An autonomous vehicle safety vulnerability assessment method and system based on data features

Abstract

The application relates to a kind of autonomous vehicle safety vulnerability evaluation method and system based on data characteristics, the method first synchronously collects the real measurement data of the current time of autonomous vehicle and the historical measurement data of the last time of controller;Then according to the absolute value of data deviation and the weight coefficient of corresponding channel, the importance index of each measurement channel is calculated;Then the output norm of real measurement data is used to dynamically build double threshold, each channel is divided into low, medium and high three importance regions, and the optimal test channel is locked in descending order;Accordingly, the cooperative evaluation range of single channel or multiple channels is automatically decided, and the channel selection matrix is constructed;Finally, the historical data retention mechanism is used, and the historical data is used to replace the real data of the target channel to generate evaluation test data.The application realizes the accurate identification of key node data and the adaptive adjustment of test boundary, and reveals the security vulnerabilities of vehicle control system in a hidden and efficient manner without forging complex signals.

Description

Technical Field

[0001] This invention relates to the field of industrial control system automation monitoring and process safety technology, and in particular to a method and system for assessing the safety vulnerability of autonomous vehicles based on data characteristics. Background Technology

[0002] With the rapid development of intelligent connected vehicles, autonomous driving technology, and vehicle-road cooperative systems, autonomous vehicles have become an important component of modern intelligent transportation systems. Autonomous vehicles achieve functions such as environmental perception, decision-making and planning, and motion control through onboard sensors, controllers, actuators, and communication networks. Their operation is highly dependent on the information interaction between measurement data, communication networks, and control systems, thus exhibiting typical characteristics of a cyber-physical system (CPS).

[0003] Due to the deep integration of vehicle control systems and communication networks, autonomous vehicles face increasingly severe cybersecurity threats during operation. In particular, cyberattacks targeting the vehicle's perception and communication layers can lead to decreased vehicle control performance, and even cause vehicle instability, lane departure, or traffic accidents. Therefore, research on cyberattacks and security protection for autonomous vehicles has become an important research direction in the fields of intelligent transportation and networked control.

[0004] Currently, the main attack methods targeting autonomous vehicles and networked control systems include Denial-of-Service (DoS) attacks, False Data Injection (FDI) attacks, Replay attacks, and Deception attacks.

[0005] Denial-of-service attacks primarily disrupt communication channels or cause network congestion, preventing the controller from obtaining measurement data in a timely manner. False data injection attacks mislead the controller into making incorrect decisions by tampering with transmitted data. Replay attacks repeatedly send historical data instead of current data, thus affecting the normal operation of the system. Spoofing attacks mislead the control system by constructing false measurement information. All of these attack methods can reduce the control performance of autonomous vehicles to some extent.

[0006] In recent years, to improve attack efficiency and stealth, researchers have begun to focus on selective attack strategies. Unlike traditional random or global attacks, selective attacks analyze the importance of different measurement data in the system and prioritize attacks on key measurement channels that have a greater impact on system performance, thereby achieving higher attack effectiveness with limited attack resources.

[0007] However, existing selective attack methods still have the following shortcomings. (1) Most methods adopt fixed attack rules or random attack strategies, lack dynamic evaluation mechanisms for different measurement data importance, and are difficult to accurately identify the key measurement information that has the greatest impact on vehicle control performance. (2) Existing attack decision-making methods usually adopt a single threshold judgment mechanism, which can only decide whether to carry out an attack, and cannot further distinguish different levels of attack targets based on the importance of the measurement data; (3) Existing methods are difficult to dynamically adjust the attack range according to the vehicle's operating status, and cannot achieve flexible switching between single-channel attacks and multi-channel coordinated attacks; (4) Existing attack strategies generally suffer from problems such as low utilization of attack resources, insufficient attack concealment, and limited attack effectiveness; (5) There is currently a lack of a selective attack method for key measurement data of autonomous vehicles that can take into account the attack's targeting, concealment, and destructiveness.

