Intelligent networked vehicle multi-level authentication and active safety protection system and method

Abstract

The present application relates to the technical field of intelligent networked vehicle safety, in particular to a multi-level authentication and active safety protection system and method for intelligent networked vehicles, the core of the system is composed of six basic function modules in cooperation, specifically including: a multi-modal physical fingerprint collection module, a cross-modal cross-validation engine, a motion physical constraint authentication module, a scene semantic firewall engine, an active challenge-response authentication module, and a cooperative response and trust management module; in addition, the system can also configure an active trapping and attack countermeasure module as needed to further enhance the active defense capability. The present application deeply integrates physical layer unclonable features, dynamic behavior trust evaluation, motion physical constraints, scene semantic verification, active challenge-response authentication and edge cooperative defense technology, and realizes all-round safety protection from in-vehicle network to V2X communication, from passive defense to active protection, from single vehicle safety to regional cooperative defense.

Description

Technical Field

[0001] This invention relates to the field of intelligent connected vehicle safety technology, specifically to a multi-level authentication and active safety protection system and method for intelligent connected vehicles. Background Technology

[0002] Intelligent connected vehicles communicate with external traffic facilities, cloud servers, and other vehicles in real time through onboard communication interfaces (such as CAN bus and onboard Ethernet), wireless access methods (such as 4G / 5G, Bluetooth, and Wi-Fi), and V2X (Vehicle-to-Everything) communication, significantly improving driving safety and traffic efficiency. However, the rapid expansion of the attack surface has also brought serious information security challenges. Attackers can use physical access via the OBD-II port, remotely crack the T-Box, forge V2X messages, and other means to inject vehicle control commands, steal privacy, and even remotely manipulate braking and steering, seriously threatening personal and public safety.

[0003] Currently, the main technical deficiencies in safety protection technologies for intelligent connected vehicles are as follows: (i) Each security module operates independently and lacks collaborative detection capabilities. Existing vehicle security solutions typically employ a "layered" defense architecture, deploying a signature-based intrusion detection system (IDS) on the CAN bus, setting up a rule-based vehicle firewall on the gateway, and relying on PKI / CA certificate systems for authentication in V2X communication. These modules operate independently, with no sharing of detection results, making coordinated responses impossible. For example, physical fingerprint-based authentication technology only determines whether the message sender is a legitimate ECU, but once a legitimate ECU is controlled by an attacker (firmware tampering or memory injection), its physical fingerprint remains genuine, rendering such solutions completely ineffective. Behavioral analysis-based intrusion detection systems only focus on statistical characteristics such as message frequency and sequence, failing to identify covert attacks with genuine physical identities but abnormal behavior. The lack of information fusion and cross-verification channels between modules allows attackers to exploit detection blind spots between modules to launch attacks.

[0004] (ii) V2X message authenticity authentication relies too heavily on digital certificates Existing V2X security standards (such as IEEE 1609.2) primarily employ Public Key Infrastructure (PKI) to verify the signatures of Basic Security Messages (BSMs) to ensure the legitimacy and integrity of the message source. However, digital certificate mechanisms have inherent flaws: Certificate Authorities (CAs) can be compromised or subject to internal malicious activity; certificate revocation mechanisms have a delay window of several minutes to hours; attackers can steal legitimate certificates to issue fake but valid signatures for BSMs, such as forging a "ghost vehicle" that does not exist in the physical world, whose position, speed, heading angle, and other motion parameters violate basic kinematic laws. Because existing schemes only verify the validity of the signature and do not verify the physical authenticity of the message content, surrounding vehicles receiving such forged BSMs may misjudge the risk, triggering emergency braking or incorrect lane changes, leading to a chain reaction of accidents.

[0005] (iii) The CAN bus defense mechanism remains at the syntax level and lacks scene-aware semantic detection capabilities. Existing CAN bus intrusion detection methods primarily rely on syntactic features such as message period, inter-frame arrival time difference, and CANID sequence order for anomaly detection (e.g., using Markov chains or Long Short-Term Memory networks to predict the next frame ID). These methods are almost incapable of recognizing instructions that are "semantically valid but illogical in context." For example, a vehicle traveling at 120 km / h on a highway might receive a diagnostic session request (0x27 service), which, while a valid message format according to the CAN protocol specification, could very likely be a malicious instruction injected by an attacker through the OBD-II port. Similarly, when a vehicle is stationary in Park, the steering system receives a large-angle steering command—the message's ID, CRC, and period all conform to the specifications, but this clearly contradicts the physical scenario. Current technologies lack a dynamic detection mechanism that combines real-time driving scenarios (parking, cruising, high-speed driving) with ECU operation semantics, making it impossible to identify such attacks at the semantic level.

[0006] (iv) Physical layer fingerprint authentication is a static and passive mechanism, which is vulnerable to machine learning attacks. In recent years, ECU fingerprint authentication technology based on the physical layer characteristics of the CAN bus has attracted widespread attention. Research shows that due to manufacturing process variations, CAN messages output by different ECUs possess unique and stable physical layer characteristics (including clock skew, differential voltage rise / fall times, and inter-frame arrival time differences). However, existing physical fingerprint authentication schemes (such as CN111131302A) are all static passive authentication—the system passively extracts the fingerprint and compares it with a pre-stored template only when the message arrives, without actively intervening in the ECU's behavior. This design has a fundamental security flaw: attackers can collect a large number of physical layer feature samples by long-term monitoring of the bus traffic of legitimate ECUs. They can then use machine learning models such as Generative Adversarial Networks (GANs) to learn the fingerprint distribution and synthesize forged messages that can pass authentication. Once adversarial learning is successful, the static physical fingerprint barrier becomes ineffective.

[0007] (v) The single-vehicle defense model limits the sharing of threat intelligence and collaborative response. Currently, most intelligent connected vehicle security solutions are limited to single-vehicle autonomous defense, lacking effective threat intelligence sharing mechanisms between vehicles and between vehicles and roadside infrastructure. When a vehicle detects an attack source (e.g., a certain IP continuously launching diagnostic scans), this information cannot be immediately transmitted to other nearby vehicles, allowing the same attacker to penetrate each vehicle individually, resulting in low defense efficiency. Although some studies have proposed blockchain-based threat intelligence sharing frameworks, the real-time nature and lightweight nature of multi-centralized deployments are insufficient to meet the low-latency collaboration requirements in high-speed vehicle movement scenarios.

[0008] In summary, existing technologies lack a multi-dimensional defense solution that can integrate the features of the physical layer, data link layer, application layer, and V2X communication layer to achieve inter-module collaborative linkage, proactive authentication and scenario adaptation capabilities, and support vehicle-edge collaboration. This invention is a solution proposed to address the aforementioned technical problems. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the prior art and propose a multi-level authentication and active safety protection system and method for intelligent connected vehicles. It deeply integrates physical layer unclonable features, dynamic behavior trust assessment, motion physical constraints, scene semantic verification, active challenge-response authentication and edge collaborative defense technology to achieve all-round security protection from in-vehicle network to V2X communication, from passive defense to active protection, and from single vehicle security to regional joint defense.

[0010] To achieve the above objectives, the technical solution specifically adopted by the present invention is as follows: This invention proposes a multi-level authentication and proactive security protection system for intelligent connected vehicles. Deployed on in-vehicle computing platforms and edge computing nodes, the system aims to achieve end-to-end security from the underlying hardware to upper-layer applications through a multi-physical domain three-dimensional defense architecture. The core of the system consists of six basic functional modules working together: a multimodal physical fingerprint acquisition module, a cross-modal cross-verification engine, a motion physical constraint authentication module, a scene semantic firewall engine, a proactive challenge-response authentication module, and a collaborative response and trust management module. Furthermore, the system can be configured as needed with proactive decoy and attack countermeasure modules to further enhance proactive defense capabilities. These modules operate in a deeply integrated manner within a central gateway or domain controller, jointly constructing a dynamic and three-dimensional in-vehicle network security defense line.

