Automobile information safety fuzzy test method and system based on large model

Abstract

The invention belongs to the technical field of automobile information safety fuzzy testing, discloses an automobile information safety fuzzy testing method based on a large model, and develops optimization application research of LLMs in an automobile software fuzzy testing scene by utilizing the advantages of an artificial intelligence technology in the aspects of data processing, natural language interaction, generalization ability and the like. The concealment and effectiveness of the test set generated by the fuzz test are optimized, and the fuzz test process efficiency is improved. Through the implementation of the invention, related key technologies are applied to the fuzz test of the automobile, and other related products and enterprises in China are radiated by writing technical standards, so that the fuzz test system can be used as an application demonstration and sample plate of the fuzz test of the intelligent driving automobile. According to market analysis, fuzzy test calculation is carried out on 100 thousand automobiles every year according to the method, and the data safety and the system safety are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of automobile information security fuzzy testing, and in particular relates to an automobile information security fuzzy testing method and system based on a large model. Background Art

[0002] With the development of intelligent and connected cars, the internal structure of cars is becoming increasingly complex, integrating a variety of communication technologies and smart devices, enabling vehicles to connect extensively with the outside world. Due to the diversification and complexity of electronic and electrical hardware in the car and the increase in interfaces, the risk of attacks on the vehicle network has increased exponentially. The car system is subject to multiple threats such as intrusion and remote control. If any electronic control power supply in the system is attacked, it is possible to paralyze the entire system. At the same time, the vehicle control software is complex, with many codes and difficult to prevent vulnerabilities. It is still difficult for manufacturers to find all defects through optimized testing.

[0003] As an important testing method and vulnerability mining method, fuzz testing is an indispensable part of vulnerability mining technology used to detect vehicle control network security and vehicle software. Fuzz testing injects random data into the target system to evaluate whether the target system has unexpected behavior. How to effectively apply fuzz testing technology to automotive security vulnerability mining is bound to become an important research direction in the process of improving the security of the Internet of Vehicles.

[0004] The development history of fuzz testing abroad is only ten years. In 2015, Matsumoto of Japan proposed a fuzz testing method for CAN (Controller Area Network) [6]. In 2017, Philip Lapczynski et al. of the United States proposed an effective automotive fuzz testing method [7]. Its conceptual framework used the existing automotive testing framework as the user front end, and also handled the vehicle-specific operation process and communication tasks. In 2019, DSFowler proposed a fuzz testing optimization technology applied to the CAN bus, which improved the automation level and efficiency of fuzz testing and avoided the impact of data explosion on fuzz testing of vehicle networking systems. In recent years, the optimization method of vulnerability detection in software using fuzz testing has mainly focused on the selection method. Fatmah YousefAssiri and Asia OthmanAljahdali studied the discovery of optimized mutation selection, which can improve code coverage and discover more vulnerabilities [9]. JohannesWillbold et al. proposed the use of automated fuzz testing technology to expand the analysis of satellite software without the participation of human experts. The article pointed out that re-hosting is the most promising method

[10] .

[0005] Fuzz testing of vehicle networks started relatively late in China. In 2016, Han Weihui proposed an Internet car safety bus model. Aiming at the security risks of CAN bus, his model used identity authentication, risk rating and bus protection methods to solve the problem of car bus security. Zhou Yunpeng et al. proposed a fuzzy test case generation method based on deep convolutional generative adversarial network. By introducing convolutional attention mechanism to improve DCGAN, the feature expression of the model and the format correctness of the generated CAN message were improved. Du Xiaojun et al. studied multi-population genetic algorithms in software fuzz testing, automatically generated deformed test data for fuzz testing, improved test efficiency, and provided a basic basis for improving software security. In order to solve the problems of efficiency and test intensity of software vulnerability fuzz testing process, Tang Chenghua et al. proposed a key branch exploration and hotspot update algorithm for software vulnerability fuzz testing. The experimental results show that the proposed method has a significant improvement in vulnerability discovery ability compared with the traditional random divergence fuzz testing method.

[0006] Large Language Models (LLMs) are considered to be one of the important paths towards artificial general intelligence (AGI). With the introduction of the BERT and GPT series models, the LLMs framework is mainly divided into two different technical paradigms: pre-training + fine-tuning paradigm and pre-training + prompting paradigm. Currently, researchers and engineers are exploring how to integrate LLMs into fuzz testing tools to improve test efficiency and coverage.

[0007] Junjie Wang et al. mainly discussed the use of large pre-trained language models (LLMs) in the field of software testing. LLMs show potential in software testing, especially in unit test case generation and program repair. Linghan Huang et al. analyzed the role of LLMs-based fuzz testers, common uses of which include enhancing the performance of fuzz testing through prompt engineering and seed mutation. Compared with traditional fuzz testers, LLMs-based fuzz testers show great potential in API and code coverage, discovering more complex errors, and improving fuzz testing automation. Chunqiu StevenXia et al. adopted a new large model-based fuzz testing tool Fuzz4All, which uses large language models (LLMs) to generate fuzz testing inputs for multiple programming languages. Fuzz4All supports targeted fuzz testing, allowing users to specify specific functions to be tested, thereby generating fuzz testing inputs for these functions.

[0008] There are more and more research means and methods for security threats to intelligent connected vehicles, but most of the existing research focuses on the security assessment of a single communication network or specific electronic devices, lacking a comprehensive study of the security of the entire vehicle network architecture and in-depth exploration of vulnerabilities in on-board functional software.

[0009] The future challenges and research hotspots of fuzz testing technology mainly include the following aspects: improving the automation level of data format preprocessing, developing a universal, convenient and highly automated preprocessing tool suite; improving the level of result analysis, including crash deduplication, vulnerability verification, vulnerability cause analysis, etc.; forming a universal fuzz testing tool integration platform, improving the versatility and portability of fuzz testing tools, etc.

[0010] Through the above analysis, the problems and defects of the prior art are as follows:

[0011] (1) Current automotive software fuzz testing has problems such as complex fuzz testing data protocols, single fuzz testing cases, and low efficiency of testing tools.

[0012] (2) During the vehicle-level automotive network fuzzy testing process, there are problems such as large amount of communication data of the object being tested, low test efficiency, and unclear directionality. Summary of the invention

[0013] In view of the problems existing in the prior art, the present invention provides a large model-based automobile information security fuzzy testing method.