[0008] Patent application CN119167255A proposes a method, system, and equipment for spoofing and concealment testing in process industry systems, belonging to the field of industrial control system automation monitoring and process safety technology. This invention addresses the problem of incomplete safety testing and drills in process industry systems. It analyzes historical and online measurement data collected from industrial processes, designs attack vectors, generates data sequences using an online adaptive update strategy, and then injects these sequences into the real-time running industrial control system. This causes the controlled system's state to deviate from the expected value without triggering additional alarms in the detection system. However, it relies on complex matrix operations to construct spoofed signals for injection and lacks a real-time evaluation mechanism for the importance of each measurement channel. It cannot achieve adaptive, accurate, and highly concealed selective testing of dynamic critical channels using a low-complexity historical data retention strategy. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the existing technology by providing a data-feature-based method and system for assessing the safety vulnerabilities of autonomous vehicles. This method combines a measurement data importance assessment and dual-threshold dynamic judgment mechanism with a historical data retention strategy to achieve intelligent identification and accurate adaptive testing of key information of autonomous vehicles, thereby effectively revealing the potential safety vulnerabilities of the system.

[0010] The objective of this invention can be achieved through the following technical solutions: A method for assessing the safety vulnerability of autonomous vehicles based on data features includes the following steps: Collect real measurement data of the autonomous vehicle at the current moment and obtain historical measurement data received by the controller at the previous moment. Both the real measurement data and the historical measurement data contain data from multiple measurement channels. Calculate the absolute value of the difference between the actual measurement data and the historical measurement data on each measurement channel, and multiply the absolute value by the importance weight coefficient of the corresponding measurement channel to obtain the importance index of each measurement channel; Calculate the output norm of the actual measurement data, multiply the output norm by a first basic threshold ratio parameter to obtain a first threshold, and multiply the output norm by a second basic threshold ratio parameter to obtain a second threshold, wherein the first basic threshold ratio parameter is less than the second basic threshold ratio parameter. The importance index of each measurement channel is compared with the first threshold and the second threshold respectively to determine the importance region to which each measurement channel belongs. The importance region includes a low importance region, a medium importance region and a high importance region. The measurement channels are sorted in descending order of importance index to determine the highest priority test channel with the highest importance index. The target measurement channel to be tested is determined based on the importance region to which each measurement channel belongs and the highest priority test channel. The number of the target measurement channel to be tested is included in the test target set, and a channel selection matrix is ​​constructed based on the test target set. Based on the channel selection matrix, the historical measurement data is used to replace the actual measurement data of the target measurement channel to generate post-attack measurement data, which is then sent to the controller. By analyzing the state response of the vehicle system to the post-attack measurement data, the security vulnerability assessment of the autonomous vehicle is completed.

[0011] Furthermore, both the actual measurement data and the historical measurement data include yaw rate measurement data, heading error measurement data, and lateral offset measurement data; The measurement channels include channels for transmitting yaw rate, heading error, and lateral offset.

[0012] Furthermore, the step of determining the importance region to which each measurement channel belongs specifically includes: When the importance index of the measurement channel is not greater than the first threshold, the measurement channel is determined to be in the low importance region; When the importance index of a measurement channel is greater than the first threshold but not greater than the second threshold, the measurement channel is determined to be in the medium importance region. When the importance index of a measurement channel is greater than the second threshold, the measurement channel is determined to be in a high importance region.

[0013] Furthermore, the step of determining the target measurement channel to be tested based on the importance region to which each measurement channel belongs and the highest priority test channel specifically includes: When all measurement channels are in the low importance region, no test is performed, and the test target set is an empty set; When the highest priority test channel is in the medium importance region, the highest priority test channel will be the only target measurement channel to be tested. When a measurement channel exists in the high importance region, all measurement channels in the high importance region are collectively used as the target measurement channel to be tested.

[0014] Furthermore, the step of constructing the channel selection matrix based on the test target set specifically includes: Construct a diagonal matrix as the channel selection matrix. When the number of a measurement channel is included in the test target set, the corresponding diagonal element in the channel selection matrix is ​​zero; when the number of a measurement channel is not included in the test target set, the corresponding diagonal element in the channel selection matrix is ​​one.

[0015] Furthermore, the step of generating post-attack measurement data specifically includes: Multiplying the channel selection matrix by the actual measurement data yields a first vector; multiplying the difference between the identity matrix and the channel selection matrix by the historical measurement data yields a second vector; and adding the first vector and the second vector yields the post-attack measurement data.

[0016] Furthermore, the importance weighting coefficients corresponding to the yaw rate measurement data, the heading error measurement data, and the lateral offset measurement data are different from each other.

[0017] Furthermore, after sorting the various measurement channels in descending order according to the importance index, the second and third most important measurement channels are determined sequentially from the remaining measurement channels that have not yet been determined as the highest priority test channels.

