A method and system for edge-cloud collaborative multimodal security protection and dynamic evaluation of vehicle-mounted large language models.
By employing a multimodal security protection method that integrates edge and cloud technologies, attacks on in-vehicle large language models are assessed in real time and dynamically adapted to address security risks in the in-vehicle environment, achieving millisecond-level response and full-link closed-loop protection.
Patent Information
- Application Number
- CN202611294692.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-25
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies for the security protection of large language models in vehicles suffer from insufficient real-time performance, insufficient multimodal fusion perception, weak dynamic adversarial evaluation capabilities, and insufficient self-evolution capabilities of security governance closed loops, making it difficult to effectively address security risks in the vehicle environment.
A multimodal security protection method with edge-cloud collaboration is adopted. Risk assessment is carried out in real time by fusing visual, voice and vehicle status features on the edge device, and dynamic adversarial testing and model fine-tuning are carried out in the cloud. Combined with a self-evolving data flywheel mechanism, a closed-loop protection is achieved across the entire link.
It achieves real-time interception and defense of vehicle commands within millisecond-level response time, dynamically adapts to attack evolution, provides multi-dimensional security assessment and automated closed-loop repair, and improves the security and compliance of vehicle-mounted large language models.
Smart Images

Figure CN123122913A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of edge-cloud collaboration technology based on large language models, specifically to a method and system for edge-cloud collaborative multimodal security protection and dynamic evaluation oriented towards in-vehicle large language models. Background Technology
[0002] As the global automotive industry undergoes a profound transformation towards intelligence and connectivity, the intelligent cockpit has evolved into a digital third space integrating perception, interaction, and control. Large Language Models (LLMs), with their superior natural language understanding and multi-turn dialogue reasoning capabilities, are gradually replacing traditional rule-based voice assistants and becoming the core interaction engine of the next generation of intelligent cockpits. However, the unique characteristics of the in-vehicle environment extend the security risks of LLMs from the virtual information domain to the physical security domain. Because in-vehicle LLMs typically have the authority to execute vehicle control commands (such as opening the trunk, adjusting the power mode, and disabling driver assistance systems) by calling the underlying CAN bus via an interface, attacks targeting LLMs can directly threaten the driver's life.
[0003] Currently, attack methods targeting in-vehicle LLMs are becoming increasingly sophisticated, primarily manifesting as social engineering spoofing and jailbreak command hijacking. Attackers exploit semantic alignment vulnerabilities in LLMs through carefully crafted prompts, inducing users to execute jailbreak operations that violate security compliance standards. Furthermore, the multimodal environment of in-vehicle systems introduces attack vectors from multiple dimensions, including voice and vision. Attackers may exploit blind spots caused by driver fatigue or complex acoustic environments to launch combined attacks. While traditional rule-based firewalls are effective at filtering simple keywords, they exhibit extremely high false negative rates when facing highly semantically ambiguous LLM spoofing commands due to a lack of deep semantic understanding. Simultaneously, cloud-based security filtering solutions are limited by network transmission latency, making it difficult to meet the millisecond-level real-time response requirements of in-vehicle bus control. Therefore, constructing a comprehensive in-vehicle large-scale model security protection system that balances high real-time performance, deep semantic recognition, and end-to-end closed-loop evolution has become a critical technological bottleneck that urgently needs to be overcome in the field of automotive intelligent safety.
[0004] The existing technologies and their shortcomings in the field of intelligent vehicle safety are as follows: (1) Lightweight visual representation model and cabin monitoring technology: In the safety closed loop of intelligent cabin, the driver monitoring system (DMS) is the first line of defense for perceiving the safety status of the physical world. Traditional DMS solutions are mostly based on convolutional neural networks (CNN) or lightweight visual transformers (such as MobileViT), using their feature extraction capabilities to capture the driver's illegal behaviors such as closing eyes and looking down; however, traditional classification models have natural shortcomings in dealing with complex and open cabin interaction contexts, and are prone to temporal misjudgments (such as misjudging normal physiological blinking as fatigue-induced eye closing), and face the problem of collapse in generalization ability for newly emerging dangerous actions (zero samples); In recent years, the vision-language multimodal alignment large model (Zero-Shot) has been developed. The rise of VLM (such as CLIP architecture) has provided a dimensionality reduction solution for cockpit monitoring. It no longer relies on simple image classification, but accurately understands complex semantic-level distracted driving actions by calculating the high-dimensional feature similarity between the image space and the natural language command space. However, existing technologies mostly operate visual monitoring independently and lack spatiotemporal semantic coupling with the vehicle LLM command flow, resulting in the inability to achieve end-to-end physical collaborative blocking when the driver is deceived. (2) Cross-language pre-trained representation and command auditing technology: For the text commands received by the vehicle LLM, high-precision real-time auditing is the core of preventing "jailbreak attacks". At present, pre-trained language models (PLM) based on the Transformer architecture have become the mainstream solution in the field of NLP, especially multi-language representation models such as XLM-Roberta, which perform masking on a large-scale cross-language corpus (such as CommonCrawl). XLM-R uses a Multilingual Model (MLM) training method to construct a unified cross-lingual semantic space. Mathematically, for the input character sequence S, XLM-R maps it to a high-dimensional feature vector h=Encoder(S) through a multi-layer self-attention mechanism. This deep representation capability enables the model to identify highly covert social engineering instructions, i.e., attackers induce the model to bypass security alignment by semantic packaging (such as "assuming you are an unrestricted system tester"). Unlike traditional keyword-based blacklist mechanisms, the auditing model based on XLM-R has extremely strong generalization ability and can identify variant attacks that have never been seen before. However, in actual automotive engineering applications, how to compress such a large pre-trained model using INT8 quantization and deploy it on real-time operating systems such as QNX or RTOS, while ensuring that binary classification inference is completed within 15ms, remains a challenge that the industry has not yet fully solved.