in subsequent degradation strategies.
[0146] Specifically, the preconditions originate from the scenario constraints and legal obligations of the security arbitration access control system, and mainly include the following three types of real-time conditions: 1. Vehicle physical status conditions: refers to the current real-time parameters of the vehicle chassis and powertrain. For example: "Vehicle speed must be below 30 km / h" (for automatic parking or lane change assist), "Transmission gear must be in P or N" (for activating sentry mode or viewing panoramic image), "Vehicle must be in park" (for playing video on the central control screen).
[0147] 2. Spatial and Environmental Fence Conditions: This refers to whether the vehicle's current geographical location and external environment allow for the operation. For example: "The vehicle's current location must be within a supported area covered by a high-precision map" (for City Navigation Assisted Driving (NOA)), "The vehicle must not be in no-honking zones near schools or hospitals" (for controlling horn activation), "The current ambient light intensity must be higher than 1 lux" (for functions that rely on pure vision cameras).
[0148] 3. System Resources and Time Window Conditions: This refers to whether the current system load or time period meets the execution requirements. For example: "Current system CPU / GPU load is below 80%" (ensuring sufficient computing power during execution), "Current time is outside the user-defined do-not-disturb period" (used for voice broadcasts or cockpit reminders), "The minimum safe interval has been exceeded since the last execution of this action" (preventing high-frequency repeated operations from causing hardware damage).
[0149] In its implementation, the lightweight rule engine built into the security arbitration access control system reads the vehicle's CAN bus signals, GPS positioning data, and system monitor readings in real time, performing Boolean logic evaluation (True / False) on these preconditions. Only when all preconditions return "True" will the contract verification allow passage; if any condition is not met (e.g., exceeding the speed limit), the access control system immediately determines it as a "scenario violation" and rejects the reasoning result.
[0150] S42, The second checkpoint: Evidence verification.
[0151] Evidence verification involves extracting the stored digest of the input data upon which the inference result depends. This compares the real-time hash value of the current vehicle data source with the stored hash value in the audit chain. Simultaneously, it verifies the discrepancy between the data collection timestamp and the current system time. If the integrity is compromised or the timeliness exceeds the limit, the evidence is deemed invalid, and the inference result is rejected. Evidence verification ensures that the original basis for generating the inference result is authentic, undisturbed, and not expired, preventing attackers from deceiving the large inference model by injecting forged sensor data or replaying expired data.
[0152] Specifically, it includes: S421, Extraction of the chain of reasoning evidence.
[0153] The reasoning result includes an evidence digest pointer, which points to the location of the key input data it depends on (such as camera image hashes, radar point cloud summaries, and historical trajectory sequence numbers) in the audit hash chain.
[0154] The security arbitration access control retrieves this pointer and queries the audit plane module 50 for the hash value and collection timestamp of these input data.
[0155] S422, Static Integrity Verification.
[0156] The security arbitration access control system performs real-time hash calculations on the currently used vehicle data source and compares it with the hash value stored in the audit plane module 50. If they match, it means the input data has not been tampered with between collection and inference, and the evidence verification passes. If they do not match, it means the input data has been tampered with between collection and inference, and the evidence verification fails.
[0157] S423, Dynamic Freshness Verification.
[0158] The security arbitration access control system checks the deviation between the timestamp of the input data and the current system time. If the deviation exceeds a preset threshold (e.g., more than 500 milliseconds for visual images), the data is determined to be "outdated" and unsuitable for the current real-time environment.
[0159] For L3 level control commands, the security arbitration access control system mandates that the sensing data it relies on must come from the current sensing cycle, and any data from historical cache is directly rejected.
[0160] S43, The third hurdle: Confidence check.
[0161] Confidence verification assesses the reliability of large-scale model inference results on the edge of the intelligent channel, preventing security incidents caused by the large model in unfamiliar scenarios. Confidence verification analyzes the confidence score and uncertainty index attached to the model output and makes decisions according to graded thresholds. If the confidence score is below the lower limit, it is directly rejected and a safety fallback is triggered. If it is in the middle range, additional constraints such as slowing down and limiting the execution intensity are applied before downgrading and allowing the process to proceed.