[0014] The present invention is implemented in this way: a large model-based automobile information security fuzzy testing method comprises:

[0015] Step 1: Build a training corpus for the large model of automotive fuzz testing;

[0016] Construct an automotive software information security fuzz testing corpus for automotive components that includes multiple input and output forms, multiple test cases, multiple test data, and multiple test interfaces, which is used to train a large model targeting the characteristics of automotive software fuzz testing scenarios;

[0017] Step 2: Optimization of large model for automotive software fuzz testing based on LLMs;

[0018] Based on the LLMs base model, we study large model fine-tuning technology, prompt engineering technology, and retrieval enhancement technology parameter optimization technology and non-parameter optimization technology, focusing on exploring the optimization effect of LLMs enhanced body model in vehicle-level fuzz testing scenarios, and building a large model generalization system for automotive software fuzz testing;

[0019] Step 3: Develop an automotive information security fuzz testing tool based on LLMs;

[0020] By utilizing the large model training corpus and large model optimization technology in the aforementioned research, combined with the advantages of large models in automatic code generation, the ability of fuzz testing tools to intelligently generate test cases, test data, and test interfaces is improved, and the development of automotive information security fuzz testing tools based on LLMs is completed.

[0021] Further, the method for constructing a training corpus for a large model of automobile fuzzy testing is as follows:

[0022] First, a training corpus for a large automotive fuzz testing model is built based on meta-learning. The autonomous driving functional software samples messages from the automotive sensor system, positioning system, and visual system under multiple communication protocols by applying physical probes or white-box software instructions and source code information provided by suppliers. The messages are preprocessed based on the communication media, application protocol, and data format standardization to obtain a basic corpus describing the automotive functional software.

[0023] The basic corpus is converted using one-hot encoding and expanded to a trainable Euclidean space. The fuzzy processing of the basic corpus and small sample supervised learning are realized by using a meta-learning algorithm based on a memory pool and model irrelevance and a GAN network, solving the problems of sparse initial corpus data sets and rapid expansion of effective data.

[0024] Then, the constructed model training corpus is used to train the fuzzy testing large model, and its fuzzy test data intelligent generation capability is trained based on the open source base model; through multiple rounds of iterations of the training corpus and multiple rounds of training iterations of the large model, based on the LLMs prototype, the ability to process and generate fuzzy data in multiple data formats, multiple communication protocols, and multiple test objects is learned and acquired; after fine-tuning the base model, an LLMs enhanced body is formed with an intelligent fuzzy test data generation engine. There is no need for testers to have an in-depth understanding of the protocols and interfaces of the system under test. By collecting data and providing corresponding documents, the LLMs enhanced body can autonomously identify data formats and communication protocol data characteristics, and intelligently generate fuzzy test data for automotive software functional information security fuzz testing in the current scenario.

[0025] Furthermore, the LLMs-based large model optimization method for automotive software fuzz testing is as follows:

[0026] First, we study the large model parameter update optimization technology based on fine-tuning; improve the anti-interference degree of the test model by adjusting the model optimization solution algorithm, learning rate and regularization parameters; and use specific function calls, error handling mechanisms, prompt guidance, and security check content to form a fuzzy test model monitor;

[0027] Secondly, carry out large-model prompt engineering optimization technology based on the vehicle network structure knowledge graph; use the entities and entity attributes in the vehicle electronic and electrical architecture network, including all entities in the vehicle network including gateways, domain controllers, and ECUs in the domain, ECU functions, domain control ranges, and domain control security level entity attributes to establish a vehicle network structure knowledge graph; combine prompt engineering technology to improve the fuzzy test large model's ability to understand the vehicle network structure, and generate more targeted fuzzy test data, fuzzy test cases, and fuzzy test paths;

[0028] Then, the large model non-parametric update optimization technology based on the retrieval enhanced generation method is studied; a knowledge base of information security standards for the automotive industry is constructed, and according to the "Technical Requirements for Vehicle Information Security GB 44495-2024" automotive information security-related national standards, group standards and international standards, public test libraries, and commonly used communication protocols, security protocols, and data format contents in the vehicle network, the retrieval and generalization capabilities of large models are improved; clue learning of the automotive standard corpus is performed based on the data in the existing corpus.

[0029] Further, the method for developing an automobile information security fuzz testing tool based on LLMs is as follows:

[0030] Generate fuzz testing tools for automotive software based on the effective test data sets and test paths in the above links, including test data sets, test cases, test interfaces, and test code automatic generation content; Based on the universal test template, use the large language model technology to generate test cases for specific test scenarios, and complete the API interface protocol design of the fuzz testing software according to different application scenarios and communication requirements;

[0031] Finally, the test cases, interface protocols, and test data are visually packaged, and specific test requirements and data input and output formats are set to complete the automatic generation of test code. The established test scenarios are tested by launching test cases, and the test process data is recorded and the test report is generated. The report content includes test cases, test data, test interfaces, test codes, attack paths, attack times, and abnormal response behaviors.

[0032] Another object of the present invention is to provide a large model-based automobile information security fuzzy testing system comprising:

[0033] A construction module is used to build a training corpus for a large model of automotive fuzz testing; a corpus of automotive software information security fuzz testing that contains multiple input and output forms, multiple test cases, multiple test data, and multiple test interfaces at the automotive component level is built to train a large model that is tailored to the characteristics of automotive software fuzz testing scenarios;

[0034] Optimization module, used for large model optimization of automotive software fuzz testing based on LLMs; based on the LLMs base model, research on large model fine-tuning technology, prompt engineering technology, retrieval enhancement technology, parameter optimization technology, and non-parameter optimization technology, focusing on exploring the optimization effect of LLMs enhanced body model in vehicle-level fuzz testing scenarios, and building a large model generalization system for automotive software fuzz testing;

[0035] Development module for developing automotive information security fuzz testing tools based on LLMs; using the large model training corpus and large model optimization technology in the aforementioned research, combined with the advantages of large models in automatic code generation, improve the ability of fuzz testing tools to intelligently generate test cases, test data and test interfaces, and complete the development of automotive information security fuzz testing tools based on LLMs.

[0036] Another object of the present invention is to provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the large model-based automobile information security fuzzy testing method.

