Dynamic test strategy generation method for automobile network and data security

Abstract

The invention relates to the technical field of intelligent networked automobile network security, in particular to a dynamic test strategy generation method for automobile network and data security, which comprises the following steps of: acquiring a vehicle-mounted CAN bus message containing a CAN ID, a timestamp and effective load data, performing dynamic segmentation processing by adopting a self-adaptive sliding window mechanism, and generating a dynamic test strategy for the automobile network and data security; a continuous feature vector is generated in combination with semantic coding with an attention mechanism; inputting the feature vector into a bimodal deep neural network, and dynamically adjusting network weight according to a real-time threat level to generate a test strategy; finally, efficient utilization of computing resources is achieved through a hierarchical optimization mechanism, and network parameters are updated in an incremental learning mode. The technical bottlenecks of poor static pruning adaptability, low model reasoning efficiency, high-frequency scene response delay and the like in the background technology are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent networked vehicle network security technology, and particularly relates to a dynamic test strategy generation method for vehicle network and data security. BACKGROUND

[0002] With the rapid development of intelligent networked vehicles, vehicle network and data security are facing increasingly severe challenges. Traditional vehicle network security test strategies usually use static rules or predefined test cases, which are difficult to adapt to dynamically changing attack methods and complex vehicle communication environments. Existing methods mainly rely on manual experience or fixed detection models, resulting in low test coverage, high false positive rate, and inability to effectively capture long-term dependencies in CAN bus data. In addition, due to the diversity of vehicle network protocols and the continuous evolution of attack patterns, traditional methods are insufficient in unknown attack detection and cross-model adaptability, severely restricting the effectiveness of vehicle network security protection.

[0003] The existing patent (CN118282780B) proposes a new energy vehicle vehicle network intrusion detection method, device and storage medium, which significantly improves the adaptability and accuracy of the detection model through dynamic gradient proxy strategy and semantic encoding technology. The core improvements of this patent include: dynamic gradient proxy strategy: intelligently selects the optimal state input LSTM from the historical state set, enhances the robustness of the model to abnormal data, and reduces false positives. Word2Vec semantic encoding: maps discrete CANID and payload data to continuous vectors, effectively capturing semantic relationships between data and improving feature expression capabilities. Improved LSTM structure: optimizes the modeling ability of long-term dependencies, making it more suitable for time series data analysis of vehicle CAN networks. This scheme has made significant progress in the accuracy and adaptability of intrusion detection, providing important technical support for the generation of dynamic test strategies.

[0004] Although the above method has made significant progress in the accuracy and adaptability of intrusion detection, its real-time performance and lightweight still have room for improvement. The current pruning strategy is based on static thresholds and cannot dynamically adjust the computational load according to real-time data flow, which may cause response delays in high-frequency attack scenarios. In addition, the inference efficiency of the compressed model is still limited by the performance of the vehicle hardware, and further optimization is needed to adapt to more stringent real-time requirements. Therefore, there is an urgent need for a more efficient dynamic test strategy generation method to address the shortcomings of existing technologies. SUMMARY

[0005] The purpose of this invention is to provide a dynamic test strategy generation method for automotive network and data security, which solves the problems that existing methods still have room for optimization in terms of real-time performance, the static pruning strategy used cannot dynamically adjust the computational load according to the real-time data stream, may have response delays in high-concurrency attack scenarios, and the computational efficiency after model compression is still limited by the performance of the vehicle hardware.

[0006] To achieve the above objectives, this invention provides a method for generating dynamic testing strategies for automotive network and data security, comprising the following steps: Collect vehicle CAN bus message data; Preprocess the CAN bus message data to generate continuous feature vectors; Continuous feature vectors are input into a dual-modal deep neural network for processing to generate a test strategy; The computational resource utilization of the bimodal deep neural network is monitored in real time, and the bimodal deep neural network is dynamically optimized when the computational load exceeds a preset threshold.

[0007] The collection of vehicle CAN bus message data specifically includes: CAN bus message data includes CAN ID, timestamp, and payload data.

[0008] The preprocessing of CAN bus message data to generate continuous feature vectors includes the following steps: A sliding window mechanism is used to segment CAN bus message data, dynamically adjust the window size according to the current network traffic, and semantically encode the CAN ID and payload data to generate continuous feature vectors. The sliding window mechanism is implemented through FPGA hardware acceleration and includes parallel window adjustment logic and message sorting buffer.