[0162] Specifically, it includes: S431, Confidence score analysis.
[0163] The security arbitration access control extracts the confidence score (probability value) and uncertainty index (such as entropy value or MC-Dropout variance) of the large model output from the inference results.
[0164] For the inference results output by the rule engine / distillation small model of the hard real-time channel, the confidence level is preset to the maximum value by default due to its high logical determinism, and the confidence level verification is quickly approved.
[0165] S432, Graded threshold decision.
[0166] Security arbitration access control enforcement three-layer confidence level judgment logic: High confidence interval (e.g., ≥90%): The result is reliable, and the process is directly allowed to proceed to the next checkpoint; For inference results output by hard real-time channels, a direct response is given, and security-related issues are covered by security rules and fallback strategies. Medium confidence interval (e.g., 70%~90%): The result has a certain degree of uncertainty. The security arbitration access control does not directly reject it, but downgrades the instruction issuance mode to semi-automatic execution, that is, adds strong constraint parameters (such as limiting the execution speed or execution range), or requires the driver to manually confirm before execution. Low confidence interval (e.g., <70%): The result is unreliable, the inference result is rejected directly, and the safety rule fallback strategy is automatically triggered (e.g., the backup rule of the hard real-time channel or the conservative back-off action of the vehicle).
[0167] S44, Fourth checkpoint: Control and isolation.
[0168] Control isolation is the last hard barrier for safety arbitration, specifically targeting L3 level control commands (involving steering, braking, torque, and high voltage switching). Its core purpose is to safely isolate logical decisions from physical execution, ensuring that even if the previous barriers are breached, the vehicle's underlying system still has the final safety veto power.
[0169] Control isolation involves comparing the critical control commands to be issued with the physical safety boundary table item by item, simultaneously verifying them through dual-channel heterogeneous redundancy of the main control SOC and independent MCU coprocessor, and resolving conflicts with the driver's real-time operation signals before release. If any step fails, the issuance path of the command is physically cut off.
[0170] Specifically, it includes: S441, Instruction security envelope check.
[0171] The security arbitration access control system maintains a physical safety boundary table for the chassis hardware, which includes the maximum rate of change, maximum absolute value, and safety locking conditions for each actuator (e.g., the steering wheel angle must not exceed ±450°, and the torque change per second must not exceed 500 N·m).
[0172] The security arbitration access control compares each L3 control value to be sent with the security boundary. If any parameter exceeds the boundary, the security arbitration access control immediately intercepts the command, preventing it from entering the bus.
[0173] S442, Dual-channel independent verification.
[0174] For critical control commands, the security arbitration access control system synchronously sends the command to two independent logic verification units (one located in the main control SOC, and the other in a separate MCU security coprocessor). The two units perform validity verification according to their respective independent algorithms.
[0175] The security arbitration access control will only release the instruction if the verification results of the two channels are consistent and pass; otherwise, it will be forcibly discarded to ensure that the hardware or software failure of a single chip will not lead to the execution of dangerous instructions.
[0176] S443, Driver priority hard overlay.
[0177] The security arbitration access control continuously monitors the physical signals from the steering wheel torque sensor, brake pedal, and accelerator pedal.
[0178] At the very last moment before releasing the L3 control command, the safety arbitration access control system executes a conflict resolution: if it detects that the driver is actively applying force (gripping the steering wheel or pressing the brake pedal), the AI control command is automatically suspended or downgraded, and the vehicle response takes priority over the human driver's operation. This resolution is executed at the signal level, without relying on the scheduling of the upper-level operating system, ensuring hard real-time determinism.
[0179] The security arbitration access control system makes a comprehensive arbitration decision and outputs the results based on the outputs of the above four checkpoints, specifically including: 1. Comprehensive scoring across four stages and action decision-making.
[0180] Each of the four checkpoints outputs a "pass / fail / downgrade suggestion". The security arbitration access control engine makes the final decision according to priority: Hard Rejection: If the contract fails to pass verification, the control isolation limit is exceeded, or the integrity of the evidence is compromised, execution is refused, triggering an alarm. Degraded execution: If the confidence level is moderate or the freshness of the evidence is slightly off, reduce the response speed or amplitude of the command and increase human-machine co-driving prompts; Full clearance: After all four checkpoints are passed, the authorization timestamp and arbitration signature of the security arbitration access control are added, and the final secure and reliable instruction is issued to the actuator.