[0018] Furthermore, the process of replacing the actual measurement data of the target measurement channel with the historical measurement data adopts a historical data retention mechanism, so that the untested measurement channel can normally transmit the actual measurement data to the controller.

[0019] An autonomous vehicle safety vulnerability assessment system based on data features, comprising: The measurement data acquisition module is used to acquire the real measurement data of the autonomous vehicle at the current moment, and the real measurement data includes data from multiple measurement channels; The historical data acquisition module is used to acquire the historical measurement data received by the controller at the previous moment. The historical measurement data includes data corresponding to multiple measurement channels. The data importance assessment module is used to calculate the absolute value of the data difference between the real measurement data and the historical measurement data in each measurement channel, and multiply the absolute value by the importance weight coefficient of the corresponding measurement channel to obtain the importance index of each measurement channel. The dual threshold determination module is used to calculate the output norm of the real measurement data, multiply the output norm by a first basic threshold ratio parameter to obtain a first threshold, multiply the output norm by a second basic threshold ratio parameter to obtain a second threshold, and compare the importance index of each measurement channel with the first threshold and the second threshold respectively to determine the importance region to which each measurement channel belongs. The test target sorting module is used to sort each measurement channel in descending order according to the importance index, and determine the highest priority test channel with the highest importance index. The test range determination module is used to determine the target measurement channel to be tested based on the importance region to which each measurement channel belongs and the highest priority test channel, and to include the number of the target measurement channel to be tested into the test target set, and to construct a channel selection matrix based on the test target set; The selective test execution module is used to generate post-attack measurement data by replacing the real measurement data of the target measurement channel with the historical measurement data based on the channel selection matrix. The post-test data output module is used to send the post-attack measurement data to the controller.

[0020] Compared with the prior art, the present invention has the following beneficial effects: 1. Accurately identify key vulnerable nodes and improve test targeting: Collect current real measurement data and historical measurement data and calculate the absolute value of the difference between the corresponding channels. Multiply the absolute value by the importance weight coefficient of each channel to obtain the importance index. Then sort the channels in descending order of importance index to lock the highest priority test channels. This achieves the accurate extraction of vehicle status information with high vulnerability value and concentrates limited test resources on the key channels that have the greatest impact on the lateral control of the vehicle.

[0021] 2. Achieve refined classification of test levels and adaptive adjustment of thresholds: Calculate the output norm of the current real measurement data, multiply it by the first and second basic threshold ratio parameters respectively to obtain the real-time updated double thresholds, and then compare the importance index of each channel with the dynamic double thresholds and classify them into three importance regions: low, medium and high. This achieves the dynamic constraint effect of the threshold boundary adaptively expanding and contracting with the real-time operating status of the vehicle, overcoming the defect of the single threshold mechanism that cannot subdivide the target level.

[0022] 3. Automatically determine dynamic test scope while balancing concealment and evaluation performance: Based on the importance region division results, deviation testing is not performed when all channels are in low importance; when the highest priority channel is in medium importance, data preservation is only performed on that single channel; when a high importance channel exists, multi-channel collaborative testing is triggered, and the target number is extracted to construct a channel selection matrix. This achieves intelligent switching between regular transmission, single-channel boundary testing, and multi-channel collaborative evaluation, maximizing vulnerability disclosure performance without disrupting the communication architecture.

[0023] 4. No complex signal forgery is required and the engineering implementation is simple, with extremely high test concealment: By using a historical data retention mechanism combined with a selection matrix to perform matrix operations, the historical measurement data received and cached by the vehicle controller at the previous moment is retrieved and directly replaces the real data of the target measurement channel at the current moment. It is not transmitted normally by the test channel, thereby achieving the effect of injecting distorted feedback information into the controller without forging complex false data, avoiding the high exposure and irreversible damage risks caused by completely blocking communication. Attached Figure Description

[0024] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram illustrating the principle of importance assessment of measurement data in this invention. Figure 3 This is a schematic diagram of the dual-threshold test area division of the present invention; Figure 4 This is a block diagram of the overall system structure of the present invention; Figure 5 This is a state response diagram of the autonomous vehicle system under no-test conditions according to the present invention; Figure 6 The system state response diagram of an autonomous vehicle under the existing data perception testing scheme; Figure 7 This is a state response diagram of an autonomous vehicle system under the static weighted selective testing scheme of the present invention; Figure 8 This is a state response diagram of an autonomous vehicle system under the dynamic weighted selective testing scheme of the present invention; Figure 9 The graph shows the results of test node selection and the changes in system robustness under different attack schemes. Detailed Implementation