[0009] The continuous feature vectors are input into a dual-modal deep neural network for processing to generate a test strategy, specifically including: The bimodal deep neural network uses INT8 quantization precision during the inference phase and is implemented using the TensorRT tool. The average inference latency on the embedded GPU platform is less than 10 milliseconds, and the detection accuracy decreases by no more than 1.5%. The bimodal deep neural network includes a temporal feature extraction branch and a topological feature extraction branch.

[0010] The process involves inputting continuous feature vectors into a dual-modal deep neural network for processing to generate a test strategy. Specific steps include: The temporal feature extraction branch is used to process the long-term dependencies of continuous feature vectors, and the topological feature extraction branch is used to model the communication relationships between vehicular network nodes. The weights of each branch are dynamically adjusted according to the real-time threat level to generate a test strategy. The temporal feature extraction branch adopts a dynamically sparsified LSTM structure, and its sparsity threshold θ is calculated based on the statistical distribution of gradient magnitude. The formula is: θ=μ-kσ, where μ is the gradient mean, σ is the gradient standard deviation, and k is an adjustable parameter with a value range of 1.5 to 2.0.

[0011] This includes real-time monitoring of the computational resource utilization of the bimodal deep neural network. When the computational load exceeds a preset threshold, the bimodal deep neural network is dynamically optimized, specifically including: Dynamic optimization includes reducing computational precision and selectively pruning non-critical feature channels, while prioritizing high-threat-level test tasks.

[0012] The process includes real-time monitoring of the computational resource utilization of the bimodal deep neural network. When the computational load exceeds a preset threshold, the bimodal deep neural network is dynamically optimized. Following this step, the process further includes: The parameters of the bimodal deep neural network are updated using an incremental learning approach, optimizing only the neurons most affected by new data, and performing corresponding response operations based on the threat level when an attack is detected.

[0013] This invention discloses a dynamic test strategy generation method for automotive network and data security. First, it collects vehicle CAN bus messages containing CANID, timestamps, and payload data. An adaptive sliding window mechanism is used for dynamic segmentation, and semantic encoding with attention mechanism is combined to generate continuous feature vectors. Then, the feature vectors are input into a dual-modal deep neural network, and network weights are dynamically adjusted according to the real-time threat level to generate the test strategy. Finally, a hierarchical optimization mechanism is used to achieve efficient utilization of computing resources, and incremental learning is used to update network parameters. Through dynamic sparse LSTM and hierarchical optimization strategies, response latency is reduced in high-concurrency attack scenarios. Channel-selective pruning and dynamic precision adjustment reduce model computational resource consumption, meeting vehicle hardware limitations. Incremental learning and threat level adaptive mechanisms improve the detection accuracy of new attacks, comprehensively solving the technical bottlenecks pointed out in the background art, such as poor adaptability of static pruning, low model inference efficiency, and response latency in high-frequency scenarios. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0015] Figure 1This is a flowchart of the steps in the method for generating dynamic testing strategies for automotive network and data security according to the first embodiment of the present invention. Detailed Implementation

[0016] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, but should not be construed as limiting the present invention.

[0017] The first embodiment of this application is as follows: Please see Figure 1 ,in, Figure 1 This is a flowchart of the steps in the method for generating dynamic testing strategies for automotive network and data security according to the first embodiment of the present invention.