[0181] 2. The arbitration result is stored on the blockchain as evidence.
[0182] Regardless of the arbitration outcome (rejection / downgrade / approval), the security arbitration access control system generates an arbitration snapshot containing independent judgments from four levels, details of the violation, and the final decision.
[0183] The snapshot is sent to the audit plane module 50 through the internal interface, bound to TraceID, and written into the hash chain ledger as the audit payload of the contract verification phase, realizing full-link traceability.
[0184] 3. Downgrade and timeout escape mechanisms.
[0185] If the security arbitration access control system times out at a certain checkpoint (especially evidence verification or confidence verification) due to heavy computing resources, the system will not block indefinitely, but will trigger an escape mechanism: for non-security-critical operations (such as entertainment information recommendations), a friendly rejection will be returned directly; for security-critical operations, the system will immediately switch to the least risk strategy (such as slowing down and pulling over) to ensure that the vehicle's functional safety level response is not affected by the arbitration delay.
[0186] In this embodiment, the security arbitration plane module 40 includes contract verification, evidence verification, confidence level verification, and control isolation. It implements a comprehensive adjudication of the inference results, ensuring compliance with permissions, data reliability, output reliability, and physical security. This guarantees that only inference results that simultaneously satisfy all security constraints can be issued for execution, thus constructing multiple redundant security isolation barriers between nondeterministic generative AI and deterministic vehicle control. Specifically: I. Achieving comprehensive multi-dimensional compliance adjudication before inference results are issued. By integrating four checkpoints—contract verification, evidence verification, confidence verification, and control isolation—in a serial pipeline between the inference plane and the execution gateway, a comprehensive security barrier is formed that cannot be bypassed by any single dimension. This ensures that only inference results that simultaneously meet all four types of security constraints can be authorized for execution.
[0187] Second, ensure the consistency between generative inference results and preset permissions and scenario constraints. User permission contracts, scenario constraint contracts, and mandatory legal contracts are loaded in real-time through contract verification checkpoints. The operational semantics of the inference results are matched and evaluated in real-time with dynamic conditional expressions. From the perspectives of functional authorization and spatiotemporal boundaries, this ensures that the output of the nondeterministic large model always operates within the predefined legal operation range, preventing the issuance of illegal instructions due to model illusion or unauthorized inference.
[0188] Third, ensure the authenticity and timeliness of the input data upon which the reasoning is based. An evidence verification checkpoint compares the current real-time hash of the sensors with the audit chain's stored hash to verify integrity. Simultaneously, a threshold is applied to determine the deviation between the data collection timestamp and the current system time to verify freshness. This ensures that the perceived data upon which the reasoning conclusions rely has not been tampered with and reflects the current real-time environment, thus preventing outdated or falsified data-driven erroneous reasoning results from entering the execution phase.
[0189] IV. Implementing graded handling and safety fallback for model output credibility. A rule engine fallback scheme uses a confidence verification checkpoint to classify and adjudicate the confidence score and uncertainty index attached to the inference results. Under high confidence, the process is allowed to proceed normally; under medium confidence, strong constraints are applied and execution is downgraded; and under low confidence, the process is directly rejected and a hard real-time channel is triggered. This transforms the inherently nondeterministic output of large models into a predictable and graded deterministic safety response.
[0190] 5. Construct a physical safety hard boundary for chassis actuators. By controlling isolation gates to perform safety envelope checks on critical control commands (compare each item with the physical limit boundary), dual-channel heterogeneous redundancy verification (parallel verification by the main control SOC and independent MCU), and driver priority hard overlay (real-time signal conflict adjudication), an independent physical execution layer safety barrier is constructed after the logic decision layer to ensure that even if the preceding verification mechanism is breached, the vehicle chassis actuators will not receive dangerous commands that exceed the physical safety boundary or conflict with the driver's intentions.
[0191] In one embodiment, in step S5, the entire chain formed by task request, skill capsule matching, reasoning, and verification is recorded in the hash chain ledger and periodically signed and anchored.