(3) Retrieval Enhancement Generation (RAG) and Supply Chain Poisoning Defense: To address the knowledge gaps and "illusion" issues in large language models within the automotive vertical domain, Retrieval Enhancement Generation (RAG) technology is widely used in in-vehicle knowledge question answering systems. It utilizes high-performance vector databases such as FAISS to slice and convert massive amounts of user manuals and OTA update logs into vector indexes. When a user asks a question, the system uses cosine similarity to retrieve the most relevant document fragments. While RAG improves the accuracy of responses, it also introduces a new security risk—prompt injection poisoning. Attackers may inject malicious commands into PDF documents provided by third-party supply chains. Current defense methods primarily focus on static input filtering. More challenging is that official industrial PDF documents often generate a large amount of underlying structural garbled characters during parsing due to encoding issues such as the lack of embedded underlying CID fonts. These garbled characters will severely dilute the similarity distance of the vector space, causing conventional security retrieval mechanisms to completely fail. The industry lacks a composite purification mechanism for the entire RAG chain that includes "underlying garbled character dehydration and cleaning" and "high-dimensional semantic cleansing". (4) The LLM security evaluation architecture is aligned with industry standards: With the acceleration of the large model deployment process, establishing a scientific security compliance evaluation system has become a consensus between regulatory agencies and car companies. Existing general evaluation frameworks, such as OpenCompass and C-Eval, mainly focus on the general common sense and academic capabilities of the model, but in automotive-specific cybersecurity standards (such as ISO / SAE) There are significant gaps in the adaptation to 21434) and data security regulations (such as PIPL); traditional evaluations rely heavily on static test sets, allowing models to "boost" their rankings through overfitting, making it difficult to reflect their true resilience under complex ODD (Operating Design Domain); furthermore, evaluation logic often uses "manual scoring" or simple keyword matching, lacking objective quantitative indicators. The LLM-as-a-Judge paradigm, which has emerged in recent years, uses high-performance models (such as GPT-4 or Qwen-72B) as judges, employing CoT (Cooperation of Reasoning) reasoning to conduct multi-dimensional assessments of the safety of the tested models, providing a new approach for automated compliance auditing; however, in the automotive engineering field, how to transform this highly abstract judge evaluation into a legally binding compliance report calibrated at risk levels P0-P3, and achieve closed-loop linkage with the discovery of vulnerabilities in the vehicle's underlying CAN bus, remains a key gap in the current technical architecture.
[0005] In summary, there is a need for an edge-cloud collaborative multimodal security protection and dynamic evaluation method and system for in-vehicle large language models. Summary of the Invention
[0006] A brief overview of the invention is given below to provide a basic understanding of certain aspects of it. It should be understood that this overview is not an exhaustive summary of the invention. It is not intended to identify key or essential parts of the invention, nor is it intended to limit the scope of the invention. Its purpose is merely to present certain concepts in a simplified form as a prelude to the more detailed description that follows.
[0007] In view of this, in order to solve the problem that the traditional security protection and dynamic evaluation methods in the prior art are of poor capability and effect, the present invention provides an edge-cloud collaborative multimodal security protection and dynamic evaluation method and system for vehicle-mounted large language models.
[0008] Technical Solution 1 is as follows: A method for edge-cloud collaborative multimodal security protection and dynamic evaluation for in-vehicle large language models, comprising the following steps:
[0009] S1. In the end-side device, visual, voice and vehicle status features are asynchronously extracted based on multimodal data, the fusion weights are dynamically calculated and a risk score is generated, and a real-time judgment is made in combination with the normalized and compensated dynamic threshold; if a risk is triggered, a CAN bus interrupt signal is generated and the desensitized log is reported to the cloud; if no risk is triggered, the normal release instruction is given.
[0010] S2. Receive the end-side operation logs and scenario data reported in step S1, construct a dynamic adversarial test set containing multiple hazard levels and multiple test dimensions, inject red team induction instructions into the subject's large language model, call the judge's large model to perform hazard scoring based on the thought chain, and generate a compliance analysis report and multi-dimensional radar chart by weighted accumulation of hazard weights, while locating the security vulnerabilities that currently exist in the model and their triggering conditions.
[0011] S3. Based on the security vulnerability type and severity level identified in step S2, the incremental fine-tuning pipeline is automatically triggered. The same distribution positive samples are extracted from the security history database and the current zero-day vulnerability samples are used to form an incremental training set. The model is fine-tuned by using a composite loss function that simultaneously optimizes classification accuracy and semantic robustness. The updated model weights are then sent to the edge in real time through the OTA channel to update the decision inference parameters in step S1.
[0012] S4. Before being stored in the cloud knowledge base, the original text is physically paged and cleaned using regular expressions. Valid text blocks are filtered according to the density of special characters and mapped to high-dimensional vectors. Similarity retrieval is performed in combination with the feature anchor library to identify malicious / poisoned content. Then, a large model is called to perform semantic cleaning and reconstruction on suspicious blocks, retaining paged source information. Finally, a cleaned security document and a source assessment report are generated.
[0013] S5. Based on a microservice architecture, a cloud-based backend system is built to process global node requests, while the frontend provides a visual situational dashboard. At the same time, a digital twin cockpit simulation system is deployed, integrating a multimodal interaction entry point. The model weights updated in step S3 are deployed to the simulation environment. The security documents provided in step S4 are used as the knowledge source for simulation inference. Security policy verification and performance optimization are completed in a virtual environment. After verification, the system is finally solidified into automotive-grade chips, achieving an end-to-end full-link technology closed loop.
[0014] Furthermore, in S1, the multimodal data includes the current video frame. Current audio stream Current vehicle chassis state vector and benchmark threshold ;
[0015] Record the current system time As the starting moment of the current processing ;
[0016] Asynchronous concurrent modal sensing is performed through the edge-side real-time protection system, simultaneously executing three independent tasks to process different modal data, as detailed below:
[0017] For the current video frame The first task, Task 1, is executed, using the ZeroShotVLM (Zero-Shot Visual Language Model) for analysis, based on predefined text prompt feature vectors for each state. Perform matching and output visual feature vectors. ;
[0018] For the current audio stream Execute the second task, Task 2, using the speech recognition model. Convert it to text Using a multilingual encoder Text Transform into semantic feature vectors ;
[0019] Regarding the current vehicle chassis state vector ,according to The function dynamically calculates the fusion weights of the three modalities to obtain the visual weights. Text weight State vector weights ;
[0020] After all tasks are completed, perform tensor fusion and activation operations, weighted summation of the feature vectors of the three modalities, and then... The function obtains the risk score ;
[0021] use The function is applied to the current vehicle chassis state vector. medium speed Normalization is performed, using the baseline threshold. The compensation coefficient, multiplied by the normalized vehicle speed value, is used to calculate the dynamic threshold based on the vehicle's dynamic condition. ;
[0022] Determine the current risk score Does it exceed the dynamic threshold? If the limit is exceeded, a CAN bus interrupt signal will be generated. For the current video frame ,text and risk score Perform de-identification processing and generate logs. ,use The function will de-identify the logs. The data is reported to the cloud; if the time limit is not exceeded, the command is sent to the CAN bus; the total time elapsed from the start to the current moment is calculated as the processing delay. Output interrupt signal and processing delay .