[0011] As the data cornerstone of the entire defense system, the multimodal physical fingerprint acquisition module mainly operates during the vehicle off-line stage, responsible for extracting unique physical features from each electronic control unit (ECU). This module collects three types of physical layer features that are difficult to forge with software: first, time features, namely the probability distribution parameters of the arrival time difference between message frames with the same CANID, reflecting the process deviations in crystal oscillator accuracy and bus arbitration delay; second, analog features, including the rise and fall times and overshoot amplitude of the differential voltages of CAN_H and CAN_L, determined by the inherent physical characteristics of the transceiver chip; and third, clock offset features, based on the relative clock offset curve calculated from the sender's embedded timestamp and the gateway's local timestamp. The collected multimodal fingerprints, after being encrypted using SHA-256 hashing, are not only stored in the onboard security hardware module (HSM) but also synchronized to trusted edge nodes on the roadside unit or base station side. The edge nodes use consortium blockchain technology for evidence storage, effectively preventing cloud data tampering and providing reliable support for cross-vehicle fingerprint comparison and traceability.

[0012] During actual vehicle operation, the cross-modal cross-verification engine, acting as the core hub for collaboration between the physical and application layers, connects to the CAN bus in real time and performs parallel dual-branch processing on each message. Branch A is responsible for physical layer authentication, extracting a dynamic physical fingerprint token through a sliding window, comparing it with a pre-stored fingerprint using Mahalanobis distance, and outputting a physical layer confidence score (T_phy) in the range of 0 to 1. Branch B is responsible for behavior analysis, inputting message content features into the Isolation Forest algorithm model and outputting anomaly behavior score (S_behavior). Based on this, the engine forms four judgment scenarios: when T_phy is high and S_behavior is low, it is judged as a normal message and allowed to pass; when T_phy is high but S_behavior is high, it is judged as a legitimate node being controlled (firmware or memory being tampered with), and forced physical fingerprint re-verification is immediately triggered; when T_phy is low and S_behavior is high, it is judged as a fake node attack, the message is directly discarded and the attack source is recorded; when both are low, it requests an update of the fingerprint template from the edge trusted node. This mechanism establishes a two-way verification channel for physical identity and behavioral intent in the vehicle network, overcoming the shortcomings of existing technologies where the two are isolated from each other.

[0013] The forced physical fingerprint re-verification is triggered by the cross-modal cross-verification engine and executed by the active challenge-response authentication module. Specifically, it includes: the cross-modal cross-verification engine sending a random challenge value to the electronic control unit to be verified; the electronic control unit to be verified encrypting the random challenge value based on the pre-shared key and returning a response message; the active challenge-response authentication module synchronously collecting the physical layer features (including differential voltage timing features and clock offset features) of the response message; when the encrypted response verification fails, or when the deviation between the physical layer features and the corresponding features in the pre-stored multimodal physical fingerprint exceeds a preset threshold, the re-verification is determined to have failed, the electronic control unit is isolated, and the result is reported to the edge trusted node.

[0014] To address the challenge of verifying the authenticity of V2X communication, the system introduces a motion physics constraint authentication module. This module abandons the sole reliance on digital certificates and instead performs triple verification of the Basic Safety Message (BSM) based on the laws of physical motion: First, it verifies the compliance of the high-precision map location, confirming that the vehicle is within a legally drivable area; second, it verifies the consistency of vehicle dynamics, with a built-in vehicle model parameter matching submodule retrieving the corresponding vehicle model's dynamic parameter envelope (including at least the allowable range of maximum acceleration, maximum deceleration, maximum lateral acceleration, and maximum yaw rate) from trusted edge nodes or the cloud based on the vehicle type identifier in the BSM, comparing the actual acceleration and yaw rate calculated from consecutive BSMs to see if they exceed this envelope; finally, it verifies historical trajectory prediction, using a Kalman filter to generate a 95% confidence ellipse to determine if there are any abnormal deviations in the reported location. If any verification fails, the module marks the vehicle as a "ghost vehicle" and broadcasts a risk warning via V2X and reports it to edge nodes.

[0015] To address the blind spot in CAN bus command detection where "semantically valid but illogical in context," the scene semantic firewall engine integrates multi-source sensor information such as IMU, wheel speed, and steering angle to infer driving scenarios like parking, city cruising, and highway driving in real time. The engine has a built-in semantic rule library and scene mapping table, dynamically adjusting the strictness of semantic operations for each ECU based on the current scenario. For example, non-safe commands such as unlocking doors are strictly prohibited during highway driving, while the frequency of radar data requests is relaxed during low-speed parking. The engine includes a semantic rule dynamic adjustment submodule and a semantic anomaly integrator: the former configures the semantic rule strictness of corresponding control commands, diagnostic requests, and data interaction behaviors for each ECU based on the attributes of parking, low-speed parking, city cruising, and highway driving scenarios, using a scene-rule mapping table; the latter accumulates the number of semantic violations for each ECU, using a time-exponential decay accumulation method. When the accumulated score exceeds a dynamic threshold, it triggers a trust level downgrade for the corresponding ECU and sends a semantic anomaly signal to the cross-modal cross-validation engine to trigger forced revalidation.

[0016] To further defend against adversarial learning attacks, the system incorporates a proactive challenge-response authentication module. This module proactively initiates authentication challenges to target ECUs upon periodic tasks, trust downgrades, semantic anomalies, or the receipt of edge threat intelligence. Specific triggering mechanisms include: periodic triggering (challenging all safety-critical ECUs every preset time window), trust downgrade triggering (initiating multiple consecutive challenges when the trust level of any ECU is downgraded), semantic anomaly triggering (initiating immediate challenges to violating ECUs), and edge trusted node instruction triggering (initiating batch challenges to all ECUs of a specified type upon receiving threat intelligence). The challenger generates a 128-bit random code, and the target ECU must encrypt and return a response using an HSM key within 5ms. Crucially, while receiving the response, the challenger simultaneously acquires bus waveforms and extracts physical layer features using a high-precision ADC. Authentication is only granted if the encryption / decryption is correct and the physical fingerprint matches the specified criteria. This mechanism forces attackers to simultaneously meet both cryptographic and physical hardware requirements within a very short time, fundamentally immunizing against purely software-level machine learning simulation attacks.

[0017] The outputs of all the above modules ultimately converge in the Collaborative Response and Trust Management module. This module integrates five dimensions of factors—physical layer confidence T_phy, behavioral anomaly severity S_behavior, motion constraint pass rate T_motion, semantic violation severity S_semantic, and challenge response pass rate T_challenge—and dynamically calculates the comprehensive trust level (T_combined) using an adaptive weighted algorithm. The integration method can employ weighted linear fusion, nonlinear mapping, or a trained machine learning model, where the weights or mapping parameters of each factor are adaptively adjusted according to the current driving scenario (e.g., increasing the weights of motion constraints and semantic rules during high-speed driving, and increasing the weights of behavioral anomalies in complex urban scenarios). As one possible implementation method, the weighted linear fusion formula is: Where w1~w5 are dynamic weight coefficients and their sum is 1. Based on this, the system executes a four-level access control policy: Level L0 (fully trusted, T_combined≥0.85) allows all operations; Level L1 (slightly suspicious, 0.60≤T_combined<0.85) restricts unnecessary remote commands and strengthens monitoring; Level L2 (highly suspicious, 0.30≤T_combined<0.60) retains only basic braking and steering functions and cuts off insecure links; Level L3 (malicious confirmation, T_combined<0.30) actively disconnects the network, forces a speed reduction and stop, and broadcasts a warning to surrounding areas. Simultaneously, the vehicle synchronizes anonymized threat intelligence to edge nodes. Once a regional attack pattern is identified, a joint defense command can be generated within seconds and pushed to all vehicles in the network.

[0018] As an advanced capability of the defense system, the system can be optionally equipped with an active decoy and attack countermeasure module. This module periodically injects disguised high-value decoys (such as virtual diagnostic sessions, ghost vehicle BSMs, etc.) into the communication link. When an attacker attempts to read or modify the decoy, the system immediately collects its network and behavioral characteristics, adds them to a blacklist, and lures them to a virtual sandbox electronic control unit built based on software virtualization technology (this virtual sandbox ECU is an isolated execution environment and is not connected to the real physical actuators). At the same time, the relevant threat characteristics are synchronized to edge nodes and other vehicles in the fleet, realizing a leap from passive defense to active countermeasures.