[0192] Specifically, please refer to Figure 3 The processor 901 is equipped with an audit plane module 50, which is used to record the entire chain of task requests, skill capsule matching, inference, and verification into the hash chain ledger and perform periodic signature anchoring.
[0193] It is understandable that the process of recording the entire chain of task requests, skill capsule matching, reasoning, and verification into the hash chain ledger in step S5, and performing periodic signature anchoring, is also the specific processing process of the audit plane module 50.
[0194] Specifically, the audit plane module 50 uses TraceID for full-chain tagging, and employs vehicle-side hash chain ledger storage and periodic TEE signature anchoring to achieve traceability, verifiability, and tamper-proofing. The specific process includes: TraceID full-chain tagging, hash chain ledger, and TEE signature anchoring. The specific implementation process is as follows: S51, TraceID full-chain tagging: full-chain TraceID injection and context passing process.
[0195] Upon receiving an external task request, a globally unique TraceID is generated at the task entry point and persists throughout the entire process of "request → matching → inference → verification," enabling audit records at each stage to achieve logical association and call topology reconstruction through the same identifier. The specific implementation process includes: S511, TraceID generation and injection.
[0196] When the system receives an external task request, the audit plane module 50 immediately generates a globally unique TraceID (containing a timestamp, node identifier, and random number) at the ingress gateway. This TraceID is injected into the task request context and flows with the task request to the subsequent three stages: skill capsule matching, inference execution, and security arbitration.
[0197] S512, end-to-end marking.
[0198] At each processing node in each stage (skill capsule matching, inference execution, and security arbitration), the audit plane module 50 extracts the TraceID from the context. Each stage, when generating its own audit sub-record, forcibly binds the TraceID to ensure that the logs of each processing node logically belong to the same task request transaction.
[0199] S52, Hash Chain Ledger: The process of building and writing the vehicle-side hash chain ledger.
[0200] After obtaining the full-link TraceID binding, the audit plane module 50 writes the structured audit records into the tamper-proof vehicle-side hash chain ledger. Specifically, the audit plane module 50 embeds the hash value of the previous record as a chain pointer when writing each audit record, and the hash value and the current record payload participate in the hash operation to generate their own hash fingerprint. This is then written sequentially to the vehicle-side hash chain ledger in an append-only manner, forming a chain data structure where records are strongly coupled through hash pointers. This ensures that any single-point tampering with any historical record will be detected by the chain verification of subsequent records due to hash pointer breakage. The specific implementation process includes: S521, Structured assembly of audit records.
[0201] The audit plane module 50 assembles the task request summary, skill capsule matching result, inference execution summary, contract verification conclusion and TraceID together into the current audit payload; at the same time, it adds the current vehicle-side timestamp and operator identity identifier (such as user account de-identified ID or system process ID).
[0202] S522, Hash chain pointer generation.
[0203] The audit plane module 50 reads the hash value (i.e., chainhead hash) of the latest record in the local ledger. After concatenating the previous chainhead hash, the current audit payload, TraceID, and timestamp, it performs an SM3 or SHA-256 hash operation to generate a unique hash fingerprint for the current record.
[0204] S523, Ledger addition.
[0205] The audit plane module 50 encapsulates the preceding hash, current payload digest, current hash fingerprint, and timestamp into a complete block record and writes it to the vehicle-side secure storage area in append-only mode. Since each new record is nested with the hash of the previous record, once the historical record is tampered with, the hash fingerprints of all subsequent records become invalid, thus forming a hash chain locally on the vehicle that cannot be tampered with by a single point of attack from either internal or external attackers.
[0206] S524, Storage Write Strategy.
[0207] To ensure low latency in the hard real-time channel, the audit plane module 50 adopts an asynchronous batch write strategy: audit records are temporarily stored in a high-reliability buffer, and then asynchronously retrieved and batch-applied to the hash chain by an independent background thread, ensuring that the millisecond-level response of the main control flow is not affected.
[0208] S53, TEE Signature Anchoring: The execution process of periodic TEE signature anchoring.