[0025] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are implemented based on the technical solution of the present invention, providing detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0026] Example 1 This embodiment provides a method for assessing the safety vulnerability of autonomous vehicles based on data features, such as... Figure 1 As shown, the specific steps are as follows: The system collects real measurement data of the autonomous vehicle at the current moment through the vehicle communication network, and obtains historical measurement data received by the controller at the previous moment based on the data retention strategy of the vehicle system. Both the real measurement data and the historical measurement data contain data from multiple measurement channels. Both the actual measurement data and the historical measurement data include yaw rate measurement data, heading error measurement data, and lateral offset measurement data. The relevant calculation formulas are as follows: The vehicle's lateral dynamic state vector is defined as: in, Indicates the vehicle's sideslip angle. Indicates the yaw rate of the vehicle. Indicates the vehicle's heading error. This indicates the lateral offset of the vehicle. This represents the discrete sampling time.

[0027] Vehicle control input is defined as: in, This indicates the input for controlling the front wheel steering angle of the vehicle.

[0028] Vehicle measurement output is defined as: in, The vehicle measurement output vector, This is for measuring the output matrix.

[0029] For the lateral control system of an autonomous vehicle, the measurement output can be expressed as: in, This represents the yaw rate measurement data. This represents the heading error measurement data. This represents the measurement data for lateral offset.

[0030] Let the measurement data received by the controller at the previous moment be: in, This indicates the number of samples received by the controller at the previous sampling time. One yaw rate measurement data, This indicates the number of samples received by the controller at the previous sampling time. One heading error measurement data, This indicates the number of samples received by the controller at the previous sampling time. One lateral offset measurement data.

[0031] Calculate the absolute value of the difference between the actual measurement data and the historical measurement data on each measurement channel, and multiply the absolute value by the importance weight coefficient of the corresponding measurement channel to obtain the importance index of each measurement channel; Importance is used to quantify the test value of a measurement channel at the current moment. Its quantification logic includes two aspects: first, the deviation between the current measurement data and the data received by the controller at the previous moment, used to characterize the degree of change or anomaly in the current state of the measurement channel; second, the measurement channel weight, used to characterize the degree of influence of the measurement on the vehicle's lateral stability and trajectory tracking performance. Therefore, for the first... Each measurement channel has an importance defined as follows: in, ; Indicates the first Each measurement channel is in The importance index of a moment; For the first The importance weighting coefficients corresponding to each measurement channel; For the first Real measurement data of each measurement channel at the current moment; The first time the controller received the data at the previous moment Data from each measurement channel.

[0032] like Figure 2 As shown, the weighting coefficients This is used to reflect the degree of influence of different measurement channels on the vehicle's lateral stability, trajectory tracking performance, and control safety. For the lateral control system of autonomous vehicles, yaw rate, heading error, and lateral offset have different importance to vehicle motion control, so different weighting coefficients can be set.

[0033] In one specific embodiment, the following may be adopted: in, Corresponding to the yaw rate measurement channel, Corresponding heading error measurement channel, Corresponding to the lateral offset measurement channel.

[0034] Importance The larger the value, the more significant the deviation between the current measurement data and the historical received data, and the greater the impact of the measurement channel on the vehicle control performance. Therefore, the measurement channel has higher testing value.

[0035] To avoid wasting testing resources by having testers perform tests at all times, and to avoid the problem of imprecise test level classification caused by using only a single threshold, this embodiment adopts a dual-threshold testing mode, such as... Figure 3 As shown, it should be noted that the dual thresholds in this embodiment are not fixed constant thresholds, but a dynamic threshold mechanism determined by the basic proportional coefficient and the system operating state.

[0036] Calculate the output norm of the actual measurement data, multiply the output norm by a first basic threshold ratio parameter to obtain a first threshold, and multiply the output norm by a second basic threshold ratio parameter to obtain a second threshold, wherein the first basic threshold ratio parameter is less than the second basic threshold ratio parameter. First, set two basic threshold ratio parameters. and ,in: Based on this, the output norm of the current measurement vector is introduced. Construct a double threshold function related to the system's operating state: First threshold: Second threshold: Therefore, the absolute value of the threshold varies with the system's operating state. It changes with the change, but its relative proportion is determined by the parameter. and This decision allows for the realization of the dual characteristics of "fixed ratio constraint + dynamic amplitude adjustment".