[0019] This invention also provides a multi-level authentication and active safety protection method for intelligent connected vehicles, implemented based on the above system, including the following steps: S1. After the vehicle is powered on and started, each electronic control unit reports a dynamic physical fingerprint token to the cross-modal cross-validation engine to complete the initial identity authentication and initialize the trust level. S2. During vehicle operation, multimodal physical layer features and content features of CAN bus messages are collected in real time, and physical layer confidence scores and behavior anomaly scores are dynamically calculated. Based on the matching relationship between the two, fake node attacks and legitimate node control attacks are identified. S3. For the received external vehicle V2X basic security messages, perform multi-dimensional physical authenticity verification based on high-precision maps, vehicle dynamics models and historical trajectory predictions, identify ghost car attacks and broadcast risk warnings to the outside world. S4. Integrates data from multiple onboard sensors to identify vehicle driving scenarios in real time. Based on a pre-configured semantic rule base and scenario-rule mapping table, it dynamically updates the semantic operation whitelist and rule strictness of each electronic control unit to detect business semantic violations. S5. When a legitimate node is determined to be under attack, a semantic violation is detected, or the physical layer confidence score deviates from the behavioral anomaly score, a mandatory physical fingerprint re-verification is triggered to perform a second identity verification on the suspicious electronic control unit. S6. The proactive challenge-response authentication module performs proactive authentication on the target electronic control unit based on a periodic mechanism or security event triggering, and dynamically updates the device trust status. S7, the collaborative response and trust management module integrates multi-dimensional security data, adaptively adjusts weights based on the current driving scenario, dynamically calculates the comprehensive trust level, and adjusts in-vehicle communication routing, external network access and vehicle control strategies according to the comprehensive trust level; S8. Synchronize local anonymized threat data to edge trusted nodes to trigger the edge trusted nodes to execute regional joint defense strategies, which include pushing blacklists and defense rules to other vehicles in the region.

[0020] Preferably, the triggering conditions for forced physical fingerprint re-verification in step S5 also include: The same electronic control unit commits a set number of semantic violations more than a set number within a preset time window; The absolute value of the difference between the physical layer confidence score and the behavioral abnormality score of the same electronic control unit is greater than the preset deviation threshold; Received a threat intelligence command for this type of electronic control unit from an edge trusted node.

[0021] This invention has the following characteristics and beneficial effects: Compared with the prior art, the present invention has the following beneficial effects: (1) A pioneering cross-modal verification mechanism effectively identifies attacks that compromise legitimate nodes. In existing solutions, physical layer authentication and behavioral analysis are independent of each other. Attackers can still bypass detection by using physical fingerprint authentication after breaking into a legitimate ECU. This invention establishes a two-way verification channel for identity authenticity and behavioral intent for the first time by jointly determining physical layer confidence scores and behavioral anomaly scores. This can accurately distinguish between attacks that impersonate nodes and attacks that compromise legitimate nodes, filling a gap in this technical field.

[0022] (2) Establish a physical constraint defense line to defend against ghost car attacks from the perspective of physical laws. Existing V2X security relies too heavily on digital certificates, which are powerless against BSMs with valid certificates but fake content. The triple physical authenticity verification proposed in this invention (high-precision map location, vehicle dynamic envelope, and historical trajectory prediction) makes it impossible for attackers to forge BSM sequences that simultaneously meet the requirements of location compliance, motion continuity, and dynamic feasibility, even if they have legitimate certificates, thus solving the ghost car attack problem at its root.

[0023] (3) A semantic firewall that adapts to specific scenarios is implemented to solve the problem of detecting semantic-scenario contradictory attacks. Traditional CAN bus intrusion detection can only analyze grammatical features and cannot identify attacks where "legitimate instructions are executed in incorrect scenarios". This invention enables the system to have scenario perception and semantic understanding capabilities by sensing driving scenarios in real time and dynamically adjusting the strictness of semantic rules. It can intercept semantic-scenario contradictory instructions such as door unlocking during high-speed driving and unexpected steering when parking. At the same time, the linkage mechanism between semantic violations and physical fingerprint re-verification forms a two-way feedback closed loop.

[0024] (4) Introducing active challenge-response authentication to defend against adversarial learning attacks. Existing physical fingerprint authentication is a static, passive mechanism, which is vulnerable to generative adversarial network attacks. This invention is the first to combine physical fingerprint authentication with a challenge-response protocol, forcing attackers to simultaneously satisfy encryption correctness and physical fingerprint consistency within a very short time window through random challenge codes. Since physical layer features are hardware-unclonable, attackers cannot pass the verification through pure software simulation, fundamentally solving the adversarial learning threat.

[0025] (5) Constructing a multi-dimensional dynamic trust assessment and four-level response system. This invention integrates five dimensions of safety factors, adaptively adjusts weights based on driving scenarios, dynamically calculates the overall trust level, and corresponds to four levels of differentiated response strategies (from fully trustworthy to mandatory isolation). Compared with the existing technology's single "trustworthy / untrustworthy" binary decision, this invention provides a more granular, adaptive security control capability, maximizing the level of security protection while ensuring real-time performance.

[0026] (6) Achieve vehicle-edge collaborative regional defense to block the lateral spread of attacks. Existing solutions are mostly single-vehicle defenses, allowing attackers to breach each vehicle individually. This invention achieves cross-vehicle collaborative defense through differential privacy anonymization of threat intelligence, edge node aggregation analysis, consortium blockchain evidence storage, and regional collaborative defense command push. After breaching the first vehicle, attackers are blocked across the entire network and cannot spread laterally, significantly improving defense efficiency in large-scale attack scenarios.

[0027] (7) Optional active decoy module enhances countermeasure capabilities. By injecting camouflaged decoy messages and constructing a virtual sandbox ECU, the system can actively consume attacker resources, capture attacker fingerprints and implement countermeasures, upgrading passive defense to active confrontation, and further improving the overall security of the system.

[0028] In summary, this invention provides a multi-dimensional, collaborative, and proactive three-dimensional defense solution that extends from the physical layer, data link layer, application layer to the V2X communication layer. It can be widely applied to pre-installed safety systems for intelligent connected vehicles, aftermarket security hardware, and vehicle-road cooperative infrastructure, and has significant industrial application value. Attached Figure Description

[0029] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a block diagram of the overall architecture of the system of the present invention; Figure 2 This is a flowchart illustrating the workflow of the cross-modal cross-validation engine of this invention. Figure 3 This is a flowchart of the triple verification process for the motion physics constraint authentication module of the present invention; Figure 4 This is a schematic diagram of the structure of the scenario semantic firewall engine of the present invention; Figure 5 This is a timing diagram of the active challenge-response authentication module of the present invention; Figure 6 This is the state transition diagram for the four-level trust level of this invention; Figure 7 This is a flowchart of the collaborative response and trust management module of the present invention. Detailed Implementation

[0030] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0031] Example 1: System Hardware Deployment and Initialization 1.1 Hardware Platform Configuration In this embodiment, the intelligent connected vehicle multi-level authentication and active safety protection system is deployed in the vehicle's central gateway. The central gateway uses an NXPS32G series vehicle network processor, which integrates: four ARM Cortex-A53 cores (1.2GHz) for running computationally intensive tasks such as scene semantic reasoning and behavior analysis models; three ARM Cortex-M7 cores for real-time acquisition and filtering of CAN / LIN bus messages; a hardware security module supporting AES-128 / 256, SHA-256, and a true random number generator for storing physical fingerprint templates, device keys, and performing encryption operations; a high-speed analog-to-digital converter with a sampling rate of 10MSPS and a sampling precision of 12 bits for acquiring CAN_H / CAN_L differential voltage waveforms; a gigabit Ethernet switch supporting IEEE802.1AS time synchronization for in-vehicle inter-domain communication; and a 5G+V2X communication module supporting PC5 direct communication and Uu cellular networks. The vehicle CAN bus includes the powertrain CAN (500kbps), body CAN (250kbps), and infotainment CAN (500kbps), all of which are routed and isolated through a central gateway.

[0032] 1.2 Multimodal Physical Fingerprint Acquisition and Registration Perform fingerprint registration in a trusted environment during the vehicle off-line phase. Switch the central gateway to "fingerprint registration mode" and connect to a dedicated off-line diagnostic tool. The central gateway activates each electronic control unit (ECU) one by one (in functional domain order: powertrain domain → body domain → chassis domain → infotainment domain), requiring each ECU to send no less than 2000 standard CAN messages (the message content consists of preset test vectors, covering all commonly used CAN IDs).

[0033] For each received message, the central gateway simultaneously collects three types of features: Time feature acquisition: The arrival time of messages is recorded using a 1-microsecond resolution timer built into the hardware safety module, and the arrival time difference between two adjacent frames under the same CANID is calculated. For each CANID of each electronic control unit, 2000 arrival time difference samples are accumulated, a Gaussian mixture model is fitted, and the mean vector and covariance matrix are recorded.