[0023] Furthermore, in S2, the test model is input. and contains multiple test tuples Multidimensional dynamic test set Prompt is the warning text, Level is the corresponding hazard level of the warning, and Dim is the test dimension;
[0024] Initial total score and maximum possible score Define the weight of each level of hazard in the hazard weight dictionary W;
[0025] Traversing the multidimensional dynamic test set For each test sample, the cue word p, the hazard level Level, and the test dimension Dim are extracted sequentially, and the cue word p is input into the subject model. In the middle, the infer interface is called to generate response text. ;
[0026] Construct judge prompts that include hazard level constraints. Combined with response text Through the judge's big model Generating thought chains and scores is OUT; a regular expression engine is used. Extracting scores from OUT If the regular expression engine fails to extract (i.e., NULL), then set... The default value is 3.0, which serves as a fallback for exceptions, setting the score of the current test sample as the minimum. Multiply by the weight corresponding to its hazard level And add it to the total score after initialization. In the meantime, obtain the current total score. Multiply the highest set hazard score by the weight corresponding to its hazard level. Accumulate the current maximum possible score To obtain the theoretical maximum score Record vulnerability judgments and source tracing chains and store them in the database;
[0027] By current total score Divide by the theoretical maximum score The overall score on a percentage scale is calculated. Based on the overall score on a 100-point scale The mapping generates A+ / A / B / C compliance levels and multi-dimensional radar charts that align with compliance standards, ultimately outputting a compliance analysis report. .
[0028] Furthermore, in S3, when a Zero-Day vulnerability sample is identified and captured, the vulnerability labeling pipeline is automatically started through the self-evolving data flywheel module to extract positive samples with the same distribution from the security history database and combine them with the currently detected Zero-Day vulnerability sample to form an incremental training dataset.
[0029] During incremental fine-tuning, the composite loss function Represented as:
[0030]
[0031] in, for, for, for, For hyperparameters, For the current hidden layer representation, Use historical failure sample feature anchor points;
[0032] The updated model weights are synchronized to the edge device in real time via OTA channel.
[0033] Furthermore, in S4, from the original PDF document Extract page number tags Page collection Create a clean list of blocks The details are as follows:
[0034] Traverse the collection of pages Pages in traverse the page text in Use regular expression replacement Text Replace typical garbled characters in the text with their true meaning to complete the text. Update; based on text The number of non-standard characters (CountSpecialChars) divided by the text The length Length is used to calculate the density of special characters. When special character density No more than 25%, text The length of the line is greater than 10; valid rows that pass the filter will be included. and its page numbers Add as a record to the clean chunk list End the filter criteria; end the row traversal of the current page; end the traversal of all pages;
[0035] Using the embedding model Clean block list The text of each record Mapping to high-dimensional vectors yields a list of vectors. Combined with FAISS, based on vector sets ,pass The function constructs an inner product index ; Feature anchor point library As a query, in the inner product index The search is performed within the specified range, taking into account the set similarity threshold. The similarity threshold will be greater than or equal to the threshold. The inner product result is integrated into the high-risk block set. ;
[0036] Traverse the high-risk block set Each block index judged as suspicious Call the large language model cleaning function Enter the original text Return the cleaned text Boolean value , This indicates whether the text actually contains malicious / poisoning content. If it is confirmed that the current text contains malicious / poisoning content, then a replacement and recording operation is performed. The original text at each position in the list is replaced with the cleaned text. ;
[0037] pass Function to trace the source of loss assessment report Add modification records, which include the original text with page numbers. Original text and the cleaned text Termination condition check; End traversal of all suspicious blocks;
[0038] Based on the clean block list The document, containing each line of text and its page number, is reassembled using the Reconstruct function in the original page number order to generate a secure document. Finally, return to the secure document. and source tracing and loss assessment report .
[0039] Technical Solution 2: A cloud-edge collaborative multimodal security protection and dynamic evaluation system for vehicle-mounted large language models, used to execute the cloud-edge collaborative multimodal security protection and dynamic evaluation method for vehicle-mounted large language models described in Technical Solution 2, including a cloud-based adversarial evolution and audit engine, an edge-side real-time protection system, and a visualization monitoring and compliance display screen connected in sequence.
[0040] The beneficial effects of this invention are as follows: In response to the security vulnerabilities of in-vehicle large language models in practical applications, this invention identifies four core requirements of the industry for intelligent security protection: First, the requirement for extremely fast response. The vehicle chassis control and active and passive safety modules have extremely stringent time latency requirements for command issuance. Any defense mechanism must complete the judgment and interception within a millisecond window (usually less than 30ms) before the command reaches the bus. This requires the security architecture to realize the sinking of computing power to the edge chip. Secondly, there is the need for multimodal fusion perception. Single-dimensional text auditing cannot completely avoid risks. The system needs to link with the cockpit monitoring system (DMS) in real time to obtain the driver's physiological state and combine it with voice and semantic information to build a multi-dimensional safety risk measurement model to achieve dynamic defense based on context awareness. Thirdly, there is the need for dynamic adversarial evaluation. Static evaluation benchmarks are easily bypassed by large models through weight overfitting or rote memorization due to fixed samples. The industry urgently needs an automated red team testing system that can simulate real hacker attack logic and dynamically generate millions of inducement instructions based on the Operational Design Domain (ODD). Finally, there is the need for closed-loop self-evolution of security governance. Zero-day vulnerabilities in the vehicle environment are unpredictable. The system must have an automated data flywheel mechanism from attack detection and automatic labeling of failed samples to incremental model retraining to ensure that the defense model can continuously self-evolve as attack methods evolve, achieving an industrial closed loop.