[0037] The importance index of each measurement channel is compared with the first threshold and the second threshold respectively to determine the importance region to which each measurement channel belongs. The importance region includes a low importance region, a medium importance region and a high importance region. Low importance areas: Medium importance area: High importance areas: in, Indicates the first The measurement channels are in a low importance state. Indicates the first Each measurement channel is in a medium importance state. Indicates the first Each measurement channel is in a state of high importance.

[0038] When a measurement channel is in a low importance region, it indicates that the current testing value of the channel is low and it is not advisable to consume testing resources. When a measurement channel is in a medium importance region, it indicates that the channel has certain testing value and can be used as a single-channel testing target. When a measurement channel is in a high importance region, it indicates that the channel is the key measurement information at the current moment and should be given priority as a testing object.

[0039] The measurement channels are sorted in descending order of importance index to determine the highest priority test channel with the highest importance index. First, determine the measurement channel with the highest importance index: in, This indicates the measurement channel number with the highest importance at the current moment.

[0040] Furthermore, a second important measurement channel is identified from the remaining measurement channels: And determine the third measurement channel: This yields the importance ranking of the measurement channels at the current moment: in, This is the highest priority testing channel. This is the second-priority testing channel. For the remaining channels.

[0041] Based on the importance region of each measurement channel and the highest priority test channel, the influence of different measurement channels on vehicle control performance is evaluated, the target measurement channel to be tested is dynamically determined, and the number of the target measurement channel to be tested is included in the test target set. A channel selection matrix is ​​constructed based on the test target set. Based on the results of the dual threshold determination and the ranking of test targets, this embodiment determines the test range.

[0042] Scenario 1: No testing is performed; if the importance index of all measurement channels does not exceed the first threshold, i.e.: At this point, for all All meet If the measurement data falls into a low importance region, it is considered that the test value of all measurement channels is low at the current moment. In order to save test resources and reduce exposure risks, no test is performed, and all measurement data is transmitted to the controller normally.

[0043] The second scenario: Implement single-channel testing; if the measurement channel with the highest importance is... satisfy: At this time, for satisfy If the value falls within the medium importance region, while the other measurements must fall within the medium / low importance region, then the current time value is considered to be... This is a relatively important measurement channel, but the conditions for multi-channel testing have not yet been met. At this time, only the measurement channel with the highest importance is tested, while the other measurement channels transmit normally.

[0044] The third scenario: Implement multi-channel collaborative testing; if one or more measurement channels satisfy: These measurement channels are considered to be of high importance, and in this case, all measurement channels that meet the above conditions are tested simultaneously.

[0045] Therefore, it can automatically switch dynamically between "no testing - single-channel testing - multi-channel testing" based on the current operating status of the vehicle.

[0046] To provide a unified description of the test target set, this embodiment defines the test target set. , When a measurement channel is selected as a test target, its number is added to the test target set. When a measurement channel is not selected, its number is not included in the test target set.

[0047] Based on the test target set Define the test selection matrix: Among them: when , ;when , .

[0048] Test selection matrix Used to characterize which measurement channels are being tested at the current moment.

[0049] For example: when hour, This indicates that no measurement channel was tested. when hour, This indicates that only the yaw rate measurement channel was tested. when hour, This indicates that only the heading error measurement channel was tested; when hour, This indicates that only the lateral offset measurement channel was tested; when hour, This indicates that the yaw rate and heading error measurement channels are being tested simultaneously. when hour, This indicates that the yaw rate and lateral offset measurement channels are being tested simultaneously. when hour, This indicates that the heading error and lateral offset measurement channels are being tested simultaneously. when hour, This indicates that all three measurement channels are being tested simultaneously.

[0050] This allows for coverage of eight test scenarios across three-channel measurement data.

[0051] Based on the channel selection matrix, the historical data retention method is used to replace the actual measurement data of the target measurement channel with the historical measurement data to generate distorted post-evaluation measurement data, and the distorted post-evaluation measurement data is sent to the controller to trigger the controller's control decision response.

[0052] For any measurement channel The actual data received by the controller after the test Defined as: like ,but: This indicates that the channel has not been tested and that the actual measurement data is being transmitted to the controller normally.