[0037] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to perform the steps of the large model-based automobile information security fuzzy testing method.

[0038] Another object of the present invention is to provide an information data processing terminal, which is used to implement the large model-based automobile information security fuzzy testing system.

[0039] In combination with the above technical solutions and the technical problems solved, please analyze the advantages and positive effects of the technical solutions to be protected by the present invention from the following aspects:

[0040] First, in view of the technical problems existing in the above-mentioned prior art and the difficulty of solving the problems, the technical solutions to be protected by the present invention and the results and data during the research and development process are closely combined to analyze in detail and deeply how the technical solutions of the present invention solve the technical problems, and some creative technical effects brought about after solving the problems. The specific description is as follows:

[0041] The present invention combines LLMs, related optimization search algorithms and prompt engineering, and uses fuzz testing to mine vulnerabilities in the vehicle network architecture, thereby discovering unnoticed security defects in the vehicle software. This interdisciplinary approach improves the automation and intelligence of the testing process by applying artificial intelligence technology to automotive fuzz testing; creates a test corpus specifically for automotive parts to provide data support for training high-efficiency AI models; optimizes the application effect of AI models in automotive software testing through fine-tuning, prompt engineering, and retrieval enhancement technologies; develops an intelligent testing tool based on LLMs to improve the ability to automatically generate test cases, data, and interfaces; and improves the ability to discover potential vulnerabilities in automotive software through more effective fuzz testing tools, thereby enhancing the security of automotive software. It helps to improve the reliability of intelligent connected vehicles at the level of electronic and electrical architecture and the security at the level of on-board functional software, which is of great significance to improving the safety level of intelligent connected vehicles.

[0042] (1) The main technical indicators completed by this invention are: developing a fuzzy test data set for the whole vehicle network, developing a fuzzy test large model software, developing a whole vehicle network fuzzy test tool that can intelligently generate test cases, test data and test interfaces, etc., completing 3 SCI / EI papers, applying for 2 patents, 3 software copyrights, and 1 test report and description document. The technical achievement indicators are specific and clear, and have direct application and promotion value.

[0043] (2) Taking advantage of the advantages of artificial intelligence technology in data processing, natural language interaction, and generalization capabilities, we will conduct research on the optimization application of LLMs in automotive software fuzz testing scenarios. We will optimize the concealment and effectiveness of the test sets generated by fuzz testing and improve the efficiency of the fuzz testing process.

[0044] (3) Through the implementation of the present invention, the relevant key technologies will be applied in the fuzzy testing of automobiles, and will be radiated to other related products and enterprises in China through the writing of technical standards, serving as an application demonstration and model for fuzzy testing of intelligent driving vehicles.

[0045] (4) According to market analysis, the present invention can perform fuzzy testing calculations on 100,000 cars each year. By improving data security and system security, it can reduce property losses caused by network security issues by hundreds of millions of yuan each year, which will not only make a positive contribution to national network security, but will also greatly promote the popularization of smart driving cars.

[0046] (5) Carry out basic theoretical research and engineering research on the training corpus of large models for automotive fuzz testing, optimization technology of large models for automotive software fuzz testing based on LLMs, and evaluation and application verification methods for automotive fuzz testing based on LLMs, and form a number of relevant theories and results.

[0047] Second, 1. Existing technical problems solved by the present invention

[0048] In the current field of automotive information security, there are many technical problems in the industrial application of fuzz testing technology. Traditional fuzz testing methods rely on manually designed test cases, which are inefficient, have limited coverage, and are difficult to adapt to the diverse communication protocols and complex interaction scenarios in automotive software. In addition, most existing fuzz testing tools are general-purpose tools that lack optimization capabilities for specific automotive needs, making it difficult to achieve efficient, safe, and comprehensive vulnerability detection in actual industrial applications. Therefore, industrial applications urgently need an intelligent, automated, and highly adaptable fuzz testing method and tool for automotive scenarios to improve testing efficiency and security assurance capabilities.

[0049] 2. Methods of the present invention for solving the problems of the prior art

[0050] The present invention fundamentally solves the problems of low efficiency, insufficient coverage and poor adaptability in traditional fuzz testing technology by introducing a fuzz testing method based on large language models (LLMs). By constructing a dedicated training corpus for automotive information security scenarios, the generalization ability and semantic understanding ability of the large model are optimized, and high-quality test cases, test data and test interface definitions are automatically generated. The present invention realizes the intelligence of the entire process from test scenario identification, case generation to result analysis, avoids the limitations of relying on manually designed test plans, and improves the ability to discover complex software vulnerabilities in industrial applications.

[0051] 3. Technological progress achieved by the invention in industrial applications

[0052] Compared with the prior art, the present invention has achieved significant technological progress in industrial applications. First, by automatically generating comprehensive test cases and data, the efficiency and coverage of fuzz testing are improved. Secondly, the tool of the present invention can dynamically adapt to different vehicle configurations and multi-control unit (ECU) collaboration scenarios, significantly enhancing the applicability of fuzz testing tools in vehicle-level testing. In addition, through retrieval enhancement and prompt engineering optimization, the tool of the present invention can quickly adapt to the latest security threats and attack scenarios, ensuring the cutting-edge and efficient nature of fuzz testing technology in industrial applications.

[0053] 4. The value and driving force of this invention in industrial applications

[0054] The present invention provides an intelligent and efficient fuzz testing solution for the field of automotive information security in industrial applications, which not only significantly reduces the testing cost, but also improves the accuracy and reliability of the test results. It has shown excellent industrial application value in vehicle-level security verification, software vulnerability mining, and security mechanism enhancement. The wide application of the present invention will promote the popularization of fuzz testing technology in the automotive industry, provide important support for improving the information security protection capabilities of automotive software, and help the safe development of intelligent connected vehicles and autonomous driving technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flow chart of the automobile information security fuzzy testing method based on a large model provided in an embodiment of the present invention.

[0056] Figure 2 It is a structural block diagram of a large model-based automobile information security fuzzy testing system provided by an embodiment of the present invention.

[0057] Figure 3 It is a diagram of the research contents and their logical relationships provided by the embodiments of the present invention.

[0058] Figure 4 This is a research diagram of the large model corpus meta-learning technology for fuzzy testing of component-level automotive software provided by an embodiment of the present invention.