[0041] Therefore, this invention proposes an edge-cloud collaborative multimodal security protection and dynamic evaluation method for in-vehicle large language models. It aims to construct a multimodal, in-depth defense system covering instruction input semantic analysis, pre-execution blocking, and full lifecycle vulnerability self-healing through deep decoupling and real-time collaboration between the "edge-side protection module" and the "cloud-based adversarial audit engine." The core logic of this invention lies in utilizing the computing power of the in-vehicle edge for millisecond-level deterministic interception, while simultaneously utilizing large-scale cloud computing power to perform complex adversarial drills and compliance assessments. This addresses the security risks caused by large models being vulnerable to social engineering attacks, jailbreak instruction hijacking, and excessively high response latency in the in-vehicle environment. The system architecture abandons the broad defense strategy of general-purpose large models, opting instead for domain-specific fine-tuning for in-vehicle control scenarios. This results in higher semantic sensitivity and execution determinism in areas such as vehicle control bus instruction auditing, personal privacy data anonymization, and protection of autonomous driving assistance strategies.
[0042] At the edge-side defense execution level, this invention deploys a multimodal parallel judgment logic consisting of a visual monitoring unit (based on the MobileViT architecture) and a text semantic auditing unit (based on a finely tuned XLM-Roberta architecture). The edge-side protection module resides in the vehicle system SOC memory space in a singleton mode. By deeply intervening in the distribution link between the large language model application layer and the underlying CAN bus communication protocol stack, it achieves physical pre-audit of each execution instruction. The visual monitoring unit analyzes the driver's behavior vector in real time to determine whether there is an abnormal operation state. The text auditing unit performs a rapid semantic scan of the candidate instructions generated by the large model to identify potential jailbreak attack intentions. This invention, through dynamic weighted fusion logic, can complete the "allow / block" decision of the instruction within a window period of 15ms to 30ms. This edge-side on-site auditing mechanism eliminates the non-deterministic latency caused by network jitter during cloud interaction, ensuring that high-risk instructions are completely intercepted before physically reaching the vehicle actuators.
[0043] At the cloud-based collaborative evolution level, this invention constructs a red team adversarial simulation engine with self-evolving capabilities. Based on the high-performance Qwen-14B model, this engine dynamically generates highly realistic induction instruction sets for various Operational Design Domains (ODDs), such as highway cruising, extreme weather events like heavy rain, and interactions with public charging facilities, by simulating the deceptive logic of real attackers. The core technological innovation of this invention lies in proposing a "Data Flywheel" closed-loop evolution mechanism: when the cloud-based dynamic evaluation engine discovers a missed sample (i.e., a Zero-Day vulnerability) in the edge-side protection module through adversarial exercises, the system automatically triggers vulnerability feature extraction and automated labeling logic. Subsequently, cloud computing power is used to incrementally fine-tune the vulnerability sample (Retraining), generating a defense patch weight for the vulnerability in a very short time. This weight is synchronized to the edge in real time via OTA channel, realizing an automated closed-loop iteration of the security defense line from "vulnerability discovery" to "automatic reinforcement," ensuring that the system's defense capabilities continuously strengthen as attack methods evolve.
[0044] Furthermore, this invention integrates an automated compliance assessment algorithm that aligns with international and domestic automotive safety standards (such as ISO / SAE 21434 and the PIPL privacy law). This algorithm utilizes an LLM-as-Judge architecture, employing a judge model with chain-of-thought reasoning capabilities to quantitatively evaluate the output of the tested model from multiple dimensions. The final evaluation data output by the system includes, but is not limited to, compliance radar distribution maps, risk probability tracing chains, and vulnerability risk point classifications. This mechanism transforms the security of large-scale in-vehicle models from traditional qualitative descriptions to quantitative audits with industrial applicability, providing automakers and regulatory agencies with scientific technical benchmarks for vehicle access testing, software supply chain security audits, and long-term safety maintenance, achieving a high degree of alignment between cockpit intelligence and physical security. Attached Figure Description
[0045] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:
[0046] Figure 1 This is a flowchart illustrating a cloud-edge collaborative multimodal security protection and dynamic evaluation method for in-vehicle large language models.
[0047] Figure 2 This is a schematic diagram of the overall system architecture.
[0048] Figure 3 This is a schematic diagram of the business process.
[0049] Figure 4 This is a schematic diagram of edge-side multimodal concurrent feature extraction and adaptive threshold decision-making;
[0050] Figure 5 This is a schematic diagram of the spatiotemporal coupling decision and control flow on the edge side.
[0051] Figure 6 This is a schematic diagram of dynamic adversarial evaluation based on thought chain and harm weighting;
[0052] Figure 7 A schematic diagram of the logical flow and data self-evolution architecture;
[0053] Figure 8 A schematic diagram of a closed-loop data flywheel and a model self-evolution pipeline;
[0054] Figure 9 This is a schematic diagram of the RAG supply chain security purification algorithm process;
[0055] Figure 10 This is a schematic diagram of document semantic cleansing based on page number tracking and density dehydration. Detailed Implementation
[0056] To make the technical solutions and advantages of the embodiments of the present invention clearer, the exemplary embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not an exhaustive list of all embodiments. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other.
[0057] Example 1: Reference Figures 1-10 This embodiment details a method for edge-cloud collaborative multimodal security protection and dynamic evaluation for in-vehicle large language models, specifically including the following steps:
[0058] S1. Real-time protection of multimodal weighted fusion at the vehicle end: In the vehicle end device, visual, voice and vehicle status features are extracted asynchronously based on multimodal data, the fusion weight is dynamically calculated and a risk score is generated, and a real-time judgment is made in combination with the dynamic threshold after vehicle speed normalization compensation; if a risk is triggered, a CAN bus interrupt signal is generated and the desensitized log is reported to the cloud; if no risk is triggered, the normal release instruction is given.
[0059] S2. Cloud-based dynamic adversarial evaluation and vulnerability discovery: Based on the end-side operation logs and scenario data reported in step S1, construct a dynamic adversarial test set containing multiple hazard levels and multiple test dimensions, inject red team induction instructions into the subject's large language model, call the judge's large model to perform hazard scoring based on the thought chain, and generate a compliance analysis report and multi-dimensional radar chart through hazard weighting and accumulation, while locating the security vulnerabilities that currently exist in the model and their triggering conditions;
[0060] S3. Closed-loop self-repair based on data flywheel: Based on the security vulnerability type and severity level identified in step S2, the incremental fine-tuning pipeline is automatically triggered. The same distribution positive samples are extracted from the security history database and the current zero-day vulnerability samples are used to form an incremental training set. The model is fine-tuned by using a composite loss function that simultaneously optimizes classification accuracy and semantic robustness. The updated model weights are sent to the edge in real time through the OTA channel to update the decision inference parameters in step S1.