[0053] like ,but: This indicates that the channel is under test, and the current actual measurement data has been replaced by the data received by the controller in the previous moment.

[0054] Therefore, the overall measurement data after the test can be uniformly represented as: in, It is the identity matrix. The measurement data received by the controller after the test. Based on current actual measurement data, This refers to the measurement data received by the controller at the previous moment.

[0055] The above testing method can maintain the normal transmission of some measurement data while only causing covert interference to the key measurement channels, making it difficult for the controller to directly determine that the communication link has been completely interrupted, thus distinguishing it from traditional denial-of-service testing.

[0056] Example 2 This embodiment provides a method for assessing the safety vulnerability of autonomous vehicles based on data features, such as... Figure 4 As shown, it includes the following modules: The measurement data acquisition module is used to acquire the real measurement data of the autonomous vehicle at the current moment, and the real measurement data includes data from multiple measurement channels; The historical data acquisition module is used to acquire the historical measurement data received by the controller at the previous moment. The historical measurement data includes data corresponding to multiple measurement channels. The data importance assessment module is used to calculate the absolute value of the data difference between the real measurement data and the historical measurement data in each measurement channel, and multiply the absolute value by the importance weight coefficient of the corresponding measurement channel to obtain the importance index of each measurement channel. The dual threshold determination module is used to calculate the output norm of the real measurement data, multiply the output norm by a first basic threshold ratio parameter to obtain a first threshold, multiply the output norm by a second basic threshold ratio parameter to obtain a second threshold, and compare the importance index of each measurement channel with the first threshold and the second threshold respectively to determine the importance region to which each measurement channel belongs. The test target sorting module is used to sort each measurement channel in descending order according to the importance index, and determine the highest priority test channel with the highest importance index. The test range determination module is used to determine the target measurement channel to be tested based on the importance region to which each measurement channel belongs and the highest priority test channel, and to include the number of the target measurement channel to be tested into the test target set, and to construct a test selection matrix based on the test target set; A selective test execution module is used to generate post-attack measurement data by replacing the real measurement data of the target measurement channel with the historical measurement data based on the test selection matrix. The post-test data output module is used to send the post-attack measurement data to the controller.

[0057] Example 3 To verify the effectiveness and advancement of the data feature-based autonomous vehicle safety vulnerability assessment method and system described in this embodiment, comparative tests were conducted on the constructed autonomous vehicle lateral dynamics control simulation platform. During the simulation, the vehicle lateral dynamics state vector includes four state variables: vehicle sideslip angle, yaw rate, heading error, and lateral offset.

[0058] like Figure 5 The figure shows the state response of the autonomous vehicle system under normal transmission conditions without external test interference, as illustrated in this embodiment. Under this baseline state, the four state variables in the vehicle's lateral dynamics state vector rapidly approach a stable value of 0 after a brief initial fluctuation. A magnified view on the left side of the figure shows that within 5 to 10 seconds, each state component has completely converged to a small oscillation range near zero. Simultaneously, the output norm of the system state vector, shown on the right, smoothly decreases from an initial peak close to 7 and converges to 0 within approximately 5 seconds. The magnified view further confirms the smooth process of the norm converging to an extremely small value below 0 within 5 to 10 seconds and ultimately approaching 0. These simulation results demonstrate that, without vulnerability testing interference, the vehicle's lateral control system exhibits good stability and path-tracking performance.

[0059] To compare and evaluate performance, this embodiment introduces a common data-aware deviation testing scheme in the prior art as a comparison group. For example... Figure 6 The diagram shows the state response of an autonomous vehicle system under a current data sensing scheme. Under this scheme's interference, although the four state components appear to stabilize after approximately 2 seconds, a magnified view reveals that the state variables do not converge to zero within the 5-10 second interval, instead exhibiting residual steady-state bias. Correspondingly, the output norm in the right-hand diagram fails to fully converge to zero after a rapid decrease. This indicates that the current conventional scheme only causes a slight decrease in vehicle control performance, demonstrating limited effectiveness in revealing vulnerabilities.