[0059] Figure 5 This is a research diagram of the multi-mode optimization technology of the fuzzy test large model for the whole vehicle level automotive network provided by the embodiment of the present invention.

[0060] Figure 6 It is a development diagram of an automotive software fuzz testing software tool based on LLM provided in an embodiment of the present invention.

[0061] Figure 7 This is a diagram of the awards given to national scientific research platforms and project teams according to an embodiment of the present invention.

[0062] Figure 8 This is a diagram of the information security attack and defense HIL test platform for intelligent connected vehicles provided by an embodiment of the present invention.

[0063] Fig. 9 This is a diagram of the fitting results of the CAN network pseudo-true data set provided by an embodiment of the present invention.

[0064] Fig.10 This is a diagram of cloud-embedded communication data processing and analysis results provided by an embodiment of the present invention.

[0065] Fig.11 It is an ECU device mapping and tampering attack detection model diagram based on hybrid feature extraction provided by an embodiment of the present invention.

[0066] Fig.12 This is a diagram of a whole-vehicle-level test platform for intelligent connected vehicles provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0067] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0068] like Figure 1 As shown, a large model-based automobile information security fuzzy testing method provided by an embodiment of the present invention includes the following steps:

[0069] S101, build a training corpus for the large model of automotive fuzz testing;

[0070] Construct an automotive software information security fuzz testing corpus for automotive components that includes multiple input and output forms, multiple test cases, multiple test data, and multiple test interfaces, which is used to train a large model targeting the characteristics of automotive software fuzz testing scenarios;

[0071] S102, Large Model Optimization for Automotive Software Fuzz Testing Based on LLMs;

[0072] Based on the LLMs base model, we study large model fine-tuning technology, prompt engineering technology, and retrieval enhancement technology parameter optimization technology and non-parameter optimization technology, focusing on exploring the optimization effect of LLMs enhanced body model in vehicle-level fuzz testing scenarios, and building a large model generalization system for automotive software fuzz testing;

[0073] S103, develop a fuzz testing tool for automotive information security based on LLMs;

[0074] By utilizing the large model training corpus and large model optimization technology in the aforementioned research, combined with the advantages of large models in automatic code generation, the ability of fuzz testing tools to intelligently generate test cases, test data, and test interfaces is improved, and the development of automotive information security fuzz testing tools based on LLMs is completed.

[0075] The method for constructing a training corpus for a large automobile fuzzy testing model provided by an embodiment of the present invention is as follows:

[0076] First, a training corpus for a large automotive fuzz testing model is built based on meta-learning. The autonomous driving functional software samples messages from the automotive sensor system, positioning system, and visual system under multiple communication protocols by applying physical probes or white-box software instructions and source code information provided by suppliers. The messages are preprocessed based on the communication media, application protocol, and data format standardization to obtain a basic corpus describing the automotive functional software.

[0077] The basic corpus is converted using one-hot encoding and expanded to a trainable Euclidean space. The fuzzy processing of the basic corpus and small sample supervised learning are realized by using a meta-learning algorithm based on a memory pool and model irrelevance and a GAN network, solving the problems of sparse initial corpus data sets and rapid expansion of effective data.

[0078] Then, the constructed model training corpus is used to train the fuzzy testing large model, and its fuzzy test data intelligent generation capability is trained based on the open source base model; through multiple rounds of iterations of the training corpus and multiple rounds of training iterations of the large model, based on the LLMs prototype, the ability to process and generate fuzzy data in multiple data formats, multiple communication protocols, and multiple test objects is learned and acquired; after fine-tuning the base model, an LLMs enhanced body is formed with an intelligent fuzzy test data generation engine. There is no need for testers to have an in-depth understanding of the protocols and interfaces of the system under test. By collecting data and providing corresponding documents, the LLMs enhanced body can autonomously identify data formats and communication protocol data characteristics, and intelligently generate fuzzy test data for automotive software functional information security fuzz testing in the current scenario.

[0079] The LLMs-based large model optimization method for automobile software fuzzy testing provided by the embodiment of the present invention is as follows:

[0080] First, we study the large model parameter update optimization technology based on fine-tuning; improve the anti-interference degree of the test model by adjusting the model optimization solution algorithm, learning rate and regularization parameters; and use specific function calls, error handling mechanisms, prompt guidance, and security check content to form a fuzzy test model monitor;

[0081] Secondly, carry out large-model prompt engineering optimization technology based on the vehicle network structure knowledge graph; use the entities and entity attributes in the vehicle electronic and electrical architecture network, including all entities in the vehicle network including gateways, domain controllers, and ECUs in the domain, ECU functions, domain control ranges, and domain control security level entity attributes to establish a vehicle network structure knowledge graph; combine prompt engineering technology to improve the fuzzy test large model's ability to understand the vehicle network structure, and generate more targeted fuzzy test data, fuzzy test cases, and fuzzy test paths;

[0082] Then, the large model non-parametric update optimization technology based on the retrieval enhanced generation method is studied; a knowledge base of information security standards for the automotive industry is constructed, and according to the "Technical Requirements for Vehicle Information Security GB 44495-2024" automotive information security-related national standards, group standards and international standards, public test libraries, and commonly used communication protocols, security protocols, and data format contents in the vehicle network, the retrieval and generalization capabilities of large models are improved; clue learning of the automotive standard corpus is performed based on the data in the existing corpus.

[0083] The method for developing an automobile information security fuzzy testing tool based on LLMs provided in an embodiment of the present invention is as follows:

[0084] Generate fuzz testing tools for automotive software based on the effective test data sets and test paths in the above links, including test data sets, test cases, test interfaces, and test code automatic generation content; Based on the universal test template, use the large language model technology to generate test cases for specific test scenarios, and complete the API interface protocol design of the fuzz testing software according to different application scenarios and communication requirements;

[0085] Finally, the test cases, interface protocols, and test data are visually packaged, and specific test requirements and data input and output formats are set to complete the automatic generation of test code. The established test scenarios are tested by launching test cases, and the test process data is recorded and the test report is generated. The report content includes test cases, test data, test interfaces, test codes, attack paths, attack times, and abnormal response behaviors.