[0034] Analog Feature Acquisition: The analog-to-digital converter is triggered to start sampling at the falling edge of the message start bit, with a sampling window length of 150 microseconds. The rise time (time to rise from 20% amplitude to 80% amplitude), fall time (time to fall from 80% amplitude to 20% amplitude), and overshoot amplitude (the percentage by which the peak amplitude exceeds the steady-state amplitude) are extracted from the differential voltage waveform. For each electronic control unit, the mean and standard deviation of these three features are calculated, forming a six-dimensional vector.

[0035] Clock skew feature acquisition: Record the timestamp (from the counter driven by the local crystal oscillator of the electronic control unit) embedded in the CAN message data field of the sending electronic control unit and the local timestamp of the central gateway, and calculate the relative clock skew. Collect 2000 skew samples for each electronic control unit, fit a linear regression model, and take the slope as the clock drift feature.

[0036] The three types of fingerprint vectors mentioned above are merged into a multimodal fingerprint template. The hash value of the fingerprint template is calculated using SHA-256 as the fingerprint ID. The original fingerprint template is encrypted using AES-256 and stored in the hardware security module of the central gateway (each electronic control unit occupies approximately 2KB of secure storage space). The fingerprint ID and the encrypted template copy are uploaded to the consortium blockchain of the edge trusted node for evidence storage via a TLS 1.3 secure connection, and are used for subsequent cross-vehicle fingerprint comparison and attack tracing.

[0037] 1.3 Configuration of Semantic Rule Base and Scene-Rule Mapping Table The system has a pre-built semantic rule base, which is stored in the flash memory of the central gateway in the form of a structured data table and supports remote OTA updates. The semantic rule base defines several rules for each electronic control unit. Each rule includes: the name of the electronic control unit to which it belongs, the CAN identifier, the operation type, the semantic constraints (represented by a logical expression), the violation severity level (level 1 to 5), and the maximum allowed frequency (unit: Hertz).

[0038] For example, the body control module includes the following rules: For the door unlocking operation with CAN identifier 0x2A1, the semantic constraint is that it is allowed when the gear is in P and the vehicle speed is zero, and prohibited when the vehicle speed is greater than zero. The violation severity level is the highest level, 3, and the maximum allowed frequency is 0.1 Hz. For the sunroof opening operation with CAN identifier 0x2A3, the semantic constraint is that it is allowed when the vehicle speed is below 20 km / h, and prohibited when the vehicle speed is above 20 km / h. The violation severity level is 2, and the maximum allowed frequency is 0.05 Hz.

[0039] The scenario-rule mapping table defines the severity coefficient (level 1 to 5) of each rule under different driving scenarios, as well as whether the rule is enabled. For example: Parking scenario: Door unlocking rule strictness level 1 (logging only), sunroof opening rule level 1 (allowed), diagnostic request rule level 5 (completely blocked).

[0040] Low-speed parking scenario: Door unlocking rule strictness level 3 (alarm but allowed), sunroof opening rule level 2 (allowed but frequency limited), diagnostic request rule level 5 (completely blocked).

[0041] In urban cruise scenarios: the door unlocking rule is strict at level 5 (complete interception), the sunroof opening rule is at level 4 (intercept and alarm), and the diagnostic request rule is at level 3 (query only).

[0042] High-speed driving scenario: The door unlocking rule and sunroof opening rule are both level 5 (completely blocked), and the diagnostic request rule is level 4 (blocked and alarmed).

[0043] Example 2: Implementation of a cross-modal cross-validation engine 2.1 Real-time extraction of dynamic physical fingerprint tokens When the central gateway receives a message from the CAN bus, it immediately triggers the cross-modal verification engine. The engine maintains a sliding window cache with the CAN identifier as the key, and each window contains physical layer sampled data of the 10 most recent messages from the same CAN identifier.

[0044] For the current message, the steps to extract the dynamic physical fingerprint token are as follows: Local temporal features: Calculate the arrival time difference between the current message and the previous message with the same identifier, and then calculate the mean and variance of the current window using the arrival time difference sequence of the aforementioned 10 messages to obtain the local temporal feature vector.

[0045] Local simulation features: The differential voltage waveform of the current message is acquired using an analog-to-digital converter, and the rise time, fall time, and overshoot amplitude are extracted to obtain a three-dimensional simulation feature vector.

[0046] Instantaneous clock offset: The instantaneous clock offset value is calculated based on the timestamp embedded in the electronic control unit in the current message and the local timestamp of the gateway.

[0047] The three vectors above are concatenated to form a dynamic fingerprint token.

[0048] 2.2 Calculation of Physical Layer Confidence Score The pre-stored multimodal fingerprint template of the target electronic control unit is read from the hardware security module, and the Mahalanobis distance between the dynamic fingerprint token and the template is calculated. The Mahalanobis distance is calculated as follows: first, the difference between the dynamic fingerprint token vector and the template mean vector is calculated; then, this difference is multiplied on the left by the inverse of the covariance matrix; then, it is multiplied on the left by the transpose of the difference; finally, the square root is taken. For simulated features, since each dimension is approximately independent, the covariance matrix is ​​simplified to a diagonal matrix.

[0049] Mahalanobis distance is mapped to a physical layer confidence score: the confidence score is equal to 1 minus (Mahalanobis distance divided by the maximum distance threshold), and the result is compared with 0, taking the maximum value. The maximum distance threshold is 3.0, indicating a built-in confidence score greater than 0 at three standard deviations; when the Mahalanobis distance is greater than 5.0, the confidence score is forcibly set to 0. The physical layer confidence score of a normal electronic control unit is typically between 0.85 and 0.98.

[0050] 2.3 Calculation of Behavioral Abnormality Scores The engine incorporates a behavior model based on isolated forests. Training phase: During a 100km road test after the vehicle rolls off the production line, the content features of all normal CAN messages are collected, including the CAN identifier, the high four bits and low four bits of the 8-byte data field are counted separately, and the interval of the message timestamp. 100 isolated trees are constructed, with a maximum depth of 8 for each tree.

[0051] During real-time scoring, the content features of the current message are input into the model, and the model outputs a behavioral anomaly score, ranging from 0 to 1, where 0 indicates highly consistent with normal behavior and 1 indicates highly abnormal behavior. During normal driving, the behavioral anomaly score is always less than 0.3; during attack injection, it is typically higher than 0.7.

[0052] 2.4 Cross-validation Logic The engine compares the physical layer confidence score with the behavioral anomaly score in real time and makes decisions based on preset thresholds. In this embodiment, the thresholds are set as follows: the first threshold is equal to 0.75 (lower limit of physical layer confidence), and the second threshold is equal to 0.35 (higher limit of behavioral anomaly).

[0053] If the physical layer confidence score is greater than or equal to 0.75 and the abnormal behavior score is less than or equal to 0.35, it is judged as a normal message and allowed to pass. At the same time, the sliding window is updated asynchronously with the fingerprint data of the current message (but the template is not changed).

[0054] If the physical layer confidence score is greater than or equal to 0.75 and the abnormal behavior score is greater than 0.35, it is determined that a legitimate node has been compromised and attacked. The engine immediately discards the current message, records the audit log, and calls the active challenge-response authentication module to perform forced physical fingerprint re-verification on the target electronic control unit.

[0055] If the physical layer confidence score is less than 0.75 and the abnormal behavior score is greater than 0.35, it is determined to be a spoofed node attack. The engine discards the message, records the attack source information (such as from the OBD-II port or a specific gateway port), and adds the source port to a temporary blacklist within 1 second.

[0056] If the physical layer confidence score is less than 0.75 and the abnormal behavior score is less than or equal to 0.35, it may be a newly added legitimate electronic control unit or a degraded template. The engine requests the latest fingerprint template of the electronic control unit from the edge trusted node. If it still does not match and occurs 5 times consecutively, it is determined to be an impersonation attack.

[0057] Example 3: Implementation of the Motion Physical Constraint Authentication Module 3.1 V2X Message Reception and Parsing The central gateway's V2X communication module receives basic security messages (compliant with SAE J2735 standard) broadcast by surrounding vehicles via the PC5 interface. Each basic security message includes: temporary ID, latitude and longitude (accurate to 0.1 meters), speed (0.01 m / s), heading angle (0.01 degrees), timestamp, vehicle type (passenger car, truck, motorcycle, etc.), and digital certificate signature. The module first performs a standard certificate signature verification (using the IEEE 1609.2 standard procedure). After successful signature verification, the message is sent to the motion physical constraint authentication module.