[0060] Furthermore, after activating the static weight selective testing strategy proposed in this embodiment, as follows: Figure 7 As shown, each state variable exhibits significant fluctuations in the initial stage of testing, subsequently stabilizing. A magnified view reveals that, in the steady-state phase, each state variable retains a larger and more pronounced residual deviation than existing schemes, triggering coordinated deviations in other channels. The corresponding output norm, after initial sharp fluctuations, eventually stabilizes at the deviation level. This steady-state deviation is significantly higher than that of existing schemes, strongly demonstrating that, under the same testing resource conditions, this embodiment, by accurately identifying and perturbing key data channels, can more effectively reveal the control vulnerabilities of the system on specific channels.

[0061] To thoroughly exploit system vulnerabilities, this embodiment activates a highly evolving dynamic weighted selective testing strategy. For example... Figure 8 As shown, the vehicle state exhibits completely different divergent characteristics. Although the first three state components barely remain near 0, the fourth key state component completely loses its ability to converge to zero after the initial stage, exhibiting a linear divergence over time and a continuously rising, drastic divergence, with its value approaching 20 at 15 seconds. Similarly, the output norm in the right figure shows a completely non-convergent and continuously rising steep characteristic, also approaching 20 at 15 seconds. This set of data indicates that this embodiment, through a dynamic weighting strategy, accurately locks the key data channels of the system's real-time changes and continuously injects distorted feedback, successfully inducing the complete instability and collapse of the autonomous vehicle control system, thereby maximizing the induction and revelation of the vehicle control system's collapse boundary.

[0062] Finally, as Figure 9 The figure shows a comprehensive comparison of node selection results and system robustness changes under different evaluation schemes. This figure elucidates the inherent adaptive mechanism of this embodiment from two dimensions: micro-level node scheduling and macro-level system performance. Figure 9 Figure 'a' shows the dynamic switching trajectory of the selected test channel number over time under the data perception deviation test scheme. Channel 2 is selected as the core interference target for a long time, while channels 1 and 3 are selected alternately at high frequency in the early stage. The dense switching details within 0 to 5 seconds are clearly shown in the magnified part, which confirms the intelligence of this embodiment in dynamically adjusting the test target according to the importance of data. Figure 9 b comprehensively compares the robust performance index evolution curves under the no-interference, data-aware deviation testing schemes, and static weight selective testing strategies. The curves show that the static weight selective testing strategy used in this embodiment can cause the system performance index to deviate significantly from the no-interference state and exceed the safety threshold. Among them, the static weight selective testing strategy causes the highest degree of performance degradation and the longest duration. Figure 9 c further demonstrates the channel node selection results under the static weight selective testing strategy. Since this strategy captures more critical features, the test points are extremely densely concentrated on channel 2, forming a large solid area on the graph. This perfectly verifies the engineering accuracy and efficiency of single-channel and multi-channel collaborative switching when performing historical data retention tests in this embodiment.

[0063] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for assessing the safety vulnerability of autonomous vehicles based on data features, characterized in that, Includes the following steps: Collect real measurement data of the autonomous vehicle at the current moment and obtain historical measurement data received by the controller at the previous moment. Both the real measurement data and the historical measurement data contain data from multiple measurement channels. Calculate the absolute value of the difference between the actual measurement data and the historical measurement data on each measurement channel, and multiply the absolute value by the importance weight coefficient of the corresponding measurement channel to obtain the importance index of each measurement channel; Calculate the output norm of the actual measurement data, multiply the output norm by a first basic threshold ratio parameter to obtain a first threshold, and multiply the output norm by a second basic threshold ratio parameter to obtain a second threshold, wherein the first basic threshold ratio parameter is less than the second basic threshold ratio parameter. The importance index of each measurement channel is compared with the first threshold and the second threshold respectively to determine the importance region to which each measurement channel belongs. The importance region includes a low importance region, a medium importance region and a high importance region. The measurement channels are sorted in descending order of importance index to determine the highest priority test channel with the highest importance index. The target measurement channel to be tested is determined based on the importance region to which each measurement channel belongs and the highest priority test channel. The number of the target measurement channel to be tested is included in the test target set, and a channel selection matrix is ​​constructed based on the test target set. Based on the channel selection matrix, the historical measurement data is used to replace the actual measurement data of the target measurement channel to generate post-attack measurement data, which is then sent to the controller. By analyzing the state response of the vehicle system to the post-attack measurement data, the security vulnerability assessment of the autonomous vehicle is completed.