[0086] like Figure 2 As shown, an automobile information security fuzzy testing system based on a large model provided by an embodiment of the present invention includes:

[0087] A construction module is used to build a training corpus for a large model of automotive fuzz testing; a corpus of automotive software information security fuzz testing that contains multiple input and output forms, multiple test cases, multiple test data, and multiple test interfaces at the automotive component level is built to train a large model that is tailored to the characteristics of automotive software fuzz testing scenarios;

[0088] Optimization module, used for large model optimization of automotive software fuzz testing based on LLMs; based on the LLMs base model, research on large model fine-tuning technology, prompt engineering technology, retrieval enhancement technology, parameter optimization technology, and non-parameter optimization technology, focusing on exploring the optimization effect of LLMs enhanced body model in vehicle-level fuzz testing scenarios, and building a large model generalization system for automotive software fuzz testing;

[0089] Development module for developing automotive information security fuzz testing tools based on LLMs; using the large model training corpus and large model optimization technology in the aforementioned research, combined with the advantages of large models in automatic code generation, improve the ability of fuzz testing tools to intelligently generate test cases, test data and test interfaces, and complete the development of automotive information security fuzz testing tools based on LLMs.

[0090] Another object of the present invention is to provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the large model-based automobile information security fuzzy testing method.

[0091] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to perform the steps of the large model-based automobile information security fuzzy testing method.

[0092] Another object of the present invention is to provide an information data processing terminal, which is used to implement the large model-based automobile information security fuzzy testing system.

[0093] The present invention is specifically implemented:

[0094] Aiming at the problems existing in the current automotive software fuzz testing, such as complicated fuzz testing data protocols, single fuzz testing cases, and low test tool efficiency, this project intends to carry out research on automotive information security fuzz testing tools based on LLMs to improve the intelligence of automotive information security fuzz testing tools. The various research contents are logically related as follows: Figure 3 shown.

[0095] (1) Research on Meta-Learning Technology of Large Model Corpus for Fuzz Testing of Component-Level Automotive Software

[0096] In order to address the problems in the component-level automotive functional software testing process, such as high degree of public uncertainty in the test software, a large variety of components, and complex input and output test data communication protocols, this section conducts research on building a self-updating fuzzy testing large model corpus based on meta-learning.

[0097] First, a training corpus for a large model of automotive fuzz testing is constructed based on meta-learning. Taking autonomous driving functional software as an example, by applying physical probes or white box software instructions, source code and other information provided by suppliers, the message messages of the automotive sensor system, positioning system and visual system under various communication protocols are sampled, and standardized preprocessing such as communication media, application protocols and data formats is performed to obtain a basic corpus describing the functional software of the automobile. Furthermore, unique hot encoding is used to convert the data of the basic corpus and expand it to a trainable Euclidean space. The meta-learning algorithm based on memory pool and model irrelevance and the GAN network are used to realize the fuzzification processing of the basic corpus and small sample supervised learning, solving the problem of sparse initial corpus data set and rapid expansion of effective data.

[0098] Then, the constructed model training corpus is used to train the fuzz test big model, and its fuzz test data intelligent generation capability is trained based on the open source base model. Through multiple rounds of iterations of the training corpus and multiple rounds of training iterations of the big model, based on the LLMs prototype, the ability to process and generate fuzzy data of multiple data formats, multiple communication protocols, and multiple test objects is learned and acquired. After fine-tuning the base model, an LLMs enhanced body with an intelligent fuzz test data generation engine is formed. There is no need for testers to have an in-depth understanding of the protocols and interfaces of the system under test. By collecting data and providing corresponding documents, the LLMs enhanced body autonomously identifies data features such as data formats and communication protocols, and intelligently generates fuzzy test data for automotive software functional information security fuzz testing in the current scenario.

[0099] Figure 4 Research content: Research on meta-learning technology of large model corpus for fuzz testing of component-level automotive software

[0100] (2) Research on multi-mode optimization technology of large fuzzy test models for vehicle-level automotive networks

[0101] In order to solve the problems of large communication data volume, low test efficiency and unclear directionality of the object under test, etc. during the vehicle-level automotive network fuzz testing, this section conducts research on multi-mode optimization technology of large fuzzy test models for vehicle-level automotive networks.

[0102] Figure 5 Research content 2: Research on multi-mode optimization technology of large models for fuzzy testing of vehicle-level automotive networks

[0103] First, we study the parameter update optimization technology of large models based on fine-tuning. By adjusting the model optimization solution algorithm, learning rate and regularization parameters, we can improve the anti-interference degree of the test model, reduce the risk of overfitting of the model, and improve the training accuracy of large models. We use specific function calls, error handling mechanisms, prompt guidance, safety checks and other contents to form a fuzzy test model monitor, improve the attention mechanism and key information extraction ability of the large model, and avoid the problem that the fine-tuned large model is over-adapted to the fuzzy test field and causes catastrophic forgetting.

[0104] Secondly, develop large-model prompt engineering optimization technology based on the knowledge graph of the vehicle network structure. Utilize the entities and entity attributes in the vehicle's electronic and electrical architecture network, such as all entities in the vehicle network including gateways, domain controllers, and ECUs within the domain, as well as entity attributes such as ECU functions, domain control ranges, and domain control security levels, to establish a knowledge graph of the vehicle network structure. Combined with prompt engineering technology, the fuzzy test large model's ability to understand the vehicle network structure is improved, and more targeted fuzzy test data, fuzzy test cases, and fuzzy test paths are generated to quickly locate vulnerability locations and risk levels, accelerating the fuzzy testing process for the vehicle network.

[0105] Then, the non-parametric update optimization technology of large models based on the retrieval enhancement generation method is studied. The automotive industry information security standard knowledge base is constructed. According to the "Technical Requirements for Vehicle Information Security GB 44495-2024" and other national standards, group standards and international standards related to automobile information security, public test libraries, common communication protocols, security protocols, data formats, etc. in the vehicle network, the retrieval and generalization capabilities of large models are improved. Clue learning is performed on the automotive standard corpus based on the data in the existing corpus, and the approximate nearest neighbor search algorithm is used for search to realize the feedback ability of the large language model on the fuzzy test results of vehicle information security.