[0018] This invention provides a method for generating dynamic testing strategies for automotive network and data security, comprising the following steps: S101: Collects vehicle CAN bus message data; Specifically, in the actual operation of intelligent connected vehicles, the onboard CAN bus, as the main communication network of the vehicle's electronic control system, carries data exchange for key functional modules including engine control, braking system, and battery management. This invention uses specialized data acquisition equipment, such as a CAN analyzer based on the SocketCAN protocol, to establish a physical connection with the vehicle's OBD-II diagnostic interface or directly connect to the target CAN bus node, acquiring raw CAN bus message data in real time. In use, firstly, a standard-compliant CAN bus transceiver, such as the TJA1050 chip, is selected at the hardware level, along with an embedded processor, such as the ARM Cortex-M series, to construct a data acquisition terminal. This terminal is connected to the vehicle's CAN bus through an electrical isolation circuit, ensuring that the acquisition process does not interfere with the vehicle's normal communication. The acquired message data strictly follows the CAN 2.0B extended frame format. Each frame contains a 29-bit identifier (CAN ID), 8 bytes of payload data, and a timestamp accurate to the microsecond level. The CAN ID is used to distinguish the communication type of different ECU nodes; for example, 0x18FEF100 represents voltage data from the battery management system. The timestamp records the precise arrival time of the message, providing a basis for subsequent timing analysis. During data acquisition, a dual-buffer storage architecture is employed: while the current buffer writes data in real time, the background buffer transmits the acquired data to the host computer processing system via USB or Ethernet interface. This design effectively avoids data loss, ensuring complete capture of all bus messages even at a high baud rate of 500kbps. Simultaneously, the bus load rate is monitored in real time. When an abnormally high load is detected, such as exceeding 80%, a data sampling rate adjustment mechanism is automatically triggered. This dynamically reduces the acquisition frequency of non-critical messages, such as switching from full acquisition to ID filtering mode, to ensure the integrity of critical safety data acquisition. To meet the adaptation needs of different vehicle models, this invention also incorporates a protocol-compliant message parsing library, capable of automatically identifying and parsing CAN message formats from different manufacturers. The acquired raw data is stored in a standardized structure, containing three core fields: CAN ID (4-byte hexadecimal), timestamp (8-byte Unix timestamp), and payload data (1-8 byte array). This structured processing lays the data foundation for subsequent feature extraction.

[0019] S102: Preprocess the CAN bus message data to generate a continuous feature vector; Specifically, an adaptive sliding window mechanism is first used to dynamically segment the time-series message data. Based on the characteristics of vehicle network communication, the initial window size is set to 50ms, corresponding to 5-10 cycles of a typical CAN bus periodic message, and network traffic changes are monitored in real time: when a sudden surge in traffic is detected, such as when diagnostic commands are sent in a concentrated manner, the window automatically shrinks to 20ms to improve timing resolution; under normal communication conditions, it expands to 100ms to enhance contextual relevance. The window adjustment algorithm comprehensively considers three factors: the current bus load rate (30%-70% is the optimal range), the dispersion of message ID distribution (calculated using Shannon entropy), and the matching degree of historical attack features. The message data within each window is strictly sorted by timestamp to form a time-series message sequence, preserving complete time dimension information for subsequent feature extraction. To meet the hard real-time requirements of the vehicle environment, this invention uses FPGA hardware acceleration of the adaptive sliding window mechanism. The FPGA chip (such as the Xilinx Artix-7 series) has a built-in parallel processing unit to implement the dynamic window size adjustment logic and message sorting buffer. Designed using hardware description languages ​​(such as Verilog), the FPGA processing unit is connected to the main processor (such as the ARM Cortex-A series) via a high-speed bus (such as PCIe) to achieve parallel data processing. At a baud rate of 500kbps, the window processing latency after FPGA acceleration is less than 2 milliseconds, ensuring real-time response in high-traffic scenarios. Next, joint semantic encoding is performed on the CAN ID and payload data within the window. Addressing the discrete nature of CAN IDs, an improved Word2Vec algorithm is used for distributed representation learning: a vocabulary containing all occurrences of CAN IDs is constructed, and the embedding vector for each CAN ID is calculated using a sliding context window (default size c=5). An attention mechanism is introduced to automatically adjust the context weights based on the message type (control commands / status data / diagnostic information), giving key IDs, such as the 0x0CF00400 braking command, higher feature representation strength. For the payload data, byte-level splitting and normalization (mapping to the [0, 1] interval) are performed first, followed by concatenation with the embedding vector of the corresponding CAN ID, ultimately generating a 128-dimensional continuous feature vector. To improve feature quality, this invention also includes three optimization measures: (1) outlier filtering: removing load values ​​that deviate significantly from the normal range based on the 3σ principle; (2) missing value compensation: using a forward filling algorithm to fill in periodic messages that are occasionally lost; (3) feature standardization: performing Z-score normalization on the final generated continuous feature vector to ensure that each dimension of the feature has zero mean and unit variance. The entire preprocessing process is optimized using fixed-point arithmetic when deployed on the embedded end, converting floating-point matrix operations into 8-bit integer calculations, so that the processing delay is controlled within 5ms, meeting the real-time requirements of the vehicle environment.The preprocessed feature vectors are retained in a circular buffer for downstream neural network modules to read as needed, forming a complete data pipeline.