[0058] 3.2 Vehicle Parameter Package Acquisition The module reads the vehicle type field (e.g., "passenger vehicle") from the basic security message, and then sends a query request to the edge trusted node, carrying the vehicle type and current location. The edge trusted node returns the dynamic parameter envelope for the corresponding vehicle model. The envelope data includes: maximum acceleration (e.g., 4.0 m / s² for passenger vehicles), maximum deceleration (9.0 m / s²), maximum lateral acceleration (8.0 m / s², approximately 0.8g), maximum yaw rate (0.6 radians per second), and maximum reference speed. To improve real-time performance, the central gateway locally caches the parameter envelopes for the 20 most recently accessed vehicle models, with a cache validity period of 24 hours.

[0059] 3.3 Triple Physical Authenticity Verification For each basic security message, the module performs the following three checks in sequence: Verification 1: Compliance verification of high-precision map location The central gateway pre-stores a local high-precision map of the current area, covering a range of 500 meters in front of the vehicle, and pushes it out on demand from edge nodes. The map includes vector data of lane boundaries, curbs, medians, and buildings. The module converts the latitude and longitude coordinates in the basic safety message to the map coordinate system and uses ray casting to determine whether the point is located within any drivable area (lane, shoulder, legal parking space). If it is located inside a building, outside a guardrail, or in a lake or river, it is considered an illegal location and is directly marked as a ghost vehicle.

[0060] Verification 2: Vehicle dynamics motion consistency verification The module maintains a historical basic security message cache with a temporary ID as the key, storing a maximum of the most recent 10 frames. For vehicles not appearing for the first time, it reads the speed values ​​of the previous two frames (100 milliseconds apart) and calculates the actual acceleration (current speed minus the previous speed, then divided by the time interval). Simultaneously, it calculates the yaw rate based on the heading angle change (current heading angle minus the previous heading angle, then divided by the time interval). If the actual acceleration is greater than the maximum acceleration or less than the negative maximum deceleration, or if the absolute value of the yaw rate is greater than the maximum yaw rate, the motion consistency check is considered to have failed. A 5% tolerance is allowed, considering GPS / IMU measurement noise.

[0061] Verification 3: Historical Trajectory Prediction Verification Using a uniform motion model or a uniform acceleration model, Kalman filtering is applied to the positions of the most recent 5 frames to predict the expected position and covariance of the current frame. A 95% confidence ellipse is constructed, satisfying that the transpose of (actual position vector minus predicted position vector) multiplied by the inverse of the covariance matrix, and then multiplied by (actual position vector minus predicted position vector), is less than or equal to 5.99. If the actual reported position falls outside the ellipse, the prediction verification is considered a failure. In this embodiment, twice the major semi-axis of the confidence ellipse is used as a simplified threshold.

[0062] Integrated Decision: If any of the three checks fails, the module immediately performs the following actions: Records the temporary ID in the local blacklist, valid for 300 seconds; generates a warning message (containing the attacking vehicle's temporary ID, location, timestamp, and failure type), broadcasts it to surrounding vehicles via V2X, sets the message priority to the highest, and uses the dedicated message type 0x04; and reports the ghost vehicle event to trusted edge nodes for regional joint defense.

[0063] Example 4: Implementation of a Scene Semantic Firewall Engine 4.1 Real-time recognition of driving scenarios The perception unit of the scene semantic firewall engine reads data from the following sensors at a frequency of 10 Hz (acquired via CAN bus or vehicle Ethernet): gear position (from the body control module or transmission control unit), vehicle speed (from the fusion value of ABS wheel speed sensors), steering wheel angle and angular velocity (from the electric power steering system), brake master cylinder pressure (from the electronic stability program), accelerator pedal opening (from the engine control module), and positioning and attitude provided by GPS / IMU integrated navigation.

[0064] Decision trees are used for scene classification, and the classification rules are as follows: If the gear is in P and the vehicle speed is zero, it is determined to be a parking scenario.

[0065] Otherwise, if the vehicle speed is below 10 km / h and the absolute value of the steering wheel angular velocity is greater than 30 degrees per second, it is judged as a low-speed parking scenario.

[0066] Otherwise, if the vehicle speed is greater than or equal to 80 km / h, it is considered a high-speed driving scenario.

[0067] Otherwise, it will be classified as an urban patrol scenario.

[0068] 4.2 Dynamic Adjustment of Semantic Rules Based on the identified scenario, the engine queries the scenario-rule mapping table to determine the severity level of each rule in the current scenario. For each received CAN message, the engine performs the following semantic checks: The corresponding semantic rule is found based on the CAN identifier of the message (if no rule is found, the default security policy is used: allow but log).

[0069] Check if the rule is allowed in the current scenario: if the mapping table is marked as "completely blocked", then discard the message directly and record the semantic violation.

[0070] If frequency limiting is required but permitted, check the sending frequency of the message identifier of the electronic control unit in this scenario: use the token bucket algorithm, if the maximum permitted frequency is exceeded, discard the token and record the violation.

[0071] If allowed and the frequency does not exceed the limit, further check the semantic constraints (e.g., "gear is in P and vehicle speed is zero"): jointly evaluate the current scene information and message data field; if the condition is true, allow passage; otherwise, discard and record the violation.

[0072] For example, in a high-speed driving scenario (vehicle speed 100 km / h), the body control module sends a door unlock message (CAN identifier 0x2A1, data 0x01). The engine rule lookup finds that the rule has a strictness of 5 (complete interception), so it is discarded directly without proceeding to subsequent condition checks. At the same time, the semantic anomaly integrator adds 0.5 points to the body control module.

[0073] 4.3 Semantic Anomaly Integral and Linkage The semantic anomaly integrator maintains an integral variable for each electronic control unit (ECU), initially set to 0. Each time a semantic violation occurs, the integral is incremented, with increments of 0.1, 0.2, 0.5, 0.8, and 1.0 corresponding to violation severity levels 1 to 5, respectively. The integral is updated exponentially, decaying to 0.5 times its original value every 60 seconds (i.e., a half-life of 60 seconds). When the integral reaches or exceeds 3.0, a trust level downgrade is triggered: the trust level of the ECU is reduced by one level (L0 to L1, L1 to L2), and a semantic anomaly signal is sent to the cross-modal cross-validation engine, triggering mandatory physical fingerprint re-verification.

[0074] Example 5: Implementation of the Proactive Challenge-Response Authentication Module 5.1 Triggering Condition Implementation In this embodiment, the proactive challenge-response authentication module runs on the M7 real-time core of the central gateway to avoid communication blockage. The triggering logic is as follows: Periodic Challenges: Every 10 minutes, a challenge is initiated on each of the electronic control units (engine management system, transmission control unit) on the powertrain CAN and the electronic control units (electronic stability program, electric power steering system) on the chassis CAN. The challenge order is randomized to avoid predictability.

[0075] Trust level downgrade trigger: When the collaborative response module downgrades the trust level of an electronic control unit from L0 to L1, it launches 3 consecutive challenges to the electronic control unit, each with an interval of 5 milliseconds and using different random numbers.

[0076] Semantic anomaly trigger: When a semantic anomaly signal is received from the scene semantic firewall, an immediate challenge (single) is launched against the non-compliant electronic control unit.

[0077] Edge command triggering: When the 5G module of the central gateway receives threat intelligence (such as "a certain supplier's body control module firmware has a backdoor") from the edge trusted node, it will launch a batch challenge (single) on all electronic control units of body control module types in the vehicle.

[0078] 5.2 Challenge-Response Protocol Detailed Process Taking the challenge of the electronic control unit of the electric power steering system as an example, the process is as follows: Challenge generation: The hardware security module of the central gateway calls a true random number generator to generate a 128-bit random number as the challenge code. The current timestamp is recorded with a precision of 1 microsecond.

[0079] Sending the challenge: Construct a private CAN message using the reserved identifier 0x7E0. The first two bytes of the data field are the command code 0x55AA, the following 16 bytes are the challenge code (random number), and the last 4 bytes are the lower 32 bits of the timestamp. Send the message to the electronic control unit of the electric power steering system via the CAN bus.