[0106] (3) Development of automotive software fuzz testing software tools based on LLM

[0107] Based on the effective test data sets and test paths in the above links, a fuzz testing tool for automotive software is generated, including test data sets, test cases, test interfaces, and automatic test code generation. Based on the universal test template, the large language model technology is used to generate test cases for specific test scenarios, and the API interface protocol design of the fuzz testing software is completed according to different application scenarios and communication requirements.

[0108] Finally, the test cases, interface protocols, and test data are visually packaged, and specific test requirements and data input and output formats are set to complete the automatic generation of test code. By launching the test cases to test the established test scenarios, the test process data is recorded and the test report generation is completed. The report content includes test cases, test data, test interfaces, test codes, attack paths, attack times, and abnormal response behaviors.

[0109] Figure 6 Research content 3: Development of automotive software fuzz testing software tool based on LLM

[0110] The unmanned vehicle intelligent control team of Beijing Institute of Technology, where the project applicant is located, relies on national and provincial platforms such as the National Engineering Research Center for Electric Vehicles, the National Key Laboratory of High-end Vehicle Integration and Control, the Key Laboratory of Unmanned Vehicle Technology of the Ministry of Industry and Information Technology, and the Beijing Key Laboratory of Clean Vehicles. In recent years, the research results of the applicant's mechanical engineering (vehicle engineering) field have had an important impact at home and abroad. It is a national key discipline / double first-class discipline and has an important impact in the field of vehicle network security and unmanned driving in China. In recent years, the research team of the project applicant has won a number of national / provincial and ministerial scientific and technological achievements awards around automotive heterogeneous network security and intrusion detection technology, and has strong scientific research capabilities. In summary, the high-level R&D team of the applicant can provide important scientific research platform support for the development of this project.

[0111] Figure 7 Awards for national scientific research platforms and project teams

[0112] The research team also has a hardware HIL test platform for intelligent connected vehicles, such as Figure 8 The test platform has a unified hardware interface, network data flow software, intelligent driving simulation software, data acquisition card, and self-developed domain controller. The unified hardware interface can realize data interaction in heterogeneous networks.

[0113] Providing necessary data support for network analysis tools, CANoe (VN5640) and other devices support the development of security protocols for vehicle Ethernet, the security attack and defense test platform develops attack and defense confrontation functions for vehicle information security, and conducts on-site testing of the real-time communication and network conditions of the vehicle. The domain controller can use the code debugging platform to carry out embedded development of the developed security algorithm, provide engineering software for real vehicle applications, and verify the vehicle-level matching performance with the help of simulated driving software. The system supports the installation of SOA systems for embedded service development, etc. In addition, the project research team has also built a high-performance parallel computing server (main frequency 3GHz, 36 cores and 72 threads, running memory 384GB), which supports automobile message information analysis and data offline analysis, as well as vehicle-cloud data interaction and hardware-in-the-loop testing of vehicle-cloud dynamic defense systems. The above conditions provide a powerful computing platform and software guarantee for the verification of automotive network security strategies.

[0114] Project Research Basis

[0115] In recent years, the project application team has conducted in-depth research on intrusion detection, secure communication protocols, deep learning frameworks, and vehicle-cloud linkage defense in heterogeneous networks of intelligent driving vehicles, focusing on deep intrusion detection technology based on message spatiotemporal feature distribution, abnormal traffic monitoring technology based on information entropy theory, ECU device mapping technology based on hybrid feature extraction, and linkage defense technology based on vehicle-cloud interaction and deep AI heterogeneous detection models. In addition, the team has published many high-level academic papers in journals such as IEEE T-ITS and IEEESensers Journal, and has won many achievement conversion and scientific and technological awards. The specific situation is as follows:

[0116] (1) Research on the method of generating abnormal threat test dataset based on self-supervised learning theory

[0117] The applicant used the method of random noise injection into the communication ID to fit part of the CAN network pseudo-true data set. The fitting characteristics are as follows: Fig. 9As shown in (a), it can be found that as the noise ratio increases, the discreteness of the pseudo-true data set will not be distorted, indicating that the generated data has a strong pseudo-true ability, that is, it will not be easily recognized by the detection system, which is conducive to improving the accuracy of the detection algorithm. In addition, a certain CAN attack test was performed on the pseudo-true data. By inserting pseudo-true data into normal data, attack data with a high degree of fitting was obtained, deceiving the car ECU to read the abnormal data, such as Fig. 9 (b) as shown.

[0118] (2) Research on active defense mechanism against vehicle operation attacks based on human-computer interaction.

[0119] The applicant has deployed research work such as driver operation prediction and vehicle instruction rule library based on cloud platform. The results obtained by the mathematical model of attack tree and attack detection method are as follows Fig.10 The project applicant has realized the construction and offline training of the driver operation prediction reinforcement learning model. In the subsequent work, it is necessary to continue to explore the behavioral differences of different drivers to improve the general performance of the prediction model.

[0120] The vehicle-side processing technologies used (such as wavelet packet analysis, cepstrum analysis, vehicle-side data upload protocol, etc.) were tested and analyzed on an embedded device (TMS320F28377D, TI), where the processing delay of the key detection algorithm did not exceed 5ms and the memory usage did not exceed 4%.

[0121] (3) Research on ECU device mapping and tampering attack detection technology based on hybrid feature extraction

[0122] Aiming at the problems of unstable physical feature extraction, long detection engine consumption and low detection accuracy in the tampering attack intrusion detection algorithm of intelligent connected vehicles, an ECU source mapping technology based on physical hybrid feature extraction and a tampering attack detection model are proposed. Fig.11 As shown. This solution deeply analyzes the performance characteristics of the voltage bit time, platform time, edge jump time and high-level mode of the differential voltage signal of the bus message, and draws the view that different ECU hardware has unique performance characteristics. Based on the bus bit filling principle, the concept of "bit group" is proposed for the first time in the industry, and the performance characteristics of all messages are divided into 5 groups that can be used for training features of the learning model. A lightweight Softmax classifier is designed to classify different ECU messages, and the message sending source mapping and low-frequency tampering attacks are effectively detected. Compared with other top solutions in the industry (such as average bit time and timestamp ramp), the proposed research solution improves the ECU source mapping accuracy by more than 6%, and the average detection accuracy of tampering attacks is as high as 97.83%.