[0209] Relying solely on the vehicle-side local hash chain is insufficient to defend against attacks that completely replace or rollback physical storage. Therefore, a periodic anchoring based on a trusted execution environment is introduced. Specifically, triggered by time or quantity, newly added records within a period are aggregated into a Merkle root hash. This hash is then signed by an automotive-grade TEE using a hardware-secured key and published to the cloud / blockchain. This achieves periodic TEE signature anchoring and provides off-chain immutable and verifiable timestamp proof. The specific implementation process is as follows: S531, Anchor Trigger and Window Establishment.
[0210] The audit plane module 50 establishes a dual trigger mechanism: a time dimension (e.g., every 5 minutes) and a quantity dimension (e.g., every 100 new records accumulated). An anchoring process is triggered once either condition is met. Upon anchoring initiation, the audit plane module 50 records the start and end sequence numbers of the current window and the hash of the starting chain head.
[0211] S532, Merkle tree aggregation (efficiency optimization).
[0212] To improve anchoring efficiency, the audit plane module 50 does not sign each record individually. Instead, it extracts the hash values of all newly added audit records in the current period, constructs a Merkle tree inside the TEE, and calculates the unique root hash of all records in the current period.
[0213] S533, TEE internal signature operation.
[0214] The audit plane module 50 transmits the "Merkle root hash + latest window serial number + latest chainhead hash + anchor timestamp" to the automotive-grade TEE security environment (e.g., TrustZone or a standalone security chip). The TEE uses its internally embedded unique device key or a key derived from the vehicle manufacturer's root certificate to digitally sign this combined data. The entire signing process is performed within the closed TEE environment, and the key is never exposed to the regular operating system or application layer. Even if the vehicle's infotainment system is compromised, attackers cannot forge the anchor signature.
[0215] S534, Outward dissemination and consolidation of anchored evidence.
[0216] After the TEE signature is completed, the audit plane module 50 uploads the signature value and anchor data packet (referred to as anchor evidence) to the vehicle manufacturer's cloud audit platform for storage via a secure channel for cloud-side backup. After successful anchoring, the audit plane module fills the anchor evidence ID and signature value back into the next-hop record of the vehicle-side hash chain, forming a two-way binding between on-chain and off-chain anchoring.
[0217] In this application, the audit plane module 50 uses TraceID to tag the entire link and uses vehicle-side hash chain ledger for evidence storage and periodic TEE signature anchoring to achieve traceability, verifiability, and tamper-proof.
[0218] 1. Tracking process.
[0219] When security auditors or compliance systems need to verify a specific task, they can retrieve it based on the TraceID by simply entering the TraceID. The query engine of the audit plane module uses an inverted index to quickly retrieve all associated audit records across four stages for that TraceID, and reconstructs the complete execution trajectory of "request → matching → inference → verification" in chronological order.
[0220] 2. Traceability process.
[0221] For the retrieved records, the audit plane module 50 provides a chain tracing function: starting from the target record, it backtracks along the preceding hash pointer to verify whether the hash connection of each level is intact, and confirms that the context of the task during execution has not been tampered with.
[0222] 3. Tamper-proof verification process, achieving dual verification.
[0223] Local chain verification: The audit plane module 50 starts from the genesis block, recalculates the hash of each block one by one, compares the stored hash fingerprints, and verifies the overall integrity of the local hash chain.
[0224] TEE Anchor Verification: Anchor evidence for the corresponding period is retrieved from the cloud or blockchain, and the signature validity is verified using the TEE public key. Simultaneously, the Merkle tree of the vehicle's local records for the current period is reconstructed, and the root of the tree is compared to the Merkle root in the anchor evidence. If both match, it mathematically proves that no new, deleted, or modified records have been added to the vehicle's local records since the last anchoring.
[0225] In this embodiment, the audit plane module 50 achieves traceable task execution trajectory, tamper-proof audit data, and externally verifiable solidification of ledger status through a triple mechanism: TraceID end-to-end tagging, vehicle-side hash chain-based tamper-proof evidence storage, and periodic TEE signature external anchoring. It also possesses adaptive degradation capabilities in privacy compliance and resource-constrained scenarios. Specifically: I. Achieve cross-stage logical association and trajectory reconstruction throughout the entire lifecycle of a task request. By generating a globally unique TraceID at the entry point of each task and passing it through all stages of the entire chain from "task request → skill capsule matching → inference execution → contract verification", audit records that were originally scattered across various independent subsystems are logically linked through the same identifier, supporting timeline reconstruction and call topology backtracking of the complete execution trajectory for each individual task.