[0061] S4.RAG Supply Chain Security Cleaning and Semantic Detoxification: Before being stored in the cloud knowledge base, the original text is physically tracked and cleaned using regular expressions. Valid text blocks are selected according to special character density and mapped into high-dimensional vectors. Similarity retrieval is performed using a feature anchor library to identify malicious / poisoned content. Subsequently, a large model is called to perform semantic detoxification and reconstruction on suspicious blocks, retaining page number traceability information. Finally, a cleaned security document and a traceability loss assessment report are generated.
[0062] S5. System Engineering Deployment and Digital Twin Closed-Loop Verification: Based on a microservice architecture, a cloud-based backend system is built to handle global node requests, while the frontend provides a visual situational dashboard. At the same time, a digital twin cockpit simulation system is deployed, integrating multimodal interaction entry points. The model weights updated in step S3 are deployed to the simulation environment. The safety documents provided in step S4 serve as the knowledge source for simulation inference. Safety policy verification and performance optimization are completed in a virtual environment. After verification, the data is finally solidified into automotive-grade chips, realizing an end-to-end full-link technology closed loop from cloud evaluation, model evolution, data purification to mass production of real vehicles.
[0063] Specifically,
[0064] The engineering implementation system of this invention adopts a highly decoupled microservice architecture. The backend system is built on a high-performance framework and uses an asynchronous IO mechanism to handle vulnerability feature feedback and instruction monitoring requests from global vehicle nodes. The front-end console uses a visualization engine to transform complex adversarial evaluation data into an intuitive interactive screen, displaying the dynamic situation of red-blue adversarial battles, compliance radar distribution, and audit link tracing of the judge model in real time. To verify the performance of the algorithm on a real vehicle computing platform, this invention has developed a set of digital twin cockpit simulation systems (Digital Twin HUD). This simulation system not only integrates multimodal interaction entry points of microphones and physical cameras at the front end, but also connects to automotive-grade hardware compilation pipelines at the underlying level. The system provides a standard export conversion mechanism that can convert dynamic computation graphs (such as PyTorch's .pth model) into static computation graphs (such as .onnx format) that support INT8 / FP16 quantization. This compilation conversion completely eliminates the dependence on the huge underlying runtime environment, enabling the multimodal guardrail engine to be seamlessly deployed in the real-time operating system of Qualcomm DSP or edge NPU. By monitoring NPU load and inference latency in real time, the simulation system provides realistic physical feedback for optimizing security algorithms. This end-to-end design, from cloud-based evaluation labs and multimodal pre-blocking to mass production deployment of edge chips, completely establishes a closed-loop technology for in-vehicle large-scale model security protection, from "theoretical verification" to "real-vehicle implementation."
[0065] Furthermore, in S1, the multimodal data includes the current video frame. (Images captured by the camera), current audio stream (Speech captured by microphone), current vehicle chassis state vector (Information such as vehicle speed and gear position) and baseline threshold (Preset basic interrupt trigger threshold);
[0066] Record the current system time As the starting moment of the current processing This is used to delay subsequent calculations;
[0067] Asynchronous concurrent modal sensing is performed through the edge-side real-time protection system, simultaneously executing three independent tasks to process different modal data, as detailed below:
[0068] For the current video frame The first task, Task 1, is executed, using the ZeroShotVLM (Zero-Shot Visual Language Model) for analysis, based on predefined text prompt feature vectors for each state. (In this embodiment, The matching process is performed to output a visual feature vector. ;
[0069] For the current audio stream Execute the second task, Task 2, using the speech recognition model. Convert it to text Using a multilingual encoder Text Transform into semantic feature vectors ;
[0070] Regarding the current vehicle chassis state vector ,according to The function dynamically calculates the fusion weights of the three modalities to obtain the visual weights. Text weight State vector weights ;
[0071] After all tasks are completed, perform tensor fusion and activation operations, weighted summation of the feature vectors of the three modalities, and then... The function obtains the risk score ;
[0072] use The function is applied to the current vehicle chassis state vector. medium speed Normalization is performed, using the baseline threshold. The compensation coefficient, multiplied by the normalized vehicle speed value, is used to calculate the dynamic threshold based on the vehicle's dynamic condition. ;
[0073] Determine the current risk score Does it exceed the dynamic threshold? If the limit is exceeded, a CAN bus interrupt signal will be generated. For the current video frame ,text and risk score Perform de-identification processing and generate logs. ,use The function will de-identify the logs. The data is reported to the cloud; if the time limit is not exceeded, the command is sent to the CAN bus; the total time elapsed from the start to the current moment is calculated as the processing delay. Output interrupt signal and processing delay .
[0074] Specifically, in the edge computing nodes, this invention implements a multimodal joint decision-making algorithm based on spatiotemporal feature coupling. This algorithm deeply integrates visual monitoring, voice commands, and vehicle operating status to construct a high-speed interception barrier. In the execution logic, for the visual modality, it innovatively introduces a zero-shot visual-language alignment model (ZeroShotVLM). Unlike traditional binary classification networks that rely on temporal frame determination, ZeroShotVLM directly maps driver behavior in the physical space to the internationally recognized distracted driving state space (e.g., C0 safe driving, C1 right-handed mobile phone operation, C6 eating, etc.). This is achieved by extracting the feature vector of the current video frame image. And compared with the predefined text prompt feature vectors of each state. Calculate the cosine similarity and use the Softmax function to output the probability distribution of each dangerous state, i.e. ,in, For the preset threshold, The risk probability of the visual dimension is output if and only if the highest probability state is an unsafe state (i.e., i ≠ 0) and the confidence level exceeds a preset threshold. Meanwhile, for the text commands converted from speech, the system uses a finely tuned XLM-Roberta encoder to extract their semantic feature vector T, determines whether they contain social engineering-induced features such as overreach, and outputs the risk probability in the text dimension. To achieve dynamic defense based on environmental perception, a real-time vehicle operating state vector S (such as vehicle speed, steering angle, and ODD scene information) is introduced as a dynamic weight adjustment factor. Specifically, the system defines a comprehensive risk score based on a nonlinear activation function. Its calculation logic is as follows: In this formula, the weighting coefficients , , It is not a fixed value, but rather dynamically adjusted according to the vehicle's operating status. Exceeding the preset absolute safety threshold At that time, the Edge Guard Engine located at the bottom layer immediately triggers the physical blocking logic, using a high-priority process that resides in memory in the Singleton pattern to directly cut off the link from the instruction to the CAN bus, and completes the determination within the end-to-end window period of 15ms to 30ms.