2. The method for assessing the safety vulnerability of autonomous vehicles based on data features according to claim 1, characterized in that, Both the actual measurement data and the historical measurement data include yaw rate measurement data, heading error measurement data, and lateral offset measurement data. The measurement channels include channels for transmitting yaw rate, heading error, and lateral offset.

3. The method for assessing the safety vulnerability of autonomous vehicles based on data features according to claim 1, characterized in that, The step of determining the importance region to which each measurement channel belongs specifically includes: When the importance index of a measurement channel is not greater than the first threshold, the measurement channel is determined to be in a low importance region. When the importance index of a measurement channel is greater than the first threshold but not greater than the second threshold, the measurement channel is determined to be in the medium importance region. When the importance index of a measurement channel is greater than the second threshold, the measurement channel is determined to be in a high importance region.

4. The method for assessing the safety vulnerability of autonomous vehicles based on data features according to claim 3, characterized in that, The step of determining the target measurement channel to be tested based on the importance region to which each measurement channel belongs and the highest priority test channel specifically includes: When all measurement channels are in the low importance region, no test is performed, and the test target set is an empty set; When the highest priority test channel is in the medium importance region, the highest priority test channel will be the only target measurement channel to be tested. When a measurement channel exists in the high importance region, all measurement channels in the high importance region are collectively used as the target measurement channel to be tested.

5. The method for assessing the safety vulnerability of autonomous vehicles based on data features according to claim 1, characterized in that, The step of constructing a channel selection matrix based on the test target set specifically includes: Construct a diagonal matrix as the channel selection matrix. When the number of a measurement channel is included in the test target set, the corresponding diagonal element in the channel selection matrix is ​​zero; when the number of a measurement channel is not included in the test target set, the corresponding diagonal element in the channel selection matrix is ​​one.

6. The method for assessing the safety vulnerability of autonomous vehicles based on data features according to claim 1, characterized in that, The steps for generating measurement data after the attack specifically include: Multiplying the channel selection matrix by the actual measurement data yields a first vector; multiplying the difference between the identity matrix and the channel selection matrix by the historical measurement data yields a second vector; and adding the first vector and the second vector yields the post-attack measurement data.

7. The method for assessing the safety vulnerability of autonomous vehicles based on data features according to claim 2, characterized in that, The importance weighting coefficients corresponding to the yaw rate measurement data, the heading error measurement data, and the lateral offset measurement data are different.

8. The method for assessing the safety vulnerability of autonomous vehicles based on data features according to claim 1, characterized in that, After sorting the various measurement channels in descending order of importance index, the second and third most important measurement channels are determined sequentially from the remaining measurement channels that have not yet been determined as the highest priority test channels.

9. The method for assessing the safety vulnerability of autonomous vehicles based on data features according to claim 1, characterized in that, The process of replacing the actual measurement data of the target measurement channel with the historical measurement data adopts a historical data retention mechanism, which enables the untested measurement channel to transmit the actual measurement data to the controller normally.

10. An autonomous vehicle safety vulnerability assessment system based on data features, characterized in that, include: The measurement data acquisition module is used to acquire the real measurement data of the autonomous vehicle at the current moment, and the real measurement data includes data from multiple measurement channels; The historical data acquisition module is used to acquire the historical measurement data received by the controller at the previous moment. The historical measurement data includes data corresponding to multiple measurement channels. The data importance assessment module is used to calculate the absolute value of the data difference between the real measurement data and the historical measurement data in each measurement channel, and multiply the absolute value by the importance weight coefficient of the corresponding measurement channel to obtain the importance index of each measurement channel. The dual threshold determination module is used to calculate the output norm of the real measurement data, multiply the output norm by a first basic threshold ratio parameter to obtain a first threshold, multiply the output norm by a second basic threshold ratio parameter to obtain a second threshold, and compare the importance index of each measurement channel with the first threshold and the second threshold respectively to determine the importance region to which each measurement channel belongs. The test target sorting module is used to sort each measurement channel in descending order according to the importance index, and determine the highest priority test channel with the highest importance index. The test range determination module is used to determine the target measurement channel to be tested based on the importance region to which each measurement channel belongs and the highest priority test channel, and to include the number of the target measurement channel to be tested into the test target set, and to construct a channel selection matrix based on the test target set; The selective test execution module is used to generate post-attack measurement data by replacing the real measurement data of the target measurement channel with the historical measurement data based on the channel selection matrix. The post-test data output module is used to send the post-attack measurement data to the controller.

Images