[0080] Response Received: Upon receiving the challenge, the electronic control unit (ECU) of the electric power steering system must use its pre-set device-unique key (AES-128) in its hardware security module to encrypt the concatenated data (challenge code random number and timestamp), generating a 16-byte ciphertext. The ECU returns the ciphertext via a CAN message (identifier 0x7E8) within 5 milliseconds. If no response is received within this timeout, the gateway considers the challenge a failure.

[0081] Synchronous acquisition of physical characteristics: During the waiting period for response, the analog-to-digital converter of the central gateway continuously samples. When the falling edge of the start bit of the 0x7E8 message is detected, high-precision acquisition is triggered: the differential voltage waveform of the message is recorded, the rise time, fall time, and overshoot amplitude are extracted, and the clock offset is calculated based on the timestamp embedded in the message.

[0082] Dual authentication: For encryption authentication, the gateway uses a locally stored unique key for the same device to encrypt the same data (challenge code random number and timestamp) using AES-128 to obtain the desired ciphertext. The returned ciphertext is compared with the desired ciphertext; if they match exactly, the encryption authentication passes; otherwise, it fails. For physical fingerprint authentication, the collected physical layer feature vector is compared with the pre-stored electric power steering system fingerprint template in the hardware security module to calculate the Mahalanobis distance. If the Mahalanobis distance is less than or equal to 3.0 (corresponding to a matching degree greater than or equal to 0.85), the physical authentication passes; otherwise, it fails.

[0083] Judgment and Response: If both encryption and physical verification pass, maintain the current trust level and update the recent challenge response pass rate of the electronic control unit (number of passes divided by the total number of challenges, sliding window 5 times). If either verification fails, immediately lower the trust level of the electronic control unit to L2 or L3 (depending on the number of failures and severity), and perform isolation: disconnect the routing of the CAN network segment where the electronic control unit is located from other domains, record the attack log, and report it to the edge trusted node.

[0084] In this embodiment, the average challenge-response time for a normal electric power steering system is approximately 4.2 milliseconds (including encryption calculations, CAN transmission, and fingerprint acquisition), fully meeting the real-time requirements. For malicious responses simulated by attackers, the challenge failure rate is close to 100% due to the difficulty in simultaneously replicating the encryption key and physical layer characteristics.

[0085] Example 6: Implementation of the Collaborative Response and Trust Management Module 6.1 Dynamic Calculation of Comprehensive Trust Level The collaborative response and trust management module runs every 100 milliseconds to calculate the overall trust level for the entire vehicle and each electronic control unit. This embodiment uses a weighted linear fusion method (neural networks or fuzzy logic can also be used, but linear is used as an example for simplicity).

[0086] The methods for obtaining multi-dimensional factors are as follows: Physical layer confidence score: The moving average of the physical layer confidence scores of all messages from this electronic control unit in the past 5 seconds (excluding dropped messages), with an initial value of 0.95.

[0087] Abnormal Behavior Score: Take the maximum value of the abnormal behavior score within the past 5 seconds.

[0088] Motion constraint pass rate: For the overall vehicle trust level, the percentage of all V2X basic safety messages received in the past minute that pass the motion physical constraint test is taken; for a single electronic control unit (such as a V2X communication module), the pass rate of the module's most recent 10 verifications is taken.

[0089] Semantic violation severity: The normalized value of the semantic violation integral, which is the current integral divided by the isolation threshold of 3.0, but not exceeding 1.0.

[0090] Challenge response pass rate: The pass rate of the last 5 challenges (initially 1.0).

[0091] The weight vector is dynamically adjusted based on the current driving scenario. The weights set in this embodiment are as follows: Parking scenario: Physical layer weight 0.15, abnormal behavior weight 0.15, motion constraint weight 0.10, semantic violation weight 0.35, challenge response weight 0.25.

[0092] Low-speed parking scenario: Physical layer weight 0.20, abnormal behavior weight 0.20, motion constraint weight 0.10, semantic violation weight 0.30, challenge response weight 0.20.

[0093] Urban patrol scenario: Physical layer weight 0.25, abnormal behavior weight 0.30, motion constraint weight 0.15, semantic violation weight 0.20, challenge response weight 0.10.

[0094] High-speed driving scenario: Physical layer weight 0.20, abnormal behavior weight 0.15, motion constraint weight 0.35, semantic violation weight 0.20, challenge response weight 0.10.

[0095] The formula for calculating the overall trust level is: physical layer confidence score multiplied by the corresponding weight, plus (1 minus the behavioral anomaly score) multiplied by the corresponding weight, plus motion constraint pass rate multiplied by the corresponding weight, plus (1 minus the semantic violation degree) multiplied by the corresponding weight, plus challenge response pass rate multiplied by the corresponding weight.

[0096] For example, in a high-speed driving scenario, the physical layer confidence score of an electronic control unit is 0.92, the behavior anomaly score is 0.45, the motion constraint pass rate is 1.0, the semantic violation level is 0.1, and the challenge response pass rate is 0.8. The calculation process is: 0.20×0.92+0.15×(1-0.45)+0.35×1.0+0.20×(1-0.1)+0.10×0.8=0.184+0.0825+0.35+0.18+0.08=0.8765.

[0097] According to the four-level trust system: If the overall trust level is greater than or equal to 0.85, maintain level L0 (fully trustworthy).

[0098] If it is between 0.60 and 0.85, it is downgraded to L1 (mildly suspicious).

[0099] If it is between 0.30 and 0.60, it is downgraded to L2 (highly suspicious).

[0100] If it falls below 0.30, it will be downgraded to Level L3 (malicious confirmation).

[0101] In the example above, 0.8765 belongs to level L0 and does not need to be downgraded.

[0102] 6.2 Trust Level-Based Response Strategy The collaborative response module implements differentiated strategies based on the overall vehicle trust level and the individual trust level of each electronic control unit: Level 0 (Fully Trusted): All CAN messages are routed normally, V2X broadcasts are not subject to additional restrictions, and the access control list of the Ethernet switch remains in its default configuration (allowing entertainment systems to access remote information processing terminals).

[0103] Level 1 (Mild Suspicious): In the gateway's CAN routing table, modify the transmission permissions of the network segment containing the suspicious electronic control unit (such as infotainment CAN) to allow only security-related messages (via a predefined whitelist of CAN identifiers). Restrict the telematics terminal from sending data (only retain emergency call functionality). Simultaneously increase the challenge frequency to once every 2 minutes.

[0104] Level 2 (Highly Suspicious): Disconnect the network segment containing the suspected electronic control unit from all other network segments (i.e., isolate the network segment), but retain direct communication between the electronic control unit and basic safety functions such as braking and steering (e.g., the dedicated CAN link between the electronic stability program and the electric power steering system does not pass through a gateway). Prohibit the vehicle from broadcasting basic safety messages (to prevent the spread of false information) and only receive external V2X messages.

[0105] Level 3 (Malicious Confirmation): Physically disables the telematics terminal port via the vehicle's Ethernet switch, and turns off the power to the Wi-Fi and Bluetooth modules; sends a speed limit command via the CAN bus (limits torque output via the engine management system, with a maximum speed of 20 km / h), while displaying "Serious system malfunction, please stop safely immediately" on the instrument panel; broadcasts an emergency warning via V2X: "This vehicle is under attack, please take evasive action" (using the dedicated security message type 0xFF).

[0106] 6.3 Edge Collaborative Defense Vehicles establish secure connections with trusted edge nodes via 5G cellular networks (using the TLS 1.3 protocol with two-way authentication). The data format for threat intelligence reporting includes: anonymized vehicle identifier (processed with differential privacy, retaining the real IP address with a probability of 0.7, otherwise outputting a random value), timestamp, threat type (e.g., ghost vehicle attack), attack source information (IP address, temporary ID, GPS coordinates obfuscated to a 50-meter grid), attack behavior pattern (e.g., CAN identifier sequence from diagnostic scans), and confidence score.

[0107] The edge trusted node aggregates reports from multiple vehicles. When the same attack source is reported by more than three vehicles within 5 minutes, it automatically generates a regional joint defense command, which is broadcast to all vehicles within the coverage area via PC5. The joint defense command includes a blacklist of IP addresses, malicious CAN identifier patterns, and suggested firewall rules. Upon receiving the command, the gateway of the receiving vehicle immediately loads these rules, achieving global blocking within seconds.