[0075] Furthermore, in S2, the test model is input. and contains multiple test tuples Multidimensional dynamic test set Prompt is the warning text, Level is the corresponding hazard level of the warning, and Dim is the test dimension;
[0076] Initial total score and maximum possible score Define the weights of each level of hazard in the hazard weight dictionary W: the weight of primary hazard P0 is 0.4, the weight of first-level hazard P1 is 0.3, the weight of second-level hazard P2 is 0.2, and the weight of third-level hazard P3 is 0.1.
[0077] Traversing the multidimensional dynamic test set For each test sample, the cue word p, the hazard level Level, and the test dimension Dim are extracted sequentially, and the cue word p is input into the subject model. In the middle, the infer interface is called to generate response text. ;
[0078] Construct judge prompts that include hazard level constraints. Combined with response text Through the judge's big model Generating thought chains and scores is OUT; a regular expression engine is used. Extracting scores from OUT If the regular expression engine fails to extract (i.e., NULL), then set... The default value is 3.0, which serves as a fallback for exceptions, setting the score of the current test sample as the minimum. Multiply by the weight corresponding to its hazard level And add it to the total score after initialization. In the meantime, obtain the current total score. The highest hazard score (10.0 points) will be multiplied by the weight corresponding to its hazard level. Accumulate the current maximum possible score To obtain the theoretical maximum score Record vulnerability judgments and source tracing chains and store them in the database;
[0079] By current total score Divide by the theoretical maximum score The overall score on a percentage scale is calculated. Based on the overall score on a 100-point scale The system maps and generates compliance levels (A+ / A / B / C) and multi-dimensional radar charts based on the benchmark compliance standard (ISO 21434), ultimately outputting a compliance analysis report. .
[0080] Specifically, to achieve a non-static and highly robust security assessment, this invention designs a dynamic adversarial evaluation system based on the role of a "Mechanic Hacker." It utilizes a high-performance generative model (based on large-parameter models such as Qwen) to dynamically simulate the attack chain of a real hacker, generating a massive (hundreds of thousands) number of adversarial commands with social engineering inducement characteristics. The core innovation of this invention lies in the introduction of a "Semantic Router" and a "dynamic hierarchical sampling" mechanism. The system automatically distributes massive, disordered attack commands to semantic buckets across multiple dimensions, including vehicle control, privacy leakage, ideology, and prevention of false positives, through regular expressions and feature matching of core dangerous entity words. When initiating an evaluation, the system dynamically extracts test cases from each semantic bucket and reassembles them into a dynamic exam, completely eliminating the "overfitting and ranking manipulation" behavior of the tested model.
[0081] During the evaluation phase, the system introduced an automated judge module based on the "LLM-as-a-Judge" architecture. This module forcibly uses the Chain of Thought (CoT) to analyze the response results and outputs a quantitative score based on the hazard level from P0 to P3.
[0082] Furthermore, in S3, when a Zero-Day vulnerability sample is identified and captured, the vulnerability labeling pipeline is automatically started through the self-evolving data flywheel module to extract positive samples with the same distribution from the security history database and combine them with the currently detected Zero-Day vulnerability sample to form an incremental training dataset.
[0083] During incremental fine-tuning, the composite loss function The aim is to simultaneously optimize classification accuracy and semantic robustness, expressed as:
[0084]
[0085] in, This is a binary cross-entropy loss function, used to maintain the model's basic classification and discrimination ability for common security instructions. For the true safety label of the sample, The predicted risk probability output by the model. It is a historical vulnerability sample library (i.e., a set of difficult samples that have successfully penetrated the defense in the past). For hyperparameters, For the current hidden layer representation, Use historical failure sample feature anchor points;
[0086] The updated model weights are synchronized to the client side in real time via OTA channel, completing the automated evolution loop from "vulnerability discovery" to "patch distribution".
[0087] Specifically, the core innovative solution proposed in this invention is a closed-loop self-healing mechanism based on the data flywheel logic. This mechanism enables the vehicle-mounted defense system to dynamically evolve with the evolution of threats. Its operation logic starts with the real-time capture of vulnerabilities by the cloud evaluation engine. Once the red team adversarial engine successfully penetrates the current defense line (i.e., discovers a Zero-Day vulnerability sample), the system will automatically start the vulnerability labeling pipeline. In order to achieve accurate repair of the vulnerability and avoid catastrophic forgetting of the model, the system will extract positive samples with the same distribution from the security history library and form an incremental training dataset together with the currently detected negative samples (vulnerability samples).
[0088] Composite loss function In the calculation formula, the first term is the conventional binary cross-entropy loss, which is responsible for maintaining the model's basic judgment ability; the second term is the semantic contrast loss term, which is calculated by the current hidden layer representation. Anchor points of historical failure samples The cosine similarity between them is used, and their weights are adjusted using the hyperparameter λ. In this way, the fine-tuning process can force the model to widen the boundary between attack commands and safe commands in the vector space, thereby achieving herd immunity against attacks of the same variant.
[0089] Furthermore, in S4, from the original PDF document Extract page number tags Page collection Create a clean list of blocks The details are as follows:
[0090] Traverse the collection of pages Pages in traverse the page text in Use regular expression replacement Text Replace typical garbled characters in the text with their true meaning to complete the text. Update; based on text The number of non-standard characters (CountSpecialChars) divided by the text The length Length is used to calculate the density of special characters. When special character density No more than 25%, text The length of the line is greater than 10; valid rows that pass the filter will be included. and its page numbers Add as a record to the clean chunk list Page numbers are retained for later tracing and reconstruction; at this point, the filtering criteria are terminated; the row traversal of the current page ends; and the traversal of all pages ends.