[0020] S103: Input the continuous feature vector into the dual-modal deep neural network for processing to generate a test strategy; Specifically, the bimodal deep neural network consists of a temporal feature extraction branch and a topological feature extraction branch. It achieves comprehensive analysis of multi-dimensional features through a dynamic weight fusion mechanism. The specific implementation process is as follows: The temporal feature extraction branch uses a dynamically sparse LSTM structure to process the time-series characteristics of continuous feature vectors. The network initialization contains 256 LSTM units, with each time step inputting a 128-dimensional feature vector. Unlike traditional LSTM, this invention introduces a gradient-sensitive dynamic sparsity mechanism: the gradient magnitude of the connection weights of each neuron is calculated in real time. When the absolute value of the gradient is lower than the threshold θ = 0.001, the connection is automatically disconnected. The sparsity rate is dynamically adjusted according to the network load (adjustable from 20% to 60%). The mathematical derivation of the dynamic sparsity threshold θ is based on the statistical distribution of gradient magnitudes. During training, the gradient values ​​of all neuron connections are recorded, and the gradient mean μ and standard deviation σ are calculated. The threshold θ is set to θ = μ - kσ, where k is an adjustable parameter, and the optimal range is determined to be 1.5 to 2.0 through grid search. When k=1.8, the model maintains over 95% accuracy while reducing computation by 40%. This design significantly reduces computation while ensuring the ability to capture long-cycle patterns of CAN messages, such as battery status reports every 200ms, through the retention of strong connections. A time-attention module is followed by the LSTM layer to automatically identify key time points, such as the moment when an abnormal message suddenly appears, and its attention weight α... t The calculation formula is: α t = softmax(W a ·h t + b a ) Where h t Let W be the hidden state at time t. a and b a These are trainable parameters. The final temporal feature is represented as a weighted sum of the hidden states at all time steps.

[0021] The topology feature extraction branch uses a graph neural network (GNN) to model the communication relationships between vehicle ECU nodes. First, a dynamic topology graph is constructed: nodes represent all detected ECUs (identified by the first 8 bits of their CAN ID), and edge weights are dynamically updated based on the frequency of message interactions between nodes. The GNN contains three layers of graph convolutions; each layer aggregates features from neighboring nodes and is activated using ReLU, ultimately outputting a 128-dimensional topology feature for each node. Specifically, considering the highly dynamic topology unique to new energy vehicles, such as the addition of charging pile nodes during charging, a topology refresh cycle of 1 second is set, and potential malicious nodes are identified through anomaly detection in node degree distribution.

[0022] Design an adaptive weighting mechanism in the dual-modal feature fusion stage: Let the temporal feature be F t The topological feature is F g The fusion weight β∈[0,1] is determined by the real-time threat level T. r Dynamic calculation: β = σ(w·T) r + b) Where σ is the sigmoid function, and w and b are learnable parameters. When a high-frequency attack (T) is detected... r When the value is greater than 0.8, the weights of temporal features are automatically enhanced (β→0.8), while the contributions of both are balanced under normal conditions (β≈0.5). The fused features are mapped to test policy vectors through a two-layer fully connected network (dimension 256→128), containing three types of outputs: a list of monitored target IDs, a sampling frequency policy (1-100Hz), and a response priority (levels 1-3). The network is trained end-to-end, with the loss function being the fusion policy effect evaluation (detection rate, false positive rate) and resource consumption indicators. After 300 rounds of training using the Adam optimizer (lr=0.001), the network reaches a stable state. To improve inference efficiency, the bimodal deep neural network uses INT8 quantization technology during deployment. The FP32 model is converted to an INT8 model using TensorRT to reduce computational resources and memory usage. Inference latency tests on embedded GPUs (such as NVIDIA Jetson AGX Xavier) show that the average processing time is reduced from 15 milliseconds to 8 milliseconds, while the detection accuracy decreases by no more than 1.5%.