[0123] In addition, the project application team has built vehicle-level test platforms such as unmanned control of intelligent connected vehicles and heterogeneous network security. Fig.12 As shown in the figure, there are three test models, each of which is equipped with a complete unmanned driving system, gateway, chassis domain control system, network data packet real-time sampling and analysis tools, vehicle-cloud interaction system, high-performance parallel computing server and other core functions. Specifically, the vehicle's unmanned driving system adopts a domestically produced autonomous driving platform (AVA-3510, RQX-59H), which supports SLAM mapping, intelligent decision-making lane changing, and robust trajectory tracking control functions with multi-source sensor fusion; the gateway system supports protocol conversion and network security protection, remote program upgrades, and security defense systems for heterogeneous communication media such as on-board Gigabit Ethernet / CANFD / CAN / LIN; the chassis domain control system supports real-time communication and active safety control functions; the network data packet real-time analysis and collection tool samples and analyzes the vehicle's communication data, providing the necessary data source for the intrusion detection engine; the vehicle-cloud interaction system deploys the AI ​​engine's computing tasks in the cloud, and sets up a vehicle-side data upload interface and a policy receiving end configuration file to realize vehicle-cloud data interaction and vehicle-side security policy updates; the high-performance parallel computing server is equipped with NVIDIA's RTX 4090 GPU, which can accelerate the training and reasoning process of large models. The training server supports distributed training, and multiple servers can work together to train a model, significantly reducing the time required to train large models. The above conditions provide strong hardware equipment and platform support for the implementation of the research and development of automotive information security fuzz testing tools based on LLMs.

[0124] The following is a detailed working principle of a large model-based automotive information security fuzz testing method, which is described in six parts:

[0125] 1. Build a car fuzz testing training corpus

[0126] In the field of automotive information security, fuzz testing is a key testing technology. First, a multi-dimensional training corpus is constructed by collecting a large amount of corpus related to automotive software. The corpus covers the following contents:

[0127] Various input and output forms, such as control signals, communication protocols and diagnostic data.

[0128] Multiple test cases covering different attack scenarios and edge cases.

[0129] A variety of test data, including normal data, random data, and malicious data.

[0130] A variety of test interfaces, such as the communication interface of the vehicle control unit (ECU), the vehicle Ethernet interface, and the OTA update interface. This corpus will serve as the core data resource for subsequent training of large models, providing a strong knowledge base and generalization capability for fuzz testing.

[0131] 2. Construction and optimization of large automotive fuzz testing models based on LLMs

[0132] When building a large model, existing large language models (LLMs) are used as base models and optimized for the scenario requirements of automotive fuzz testing.

[0133] Fine-tuning technology: Through exclusive training on the automotive fuzz testing corpus, the model's ability to understand the input, output and test logic unique to automotive scenarios is optimized.

[0134] Prompt Engineering: Design effective prompt templates to enhance the accuracy and diversity of model-generated test cases, data, and interfaces.

[0135] Retrieval enhancement technology: The integrated retrieval module enables the model to dynamically query and utilize external test data sources to improve the knowledge breadth of the model.

[0136] Non-parametric optimization technology: By optimizing the model architecture parameters, the generalization ability and efficiency of the model in vehicle-level fuzzy testing scenarios are enhanced.

[0137] Finally, a large exclusive model suitable for automotive information security fuzz testing is constructed.

[0138] 3. Automation capabilities of fuzz testing tool development

[0139] Using the trained and optimized large model, we developed a set of automotive fuzz testing tools based on LLMs to automatically improve the following capabilities:

[0140] Test case generation: Based on the big model’s understanding of automotive scenarios, test cases covering a variety of attack scenarios are automatically generated, including communication anomalies, signal tampering, and boundary value testing.

[0141] Test data generation: Generate various forms of test data through large models, including legitimate normal data and abnormal malicious data, to comprehensively detect vulnerabilities in automotive software.

[0142] Test interface identification: Combined with the semantic understanding capabilities of the large model, it can quickly locate the test interface of the automotive software and generate test logic that matches the interface.

[0143] These capabilities significantly reduce the workload of testers while improving the efficiency and quality of fuzz testing.

[0144] 4. Generalization and adaptation of vehicle-level fuzz testing scenarios

[0145] The optimization and tool development of large models are specially adapted for vehicle-level fuzz testing scenarios. Vehicle testing involves complex interactions between multiple control units (ECUs), and the diversity of data flows and communication protocols increases the complexity of testing.

[0146] Model generalization capability: Through training and fine-tuning techniques, ensure that the large model can handle the testing requirements of different vehicle systems (such as powertrain, body control, and entertainment systems).

[0147] Scenario adaptation: The large model can dynamically generate test cases adapted to different vehicle configurations, covering complex scenarios from a single ECU to multi-ECU collaboration.

[0148] Risk assessment: The tool can automatically analyze the test results based on the big model, evaluate the safety risks of the vehicle software, and provide improvement suggestions.

[0149] 5. Dynamic learning and continuous optimization mechanism

[0150] To adapt to the ever-changing threats to automotive information security, large models and tools have the ability to dynamically learn and continuously optimize.

[0151] Dynamically update the corpus: By automatically collecting and integrating new automotive vulnerability information, attack cases, and communication protocols, the corpus is updated regularly to keep the knowledge of the big model up to date.

[0152] Feedback-driven optimization: Feedback the results accumulated in actual testing into model training, and improve the performance of the model in new scenarios through continuous optimization and iteration.

[0153] Threat detection evolution: The tool can dynamically adjust fuzz testing strategies based on the latest test results and attack trends to maintain high sensitivity to new threats.

[0154] 6. Application scenarios and expected results

[0155] This method and tool can be widely used in different scenarios of automotive information security, such as component-level testing, vehicle-level testing, and OTA update security assessment.

[0156] Rapidly discover software vulnerabilities in a single ECU during component-level testing.

[0157] In vehicle-level testing, the interactive safety between multiple systems is comprehensively analyzed.

[0158] In OTA update testing, vulnerabilities and attack risks during the update process are verified.

[0159] Through large model-driven fuzz testing, the testing efficiency and coverage are significantly improved, while the testing costs are reduced, providing strong protection for the information security of automotive software.