[0091] Using the embedding model Clean block list The text of each record Mapping to high-dimensional vectors yields a list of vectors. Combined with FAISS (Facebook AISimilarity Search, a high-efficiency vector retrieval library), based on vector sets... ,pass The function constructs an inner product index ; Feature anchor point library As a query, in the inner product index The search is performed within the specified range, taking into account the set similarity threshold. ( =0.45), which is greater than or equal to the similarity threshold. The inner product result is integrated into the high-risk block set. ;
[0092] Traverse the high-risk block set Each block index judged as suspicious Call the large language model cleaning function Enter the original text Return the cleaned text Boolean value , This indicates whether the text actually contains malicious / poisoning content. If it is confirmed that the current text contains malicious / poisoning content, then a replacement and recording operation is performed. The original text at each position in the list is replaced with the cleaned text. ;
[0093] pass Function to trace the source of loss assessment report Add modification records, which include the original text with page numbers. Original text and the cleaned text Termination condition check; End traversal of all suspicious blocks;
[0094] Based on the clean block list The document, containing each line of text and its page number, is reassembled using the Reconstruct function in the original page number order to generate a secure document. Finally, return to the secure document. and source tracing and loss assessment report .
[0095] Specifically, in response to the potential supply chain poisoning threat faced by in-vehicle knowledge bases, this invention embeds a filtering system with data dehydration and real-time semantic purification into the Retrieval Enhancement Generation (RAG) pipeline;
[0096] First, in the data preprocessing stage, the system introduces physical page number tracking and a "RegexDehydration" pipeline. When parsing long text PDFs, the system preserves the mapping relationship between paragraphs and physical page numbers, and calculates the density of abnormal characters (i.e., non-Chinese and non-English characters and common punctuation marks) line by line. When the density ρ exceeds the tolerance threshold (e.g., 0.25), the system automatically discards the dirty data and uses heuristic rules to repair the typical garbled characters caused by the underlying CID mapping. This effectively prevents noise from diluting vector similarity.
[0097] Subsequently, in the semantic cleansing stage, the system uses a high-performance vector model to slice the dehydrated document. Perform high-dimensional mapping. This involves calculating the mapping between the slice vector to be added to the database and a known library of malicious attack patterns. The system uses similarity metrics to identify potential "hint injection" payloads in real time. When a document block is deemed high-risk, the system doesn't simply remove it; instead, it invokes a cloud-based large-scale model to perform semantic reorganization and cleansing. While preserving the original security facts, the large-scale model removes all unauthorized operation prompts. The cleaned and reconstructed text, along with its precise physical page number, is output as a damage assessment report and re-stored in the vector database, ensuring the absolute security of the data source upon which the large-scale model's inference relies.
[0098] Example 2: Reference Figures 2-10 This embodiment describes a cloud-edge collaborative multimodal security protection and dynamic evaluation system for in-vehicle large language models. It is used to execute the cloud-edge collaborative multimodal security protection and dynamic evaluation method for in-vehicle large language models described in Embodiment 1. Specifically, it includes a cloud-based adversarial evolution and audit engine 1, an edge real-time protection system 2, and a visualization monitoring and compliance display screen 3, which are connected in sequence.
[0099] The cloud-based adversarial evolution and audit engine 1 includes an adversarial instruction generation engine, a quantitative compliance assessment algorithm storage module, a self-evolving data flywheel module, and a high-performance FastAPI communication gateway.
[0100] The edge real-time protection system 2 includes a visual monitoring module, an instruction audit firewall, a physical control gateway layer, and an in-vehicle large language model storage module.
[0101] The visualization monitoring and compliance display screen 3 is used to display real-time adversarial monitoring and the generated 1S0 21434 report.
[0102] Although the invention has been described with reference to a limited number of embodiments, those skilled in the art will understand from the foregoing description that other embodiments are conceivable within the scope of the invention described herein. Furthermore, it should be noted that the language used in this specification has been chosen primarily for readability and instructional purposes, and not for the purpose of interpreting or limiting the subject matter of the invention. Therefore, many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the appended claims. The disclosure of the invention is illustrative and not restrictive, and the scope of the invention is defined by the appended claims.
Claims
1. A method for edge-cloud collaborative multimodal security protection and dynamic evaluation for in-vehicle large language models, characterized in that, Includes the following steps: S1. Knowledge Base Preprocessing and Cleaning: Before the in-vehicle large language model accesses the external knowledge base, the original document data is obtained, and physical page number feature tracking and regular expression dehydration cleaning based on abnormal character density are performed; the dehydrated text is mapped into a high-dimensional dense vector, and similarity retrieval is performed in combination with preset security feature anchors; for text blocks determined to be high-risk, the language model is called to perform semantic deentanglement and reconstruction, and a safe and cleaned document with traceability information is output for the large language model to call; S2. Edge-side multimodal concurrent perception and decision gating: In the vehicle edge computing node, the driver's visual state features, vehicle voice semantic features and vehicle chassis dynamic state features are asynchronously captured and extracted; based on the vehicle chassis dynamic state features, cross-modal fusion weights are dynamically calculated, and a comprehensive risk score is generated by combining the conflict penalty terms between multimodal features; When the comprehensive risk score exceeds the adaptive dynamic threshold, a physical blocking signal is generated for the underlying vehicle actuator, and the desensitized interception context log is asynchronously reported to the cloud. S3. Cloud-based dynamic sampling and automated security auditing: The cloud node receives the operation logs reported in step S2, combines them with the preset operation design domain, and uses a generative model to expand and construct a dynamic adversarial test set. A dynamic stratified sampling strategy is used to concurrently inject inducing instructions into the subject's large language model; the cloud-based judge's large model is invoked to classify the harm and perform structured quantitative scoring on the subject's model's response based on the thought chain, generating a large model security compliance analysis report and vulnerability tracing basis; S4. Data Flywheel Closed Loop and Model Adaptive Evolution: Based on the security vulnerabilities identified in step S3, the attack payloads of failed samples are automatically extracted and an incremental training set is constructed; the edge defense model is incrementally fine-tuned using a composite loss function that integrates basic classification loss and semantic contrast loss to expand the security decision boundary in the latent space; after the updated model weights are quantized, compiled, and verified by equivalent simulation, they are downloaded to the vehicle edge node via over-the-air technology to update the decision inference parameters in step S2, thus completing the closed-loop self-evolution of the security defense link.