[0226] II. Constructing a chain-like tamper-proof storage structure for vehicle-side audit data. By embedding the hash value of the preceding record as a chain pointer during the writing process for each audit record, and having it participate in the hash operation together with the current record payload to generate its own hash fingerprint, the data is sequentially written to the vehicle-side secure storage medium in an append-only manner. This forms a chain-like data structure where preceding and following records are strongly coupled through the hash pointers, ensuring that any single point of tampering with any historical record will be detected by the chain verification of subsequent records due to the breakage of the hash pointer.
[0227] Third, achieve external solidification and verifiable time anchoring of the audit ledger status. By triggering periodic anchoring based on time or quantity thresholds, newly added records within the period are aggregated into a Merkle root hash. After being signed by an automotive-grade TEE using a hardware-secured key, the hash is published to external immutable media such as cloud storage and / or blockchain, forming an externally verifiable snapshot of the vehicle-side hash chain status. This effectively resists attacks that could completely replace or roll back the vehicle-side physical storage and provides legally valid timestamp proof for audit data.
[0228] IV. Achieving Compliance Synergy Between Auditing Practices and Privacy Data Classification Strategies. By strictly implementing data classification filtering during the assembly of the audit payload—only recording de-identified identifiers or hash digests for L2-level privacy data (precise trajectory, biometrics) and L3-level control instructions, while ensuring that original sensitive data does not enter the audit ledger—it is ensured that the audit plane's own evidence storage activities do not constitute secondary collection and leakage of user privacy, thus meeting the regulatory compliance requirements for data transmission outside the vehicle.
[0229] V. Achieve adaptive resource degradation and system real-time performance assurance for the audit plane. By employing asynchronous batch writing and independent background thread processing for audit write operations, decoupling from the execution path of the hard real-time channel (≤10ms), it is ensured that audit evidence storage does not increase the response latency of the main control link. At the same time, when storage resources are close to saturation, cold and hot separation archiving is triggered. When TEE anchoring fails due to network timeout, it is marked as "to be re-anchored" and automatically re-signed after recovery. This enables the audit function to still have elastic self-healing capabilities under conditions of limited vehicle-side resources and unstable network.
[0230] It is understood that the data plane module 10, skill constraint plane module 20, reasoning plane module 30, security arbitration plane module 40 and audit plane module 50 can be implemented as hardware modules or as software modules, and the present invention does not impose any restrictions.
[0231] When implemented using hardware modules, the data plane module 10, skill constraint plane module 20, inference plane module 30, security arbitration plane module 40, and / or audit plane module 50 can be integrated into the processor 901 as one of the hardware modules of the processor 901; when implemented using software modules, the data plane module 10, skill constraint plane module 20, inference plane module 30, security arbitration plane module 40, and / or audit plane module 50 can be a functional module in the program executed by the processor 901.
[0232] Currently, a special but frequent type of fault scenario exists in vehicles: intermittent faults. These faults manifest as occasional alarms, transient communication interruptions, or brief signal anomalies. They typically disappear after a vehicle restart, power-off reset, or system self-healing. When a fault occurs, the on-site status is cleared and the rolling log is overwritten, making it impossible to reproduce the anomaly during after-sales retesting. Due to the lack of effective on-site freezing and evidence preservation mechanisms, these problems have long been trapped in a "no on-site, no evidence, no reproducibility" dilemma. Troubleshooting methods heavily rely on engineers' experience-based speculation and hardware replacement testing, resulting in low after-sales troubleshooting efficiency and high repair costs. Existing vehicle diagnostic solutions lack a systematic technical solution for this scenario.
[0233] For this purpose, please refer to Figure 7 This vehicle-mounted intelligent device skill constraint method based on a large model also includes: S6, when the fault sentinel detects an anomaly, it first triggers a site freeze; after the recovery action is executed, the fault analysis skill outputs a diagnostic report, even if the fault is no longer reproducible. Specifically, it includes: S61. When the fault sentinel detects an abnormal event (DTC / log mode / metric over-limit), it does not perform any restart or automatic recovery actions. Instead, it immediately freezes the scene and starts the diagnostic process to prevent the fault scene from being destroyed due to a hasty restart.