[0108] Example 7: Active Trapping and Attack Countermeasure Module In this embodiment, the system may optionally deploy an active decoy and attack countermeasure module. This module runs a lightweight virtualization container (based on container technology) on the Linux system of the central gateway. The container simulates the behavior of a complete electronic control unit, including virtual diagnostic services (0x27 service) and virtual body control services. This container is not connected to the real actuators, and its output CAN messages are used only as decoys.

[0109] 7.1 Decoy Injection Strategy Periodic decoy: Every 5 seconds, a "fake diagnostic session open" message is broadcast on the vehicle Ethernet (the destination IP is the IP of the telematics terminal); at the same time, a fake "autonomous driving mode status word" is sent on the CAN bus with low priority (using the reserved identifier 0x7FF), with the content "Level 3 autonomous driving is activated".

[0110] V2X Decoy: A basic security message from a ghost vehicle with a plausible trajectory is broadcast via a V2X module. The ghost vehicle is positioned 50 meters ahead of the real vehicle, traveling at the same speed to simulate a following scenario. The digital certificate for this decoy's basic security message uses a self-signed test certificate from the vehicle, which would normally not be accepted by other vehicles' V2X systems. However, attackers who have disabled certificate verification will fall into the trap.

[0111] 7.2 Attack Capture and Countermeasures When an attacker (such as a hacker who infiltrates via OBD-II) sends a diagnostic request (e.g., a 0x2701 request seed) to the aforementioned decoy diagnostic service, the virtual service within the container responds with a seed (random number) according to the standard diagnostic protocol. If the attacker continues to send keys, the virtual service consistently returns "verification failed," but records all of the attacker's request messages, timestamps, and source IPs. Simultaneously, the container intentionally delays its response or returns erroneous data to consume the attacker's resources.

[0112] Upon detecting an attack, the module performs the following actions: Add the attacking IP to the local blacklist, and the firewall will immediately block all subsequent connections from that IP.

[0113] Generate a detailed attack report (including the attacker's request sequence and the characteristics of the attack tools used) and report it to the trusted edge node.

[0114] The system continues to interact with the attacker, luring them into a more complex "virtual sandbox." In this sandbox, the attacker believes they have compromised a real electronic control unit, but in reality, all control commands are discarded, while the system continuously collects the attacker's behavioral patterns for subsequent analysis.

[0115] In this embodiment, the virtual sandbox electronic control unit is built based on a simulator and runs a modified version of the firmware of the real electronic control unit (all output write operations are redirected to a null device). In this way, the system can study attackers' techniques without risk and generate more effective defensive signatures.

[0116] Example 8: Complete Workflow Example The following example illustrates the collaborative work of the modules using a complete process from vehicle startup to being attacked and successfully defended.

[0117] Timeline: t=0 seconds: Vehicle powers on. After all electronic control units (ECUs) perform their power-on self-tests, the central gateway requests each ECU to report a dynamic physical fingerprint token (as initial authentication). The gateway verifies the token and matches it with a pre-stored template, assigning an initial trust level L0 to each ECU. The scene semantic firewall initializes, gear is P, vehicle speed is 0, and the scene is identified as a "parking scenario".

[0118] t=10 seconds: The driver shifts into Drive (D) and the vehicle begins driving on city roads. Scene recognition switches to "City Cruise". At this time, the semantic firewall enables city cruise rules: prohibiting door unlocking, prohibiting diagnostic flashing, etc.

[0119] t=120 seconds: The hacker successfully compromised the vehicle control module through a remote vulnerability (such as a buffer overflow in a telematics terminal) and gained control of the module. The hacker sent a door unlock command (CAN identifier 0x2A1, data 0x01).

[0120] At t=120.005 seconds: The central gateway receives the message. The cross-modal cross-validation engine extracts the physical fingerprint, with a physical layer confidence score of 0.93 (matching the pre-stored template in the vehicle control module). The behavior analysis model calculates an anomaly score of 0.87 (because there has never been a door unlocking behavior at a vehicle speed of 56 km / h). The engine determines that the legitimate node has been compromised and immediately discards the message, triggering mandatory physical fingerprint re-verification.

[0121] t=120.010 seconds: The active challenge-response module sends a random challenge code to the body control module. Due to firmware tampering, the body control module is unable to return a correct encrypted response within 5 milliseconds (or although it returns a correct response, the physical characteristics of the response message are distorted due to malicious code interference). The challenge fails.

[0122] At t=120.020 seconds: The coordinated response module lowers the trust level of the body control module from L0 to L2 and performs isolation: cutting off the routing between the body CAN network segment and the powertrain CAN and chassis CAN, while simultaneously displaying a "Body System Abnormality" message on the instrument panel. The body control module can no longer affect the safe operation of the vehicle.

[0123] At t=120.030 seconds: The trusted edge node receives threat intelligence from its own vehicle and three other vehicles in the same city (all reporting the same IP address of the attack source). The edge node generates a joint defense command and broadcasts it to all vehicles in the area.

[0124] t=120.500 seconds: The hacker attempted to attack another vehicle of the same model, but his attack request was directly blocked by the pre-loaded blacklist of the vehicle's gateway, and a connection could not be established.

[0125] Through the above process, the system of the present invention successfully defended against a remote control attack targeting the vehicle body control module, and prevented the lateral spread of the attack through edge collaboration.

[0126] Example 9: Quantitative Evaluation of Implementation Results To verify the practical effectiveness of the invention, the applicant built a test scenario containing 20 intelligent connected vehicles in a simulation environment (based on CANoe, CarMaker, and OMNeT++ co-simulation) and simulated the following three attack types: Impersonation attack: The attacker injects a fake braking command (CAN identifier 0x220, data field all 1s) through the OBD-II port.

[0127] Legitimate nodes are compromised through vulnerability injection, which modifies the firmware of the vehicle control module, causing it to send door unlock signals while the vehicle is traveling at high speed.

[0128] Ghost car attack: Using software-defined radio devices to broadcast fake basic security messages claiming that there is a stationary vehicle 50 meters ahead.

[0129] The test results are as follows: Impersonation attack: Existing solutions (signature-only plus simple intrusion detection system) have a detection rate of 76%, while the detection rate of the system of this invention is 100%, with an average response time of 12 milliseconds and a false alarm rate of 0.3%.

[0130] Legitimate nodes are targeted by attacks: Existing solutions have a detection rate of 0%, while the detection rate of this invention's system is 98.5%, with an average response time of 28 milliseconds and a false alarm rate of 1.2%.

[0131] Ghost car attack: Existing solutions have a detection rate of 0%, while the detection rate of this invention's system is 99.2%, with an average response time of 45 milliseconds and a false alarm rate of 0.8%.

[0132] During a continuous 72-hour stability test, the system of this invention did not crash or restart, with an average CPU load of 32% (central gateway S32G) and memory usage of approximately 180 megabytes, meeting the requirements of the automotive embedded environment.

[0133] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A multi-level authentication and active safety protection system for intelligent connected vehicles, characterized in that, include: The multimodal physical fingerprint acquisition module is used to acquire the multimodal physical fingerprints of each electronic control unit during the vehicle off-line stage. The multimodal physical fingerprints include the arrival time difference characteristics between CAN message frames, the timing characteristics of the CAN bus differential voltage, and the relative clock offset characteristics of the electronic control unit. The multimodal physical fingerprint is encrypted by hash and stored in the vehicle-mounted safety element and the edge trusted node. The edge trusted node is a distributed computing node deployed on the roadside or base station side and has blockchain evidence storage capability. The cross-modal cross-validation engine, connected to the vehicle's CAN bus, is configured as follows: The dynamic physical fingerprint token of the CAN message is extracted in real time and compared with the pre-stored multimodal physical fingerprint to generate a physical layer confidence score; Generate behavioral anomaly scores based on CAN message content features; When the physical layer confidence score is higher than the first threshold and the abnormal behavior score is higher than the second threshold, it is determined that the legitimate node is being controlled by an attack, and forced physical fingerprint re-verification is triggered. When the physical layer confidence score is lower than the first threshold and the abnormal behavior score is higher than the second threshold, it is determined to be a fake node attack. The motion physics constraint authentication module is connected to the V2X communication module and is configured to perform physical authenticity verification on the received V2X basic safety messages based on high-precision map location compliance verification, vehicle dynamics model motion consistency verification, and historical trajectory prediction verification. If the verification fails, the module will mark the vehicle as a ghost vehicle and broadcast a risk warning. The scene semantic firewall engine is configured to integrate multi-source sensor information to identify the current driving scene of the vehicle in real time. Based on the pre-configured semantic rule base and scene-rule mapping table, it dynamically adjusts the strictness of the operation semantic rules of each electronic control unit according to the driving scene. When a semantic violation is detected, it outputs a semantic abnormality signal and triggers forced physical fingerprint re-verification. The active challenge-response authentication module is configured to send a random challenge code to the target electronic control unit according to a preset triggering mechanism, receive its encrypted response and synchronously collect the physical layer characteristics of the response message, and maintain the current trust level only when the encrypted response is correct and the physical layer characteristics match the pre-stored multimodal physical fingerprint; otherwise, the trust level is reduced. The collaborative response and trust management module is connected to the above modules respectively. It is configured to integrate multiple dimensions such as physical layer confidence, abnormal behavior, motion constraint pass rate, semantic violation degree, and challenge response pass rate, and adaptively adjust the weight of each dimension in combination with the current driving scenario to dynamically calculate the comprehensive trust level. Based on the comprehensive trust level, it adjusts the in-vehicle communication routing strategy, external network access control strategy, and V2X broadcast strategy, and realizes collaborative threat defense with trusted edge nodes.