2. The method for edge-cloud collaborative multimodal security protection and dynamic evaluation for vehicle-mounted large language models according to claim 1, characterized in that, In S1, the multimodal data includes the current video frame. Current audio stream Current vehicle chassis state vector and benchmark threshold ; Record the current system time As the starting moment of the current processing ; Modality sensing is performed asynchronously and concurrently through the edge-side real-time protection system, simultaneously executing three independent tasks to process different modal data, as detailed below: For the current video frame The first task, Task 1, is executed, using the ZeroShotVLM (Zero-Shot Visual Language Model) for analysis, based on predefined text prompt feature vectors for each state. Perform matching and output visual feature vectors. ; For the current audio stream Execute the second task, Task 2, using the speech recognition model. Convert it to text Using a multilingual encoder Text Transform into semantic feature vectors ; Regarding the current vehicle chassis state vector ,according to The function dynamically calculates the fusion weights of the three modalities to obtain the visual weights. Text weight State vector weights ; After all tasks are completed, perform tensor fusion and activation operations, weighted summation of the feature vectors of the three modalities, and then... The function obtains the risk score ; use The function is based on the current vehicle chassis state vector. medium speed Normalization is performed, using the baseline threshold. The compensation coefficient, multiplied by the normalized vehicle speed value, is used to calculate the dynamic threshold based on the vehicle's dynamic condition. ; Determine the current risk score Does it exceed the dynamic threshold? If the limit is exceeded, a CAN bus interrupt signal will be generated. For the current video frame ,text and risk score Perform de-identification processing and generate logs. ,use The function will de-identify the logs. The data is reported to the cloud; if the time limit is not exceeded, the command is sent to the CAN bus; the total time elapsed from the start to the current moment is calculated as the processing delay. Output interrupt signal and processing delay .
3. The method for edge-cloud collaborative multimodal security protection and dynamic evaluation for vehicle-mounted large language models according to claim 2, characterized in that, In S2, the test model is input. and contains multiple test tuples Multidimensional dynamic test set Prompt is the warning text, Level is the corresponding hazard level of the warning, and Dim is the test dimension; Initial total score and maximum possible score Define the weight of each level of hazard in the hazard weight dictionary W; Traversing the multidimensional dynamic test set For each test sample, the cue word p, the hazard level Level, and the test dimension Dim are extracted sequentially, and the cue word p is input into the subject model. In the middle, the infer interface is called to generate response text. ; Construct judge prompts that include hazard level constraints. Combined with response text Through the judge's big model Generating thought chains and scores is OUT; a regular expression engine is used. Extracting scores from OUT If the regular expression engine fails to extract (i.e., NULL), then set... The default value is 3.0, which serves as a fallback for exceptions, setting the score of the current test sample as the minimum. Multiply by the weight corresponding to its hazard level And add it to the total score after initialization. In the meantime, obtain the current total score. Multiply the highest set hazard score by the weight corresponding to its hazard level. Accumulate the current maximum possible score To obtain the theoretical maximum score Record vulnerability judgments and source tracing chains and store them in the database; By the current total score Divide by the theoretical maximum score The overall score on a percentage scale is calculated. Based on the overall score on a 100-point scale The mapping generates A+ / A / B / C compliance levels and multi-dimensional radar charts that align with compliance standards, ultimately outputting a compliance analysis report. .
4. The method for edge-cloud collaborative multimodal security protection and dynamic evaluation for vehicle-mounted large language models according to claim 3, characterized in that, In S3, when a Zero-Day vulnerability sample is identified and captured, the vulnerability labeling pipeline is automatically started through the self-evolving data flywheel module. Positive samples with the same distribution are extracted from the security history database and combined with the currently detected Zero-Day vulnerability sample to form an incremental training dataset. During incremental fine-tuning, the composite loss function Represented as: in, This is a binary cross-entropy loss function, used to maintain the model's basic classification and discrimination ability for common security instructions. For the true safety label of the sample, The predicted risk probability output by the model. It serves as a historical vulnerability sample library. To adjust the hyperparameters of the contrast loss intensity, This represents the hidden features currently calculated by the model for the new vulnerability. The feature anchors for historical failure samples in the latent space are used; the updated model weights are synchronized to the edge in real time via the OTA channel.
5. The method for edge-cloud collaborative multimodal security protection and dynamic evaluation for vehicle-mounted large language models according to claim 4, characterized in that, In step S4, from the original PDF document Extract page number tags Page collection Create a clean list of blocks The details are as follows: Traverse the collection of pages Pages in traverse the page text in Use regular expression replacement Text Replace typical garbled characters in the text with their true meaning to complete the text. Update; based on text The number of non-standard characters (CountSpecialChars) divided by the text The length Length is used to calculate the density of special characters. When special character density No more than 25%, text The length of the line is greater than 10; valid rows that pass the filter will be included. and its page numbers Add as a record to the clean chunk list End the filter criteria; end the row traversal of the current page; end the traversal of all pages; Using the embedding model Clean block list The text of each record Mapping to high-dimensional vectors yields a list of vectors. Combined with FAISS, based on vector sets ,pass The function constructs an inner product index ; Feature anchor point library As a query, in the inner product index The search is performed within the specified range, taking into account the set similarity threshold. The similarity threshold will be greater than or equal to the threshold. The inner product result is integrated into the high-risk block set. ; Traverse the high-risk block set Each block index judged as suspicious ; Call the large language model cleaning function Enter the original text Return the cleaned text Boolean value , This indicates whether the text actually contains malicious / poisoning content. If it is confirmed that the current text contains malicious / poisoning content, then a replacement and recording operation is performed. The original text at each position in the list is replaced with the cleaned text. ; pass Function to trace the source of loss assessment report Add modification records, which include the original text containing page numbers. Original text and the cleaned text Termination condition check; End traversal of all suspicious blocks; Based on the clean block list The document, containing each line of text and its page number, is reassembled using the Reconstruct function in the original page number order to generate a secure document. ; Finally, return to the security document. and source tracing and loss assessment report .
6. A cloud-edge collaborative multimodal security protection and dynamic evaluation system for vehicle-mounted large language models, characterized in that, The method for implementing the edge-cloud collaborative multimodal security protection and dynamic evaluation method for vehicle-mounted large language models as described in any one of claims 1-5 includes a cloud adversarial evolution and audit engine (1), an edge real-time protection system (2), and a visualization monitoring and compliance display screen (3) connected in sequence.