[0234] S62. At the moment the freeze is triggered, the on-site freezer automatically extracts the following key data: 1) Snapshot Freeze: Automatically extracts critical operational data from the N seconds prior to a failure, including: Logs, Signals, On-Board Diagnostic Trouble Codes (OTC), and process status. This data is forcibly written to an isolated protected area, which is separate from the regular logs. Even if the system is subsequently reset, this data will not be overwritten.
[0235] 2) Environmental fingerprint: Simultaneously captures the external environment and internal state at the moment the fault occurred, including but not limited to: Time: A timestamp accurate to milliseconds; Temperature / Voltage: The operating temperature and supply voltage of the chip or module; Network status: signal strength, packet loss rate, latency; Load / Software Version: CPU / memory usage, currently running firmware or container version.
[0236] This data is used to rule out occasional failures caused by environmental factors such as overheating, undervoltage, and network jitter.
[0237] 3) Context association: Spatiotemporally associate the data N seconds before the failure with the current environmental fingerprint snapshot to construct a complete fact map at the moment of the failure, such as whether a sudden change in a signal value occurred at the same time as a network disconnection or a temperature spike.
[0238] S63. After freezing is complete, recovery actions, including restarting, are allowed.
[0239] S64. After recovery, the fault analysis skill capsule is triggered asynchronously. The fault analysis skill capsule packages all the above structured data (frozen snapshot, environmental fingerprint, associated context) and inputs them into the large model.
[0240] S65. The large model makes the following inferences based on all the structured data mentioned above: Timeline reconstruction: The large model reconstructs the complete timeline from the occurrence of the anomaly to the triggering of the alarm based on this data.
[0241] Anomaly pattern recognition and root cause hypothesis: With the assistance of a knowledge base (including historical failure cases and a known software defect database), the large model performs pattern matching and proposes one or more root cause hypotheses.
[0242] Evidence Verification: The large model verifies the hypotheses against the collected chain of evidence one by one. If the evidence is insufficient, it generates reproduction condition suggestions (such as "Please retry with XX load under XX firmware version") for subsequent testing and verification.
[0243] S66, Output and Closed Loop.
[0244] Regardless of whether the fault resolves itself on-site (such as reconnection caused by momentary network fluctuations), a complete diagnostic report will be forcibly generated.
[0245] The diagnostic report conclusions (confirmed root cause, confidence level, and recommended remediation) will be sent back to the audit plane module 50 and archived along with the TraceID of this fault.
[0246] In this embodiment, for intermittent faults that do not recur after a restart, when more detailed fault analysis is needed, a fault analysis skill capsule can be triggered, employing a "freeze first, then restore, then analyze" mechanism: After the fault sentinel detects an anomaly, it freezes the log stream, signal snapshots, DTC status, and environmental fingerprints within the time window before the restart / self-healing action to an independent protected area. After restoration, the fault analysis skill capsule (large model) performs timeline reconstruction, anomaly pattern identification, root cause hypothesis testing, and evidence verification on the snapshots, outputting root cause conclusions with evidence chains, confidence levels, and suggested reproducibility conditions. Even if the fault does not recur after a restart, the problem can still be quickly located and identified. This avoids the operational dilemma of "the cause of the fault remaining a mystery forever" after a traditional restart.
[0247] This invention also provides a computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a computer processor, cause the processor to perform the method described in the foregoing embodiments.
[0248] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.
[0249] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.
[0250] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; under the concept of the present invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the present invention as described above, which are not provided in detail for the sake of brevity; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for constraining the skills of an in-vehicle intelligent device, characterized in that, Applied to in-vehicle intelligent devices, the method includes: Collect vehicle data sources, and store them securely after classifying and de-identifying the vehicle data sources according to a preset data classification strategy. Each capability of the large model is encapsulated as a digitally signed skill capsule, and the skill capsule is activated and matched when a task request is generated to output a secure task request, wherein the task request carries a vehicle data source that has been hierarchically de-identified. The processed and secure task requests are routed to the corresponding channels for inference to obtain the inference results. The reasoning results are verified, and the reasoning results that pass the verification are output to the reasoning output. The entire chain of task requests, skill capsule matching, reasoning, and verification is recorded in the hash chain ledger and periodically signed and anchored.