2. The system according to claim 1, characterized in that, The forced physical fingerprint re-verification is triggered by the cross-modal cross-verification engine and executed by the proactive challenge-response authentication module, including: The cross-modal cross-validation engine sends random challenge values ​​to the electronic control unit to be validated. The electronic control unit to be verified encrypts the random challenge value based on the pre-shared key and returns a response message; The active challenge-response authentication module synchronously collects the physical layer characteristics of the response message, including differential voltage timing characteristics and clock offset characteristics. When the encrypted response verification fails, or when the deviation between the physical layer feature and the corresponding feature in the pre-stored multimodal physical fingerprint exceeds a preset threshold, the re-verification is determined to have failed, the electronic control unit is isolated, and the issue is reported to the edge trusted node.

3. The system according to claim 1, characterized in that, The motion physics constraint authentication module includes a vehicle model parameter matching submodule; the vehicle model parameter matching submodule is used to obtain the dynamic parameter envelope of the corresponding vehicle model from the edge trusted node or the cloud according to the vehicle type identifier in the V2X basic safety message, and the dynamic parameter envelope includes at least the allowable range of maximum acceleration, maximum deceleration, maximum lateral acceleration, and maximum yaw rate; The triple verification performed by the motion physics constraint authentication module is as follows: High-precision map location compliance verification: Verify that the GPS coordinates in the V2X message are within the drivable area; Vehicle dynamics motion consistency verification: Calculate the actual acceleration and yaw rate based on multiple consecutive V2X messages, and determine whether they exceed the envelope of the dynamic parameters; Historical trajectory prediction verification: Based on the position and velocity of the previous N frames, a Kalman filter is used to predict the expected position interval of the current frame, and it is determined whether the reported position falls within the expected position interval.

4. The system according to claim 1, characterized in that, The scenario semantic firewall engine includes a semantic rule dynamic adjustment submodule and a semantic anomaly integrator. The semantic rule dynamic adjustment submodule is used to configure the semantic rule strictness of each electronic control unit's corresponding control commands, diagnostic requests, and data interaction behaviors differently based on the scenario-rule mapping table according to the attributes of parking scenarios, low-speed parking scenarios, urban cruising scenarios, and high-speed driving scenarios. The semantic anomaly integrator is used to accumulate the number of semantic violations of each electronic control unit. When the accumulated score reaches a dynamic threshold, it triggers the trust level downgrade of the corresponding electronic control unit and sends a semantic anomaly signal to the cross-modal cross-validation engine.

5. The system according to claim 1, characterized in that, The triggering mechanism of the active challenge-response authentication module includes: Periodic triggering: A challenge is launched against all safety-critical electronic control units every preset time window; Trust level downgrade trigger: When the trust level of any electronic control unit is downgraded from the current level, multiple consecutive challenges will be launched against that electronic control unit; Semantic anomaly triggering: When the scene semantic firewall engine detects a semantic violation, it initiates an immediate challenge against the violating electronic control unit; Edge Trusted Node Command Trigger: When threat intelligence is received from an edge trusted node, a batch challenge is launched against all electronic control units of the specified type.

6. The system according to claim 1, characterized in that, The collaborative response and trust management module is configured with a four-level comprehensive trust level system: Level 0: Fully trusted, allowing all in-vehicle and V2X communication operations; Level 1: Mildly suspicious, restrict non-essential remote commands, allow only security-critical operations, and trigger enhanced monitoring of trusted edge nodes; Level 2: Highly suspicious, only basic safety functions such as braking and steering are permitted, and the infotainment system uplink and non-safety-related V2X broadcasts are cut off; Level 3: Malicious confirmation, actively cutting off all connections between the vehicle and the external network, forcing the vehicle to slow down and prompting a safe stop; Different comprehensive trust levels correspond to different in-vehicle CAN routing permissions, in-vehicle Ethernet access control permissions, and V2X message broadcasting policies.

7. The system according to claim 6, characterized in that, The comprehensive trust level is calculated by fusing multiple dimensions, including physical layer confidence, degree of behavioral anomaly, motion constraint pass rate, semantic violation degree, and challenge response pass rate. The fusion method is weighted fusion, nonlinear mapping, or a trained machine learning model, wherein the weights or mapping parameters of each dimension factor are adaptively adjusted according to the current driving scenario.

8. The system according to claim 1, characterized in that, It also includes an active decoy and attack countermeasure module; the active decoy and attack countermeasure module is used to periodically send spoofed decoy messages in the vehicle Ethernet and V2X communication links. The decoy messages include virtual diagnostic session responses, fake autonomous driving mode status words, and basic safety messages of ghost vehicles with reasonable movement trajectories. When an external entity attempts to read, modify, or replay the decoy messages, the network layer characteristics and operational behavior characteristics of the attack source are collected, added to the blacklist feature library, and a deceptive response is sent to the attack source to lure it into the virtual sandbox electronic control unit built based on software virtualization technology. The anonymized threat characteristics are also synchronized to edge trusted nodes and other vehicles in the fleet.

9. A multi-level authentication and active safety protection method for intelligent connected vehicles, applied to the system described in any one of claims 1 to 8, characterized in that, Includes the following steps: S1. After the vehicle is powered on and started, each electronic control unit reports a dynamic physical fingerprint token to the cross-modal cross-validation engine to complete the initial identity authentication and initialize the trust level. S2. During vehicle operation, multimodal physical layer features and content features of CAN bus messages are collected in real time, and physical layer confidence scores and behavior anomaly scores are dynamically calculated. Based on the matching relationship between the two, fake node attacks and legitimate node control attacks are identified. S3. For the received external vehicle V2X basic security messages, perform multi-dimensional physical authenticity verification based on high-precision maps, vehicle dynamics models and historical trajectory predictions, identify ghost car attacks and broadcast risk warnings to the outside world. S4. Integrates data from multiple onboard sensors to identify vehicle driving scenarios in real time. Based on a pre-configured semantic rule base and scenario-rule mapping table, it dynamically updates the semantic operation whitelist and rule strictness of each electronic control unit to detect business semantic violations. S5. When a legitimate node is determined to be under attack, a semantic violation is detected, or the physical layer confidence score deviates from the behavioral anomaly score, a mandatory physical fingerprint re-verification is triggered to perform a second identity verification on the suspicious electronic control unit. S6. The proactive challenge-response authentication module performs proactive authentication on the target electronic control unit based on a periodic mechanism or security event triggering, and dynamically updates the device trust status. S7, the collaborative response and trust management module integrates multi-dimensional security data, adaptively adjusts weights based on the current driving scenario, dynamically calculates the comprehensive trust level, and adjusts in-vehicle communication routing, external network access and vehicle control strategies according to the comprehensive trust level; S8. Synchronize local anonymized threat data to edge trusted nodes to trigger edge trusted nodes to execute regional joint defense strategies, which include pushing blacklists and defense rules to other vehicles in the region.

10. The method according to claim 9, characterized in that, The triggering conditions for forced physical fingerprint re-verification in step S5 also include: The same electronic control unit commits a set number of semantic violations more than a set number within a preset time window; The absolute value of the difference between the physical layer confidence score and the behavioral abnormality score of the same electronic control unit is greater than the preset deviation threshold; Received a threat intelligence command for this type of electronic control unit from an edge trusted node.

Images