[0160] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0161] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A large model-based automobile information security fuzzy testing method, characterized in that: The following steps are involved: Step 1: Build a training corpus for the large model of automotive fuzz testing; Construct an automotive software information security fuzz testing corpus for automotive components that includes multiple input and output forms, multiple test cases, multiple test data, and multiple test interfaces, which is used to train a large model targeting the characteristics of automotive software fuzz testing scenarios; Step 2: Optimization of large model for automotive software fuzz testing based on LLMs; Based on the LLMs base model, we study large model fine-tuning technology, prompt engineering technology, and retrieval enhancement technology parameter optimization technology and non-parameter optimization technology, focusing on exploring the optimization effect of LLMs enhanced body model in vehicle-level fuzz testing scenarios, and building a large model generalization system for automotive software fuzz testing; Step 3: Develop an automotive information security fuzz testing tool based on LLMs; By utilizing the large model training corpus and large model optimization technology in the aforementioned research, combined with the advantages of large models in automatic code generation, the ability of fuzz testing tools to intelligently generate test cases, test data, and test interfaces is improved, and the development of automotive information security fuzz testing tools based on LLMs is completed.

2. The automobile information security fuzzy testing method based on a large model as claimed in claim 1 is characterized in that: The method for constructing a training corpus for a large model of automobile fuzzy testing is as follows: First, a training corpus for a large automotive fuzz testing model is built based on meta-learning. The autonomous driving functional software samples messages from the automotive sensor system, positioning system, and visual system under multiple communication protocols by applying physical probes or white-box software instructions and source code information provided by suppliers. The messages are preprocessed based on the communication media, application protocol, and data format standardization to obtain a basic corpus describing the automotive functional software. The basic corpus is converted using one-hot encoding and expanded to a trainable Euclidean space. The fuzzy processing of the basic corpus and small sample supervised learning are realized by using a meta-learning algorithm based on a memory pool and model irrelevance and a GAN network, solving the problems of sparse initial corpus data sets and rapid expansion of effective data. Then, the constructed model training corpus is used to train the fuzzy testing large model, and its fuzzy test data intelligent generation capability is trained based on the open source base model; through multiple rounds of iterations of the training corpus and multiple rounds of training iterations of the large model, based on the LLMs prototype, the ability to process and generate fuzzy data in multiple data formats, multiple communication protocols, and multiple test objects is learned and acquired; after fine-tuning the base model, an LLMs enhanced body is formed with an intelligent fuzzy test data generation engine. There is no need for testers to have an in-depth understanding of the protocols and interfaces of the system under test. By collecting data and providing corresponding documents, the LLMs enhanced body can autonomously identify data formats and communication protocol data characteristics, and intelligently generate fuzzy test data for automotive software functional information security fuzz testing in the current scenario.

3. The automobile information security fuzzy testing method based on a large model as claimed in claim 1 is characterized in that: The LLMs-based large model optimization method for automotive software fuzz testing is as follows: First, we study the large model parameter update optimization technology based on fine-tuning; improve the anti-interference degree of the test model by adjusting the model optimization solution algorithm, learning rate and regularization parameters; and use specific function calls, error handling mechanisms, prompt guidance, and security check content to form a fuzzy test model monitor; Secondly, carry out large-model prompt engineering optimization technology based on the vehicle network structure knowledge graph; use the entities and entity attributes in the vehicle electronic and electrical architecture network, including all entities in the vehicle network including gateways, domain controllers, and ECUs in the domain, ECU functions, domain control ranges, and domain control security level entity attributes to establish a vehicle network structure knowledge graph; combine prompt engineering technology to improve the fuzzy test large model's ability to understand the vehicle network structure, and generate more targeted fuzzy test data, fuzzy test cases, and fuzzy test paths; Then, the present invention is based on the large-model non-parametric update optimization technology of the retrieval enhanced generation method; constructs an automobile industry information security standard knowledge base, according to the "Technical Requirements for Vehicle Information Security GB 44495-2024" automobile information security-related national standards, group standards and international standards, public test libraries, commonly used communication protocols, security protocols, and data format contents in the vehicle network; and performs clue learning on the automobile standard corpus based on the data in the existing corpus.

4. The automobile information security fuzzy testing method based on a large model as claimed in claim 1 is characterized in that: The method for developing an automobile information security fuzz testing tool based on LLMs is as follows: Generate a fuzz testing tool for automotive software based on the effective test data sets and test paths in the above links, including test data sets, test cases, test interfaces, and test code automatic generation content; Based on the universal test template, the large language model technology is used to generate test cases for specific test scenarios, and the API interface protocol design of the fuzz testing software is completed according to different application scenarios and communication requirements; Finally, the test cases, interface protocols, and test data are visually packaged, and specific test requirements and data input and output formats are set to complete the automatic generation of test code; By launching the test cases, the established test scenarios are tested, and at the same time, the test process data is recorded and the test report is generated. The report content includes test cases, test data, test interfaces, test codes, attack paths, attack times and abnormal response behaviors.

5. A large model-based automobile information security fuzzy testing system implementing the testing method according to any one of claims 1 to 4, characterized in that: The automobile information security fuzzy testing system based on a large model includes: A construction module is used to build a training corpus for a large model of automotive fuzz testing; a corpus of automotive software information security fuzz testing that contains multiple input and output forms, multiple test cases, multiple test data, and multiple test interfaces at the automotive component level is built to train a large model that is tailored to the characteristics of automotive software fuzz testing scenarios; Optimization module, used for large model optimization of automotive software fuzz testing based on LLMs; based on the LLMs base model, research on large model fine-tuning technology, prompt engineering technology, retrieval enhancement technology, parameter optimization technology, and non-parameter optimization technology, focusing on exploring the optimization effect of LLMs enhanced body model in vehicle-level fuzz testing scenarios, and building a large model generalization system for automotive software fuzz testing; Development module for developing automotive information security fuzz testing tools based on LLMs; using the large model training corpus and large model optimization technology in the aforementioned research, combined with the advantages of large models in automatic code generation, improve the ability of fuzz testing tools to intelligently generate test cases, test data and test interfaces, and complete the development of automotive information security fuzz testing tools based on LLMs.

6. A computer device, characterized in that: The computer device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the large model-based automobile information security fuzzy testing method as described in any one of claims 1-4.

7. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the large model-based automobile information security fuzzy testing method as described in any one of claims 1 to 4.

8. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the automobile information security fuzzy testing system based on a large model as described in claim 5.

Images