2. The method according to claim 1, characterized in that, The process of collecting vehicle data sources, and then securely storing the vehicle data sources after performing hierarchical de-identification according to a preset data hierarchical strategy, includes: Collect vehicle data sources, wherein the vehicle data sources include at least one of the following: vehicle static data, bus signals, operation logs, diagnostic data, cockpit inputs, and environmental data; The vehicle data source is classified according to a preset data classification strategy, wherein the preset data classification strategy includes L0 public data, L1 in-vehicle sensitive data, L2 privacy data and L3 control data. For vehicle data sources that have already been classified, de-identify them according to their corresponding de-identification strategies; Securely store the vehicle data source after data anonymization.
3. The method according to claim 2, characterized in that, The process of de-identifying vehicle data sources that have already undergone data classification according to their corresponding de-identification strategies includes: The desensitization strategy for the L0 level public data includes: no desensitization required; The desensitization strategy for L1-level in-vehicle sensitive data includes: desensitization before transmission outside the vehicle, and desensitization using technical measures such as accuracy downgrading, feature extraction, and / or aggregation statistics; The de-identification strategy for the L2 level privacy data includes: in-vehicle processing, strict de-identification, and the use of techniques such as obfuscation, differential privacy, pseudonymization, and / or local encryption for de-identification; The desensitization strategy for the L3 level control data includes: not applying desensitization, and using technical measures such as digital signatures, end-to-end encryption, and / or physical isolation for protection.
4. The method according to claim 1, characterized in that, The process of encapsulating each capability of the large model into a digitally signed skill capsule, and activating and matching the skill capsules when a task request is generated, includes: Each capability of the large model is encapsulated into a digitally signed skill capsule and stored. Activate the skill capsule when a task request is generated; The system matches task requests with activated skill capsules. If a match fails, the task request is rejected and audited. If a match is successful, the system processes the successful task request based on five preset rules and outputs a safe task request.
5. The method according to claim 1, characterized in that, The channel includes a hard real-time channel and an intelligent channel; wherein, the hard real-time channel consists of a rule engine and a distillation small model, and the intelligent channel consists of an edge-side quantization large model; The process of routing the processed secure task requests to corresponding channels for inference and obtaining inference results includes: Tasks that can be completed through the rule engine or distilled small models should be prioritized for use in the hard real-time channel; tasks involving driving safety should be forced to use the hard real-time channel.
6. The method according to claim 1, characterized in that, The verification of the reasoning result includes: performing contract verification, evidence verification, confidence verification, and control isolation on the reasoning result.
7. The method according to claim 1, characterized in that, The process of recording the entire chain of task requests, skill capsule matching, inference, and verification into a hash chain ledger and periodically signing and anchoring it includes: When an external task request is received, a globally unique TraceID is generated at the task entry point and follows the request through all stages of request, matching, inference, and verification. Each audit record is embedded with the hash value of the previous record as a chain pointer when it is written, and it participates in the hash operation together with the current record payload to generate its own hash fingerprint, and is sequentially written to the vehicle-side hash chain ledger in an append-only manner. Triggered by time or quantity, newly added records within a period are aggregated into a Merkle tree root hash, which is then signed by an automotive-grade TEE using a hardware-secured key, achieving periodic TEE signature anchoring.
8. The method according to claim 1, characterized in that, The method also includes: when the fault sentinel detects an anomaly, it first triggers a site freeze, and after the recovery action is executed, the fault analysis skill outputs a diagnostic report.
9. A vehicle-mounted intelligent device, characterized in that, It includes a processor and a memory, the memory storing a computer program, and the processor, when executing the computer program, implements the in-vehicle intelligent device skill constraint method according to any one of claims 1 to 8.
10. A computer storage medium, characterized in that, The computer storage medium stores a computer program, which includes program instructions that, when executed by a processor, cause the processor to perform the in-vehicle intelligent device skill constraint method according to any one of claims 1 to 8.