[0023] S104: Real-time monitoring of the computational resource utilization of the bimodal deep neural network; when the computational load exceeds a preset threshold, dynamic optimization of the bimodal deep neural network is performed. Specifically, during the operation of the dual-modal deep neural network, the usage of computing resources is continuously monitored. Key metrics such as CPU / GPU utilization, memory usage, inference latency, and cache hit rate are precisely collected every 10 milliseconds using an embedded performance counter. This monitoring data is analyzed using a sliding window. When CPU utilization consistently exceeds 75% and inference latency is greater than 15 milliseconds, or GPU memory usage exceeds 85% and cache hit rate is below 60%, a three-level progressive optimization strategy is automatically triggered. First, the computational precision is adjusted, achieving a lossless switch from FP32 to FP16 and then to INT8 via the TensorRT interface. Each adjustment is followed by a 5-second effect verification to ensure that the detection accuracy decrease does not exceed 3%. When the precision adjustment still fails to meet the requirements, the second level of optimization is implemented. Selective channel pruning based on gradient magnitude is applied to the LSTM network in the temporal feature extraction branch, with a maximum pruning ratio of 40%. Simultaneously, node-level pruning is performed on the GNN network in the topology feature extraction branch, preserving the connected subgraph composed of the core ECUs.

[0024] If resource constraints persist, a third-level task priority scheduling mechanism will be activated, establishing three queues for critical, important, and general tasks. Priority will be given to the detection needs of critical ECUs such as brakes and steering, while detection of general tasks like in-vehicle infotainment will be paused if necessary. All optimization operations will undergo a 10-second effectiveness evaluation, with real-time tracking of key metrics such as attack detection rate, false positive rate, and response latency. If any metrics deteriorate, the system will immediately roll back to the most recent stable configuration. The entire optimization process utilizes hardware acceleration technology for millisecond-level switching, saving the complete model state every 5 seconds as a recovery point, and rigorously logging all optimization operations.

[0025] S105: The parameters of the bimodal deep neural network are updated using an incremental learning approach, optimizing only the neurons most affected by new data, and performing corresponding response operations based on the threat level when an attack is detected.

[0026] Specifically, after dynamic optimization, the model enters the incremental learning and threat response phase. This process achieves continuous model evolution and real-time protection through intelligent parameter updates and a multi-level response mechanism. First, a neuron impact assessment system is established, using a sliding window with 100 samples. The gradient activity of neurons in each layer is statistically analyzed, and the contribution score of each neuron in the loss function is calculated using the formula: , among which, S i Neuron Influence Score: This score represents the importance of the i-th neuron to model performance updates. A higher score indicates a stronger learning need for that neuron under the current data, and it should be prioritized for incremental updates. The absolute value of the gradient: The weights W of the loss function L on neuron i iThe partial derivative (gradient) reflects the strength of the impact of weight changes on the loss. Taking the absolute value |·| is to ignore the gradient direction and only focus on the sensitivity of the neuron parameters. ΔW i Weight change: Records the actual adjustment magnitude (i.e., W) of the weights of neuron i in the previous incremental update. i new - W i old This measure is used to assess the neuron's "plasticity" and avoids frequent updates to already stable parameters.

[0027] The top 20% of high-impact neurons are selected for incremental updates. The update process employs a restricted backpropagation algorithm, allowing only these neurons to change their parameters while freezing other neurons. This selective update strategy reduces model update time and memory usage. To ensure update stability, a two-stage verification mechanism is implemented: first, the updated model performance is evaluated on an isolated validation set (representing 15% of the real-time data stream), requiring detection accuracy fluctuations to not exceed 2%; then, a shadow mode is run in actual traffic, continuously comparing the output differences between the old and new models to ensure accuracy before formal deployment.

[0028] For detected attacks, a three-tiered response strategy based on threat level is implemented. Threat level assessment comprehensively considers four dimensions: attack type (0-1 point), target ECU criticality (0-1 point), frequency of occurrence (0-1 point), and load anomaly (0-1 point), and obtains a comprehensive score of 0-4 points through weighted calculation (weighting coefficients optimized through real-vehicle testing). For Level 1 threats (score 1-2 points), logging and local alarms are executed, and a standard alarm frame of 0x18FFA001 is sent via the CAN bus; for Level 2 threats (score 2-3 points), progressive isolation of the target ECU is triggered, first restricting its non-critical communication, and then gradually cutting off the bus connection; for Level 3 threats (score 3-4 points), the vehicle's safety mode is activated, including: 1) sending a 0x0CFE11E0 emergency stop command; 2) cutting off power to non-safety-related ECUs; 3) activating the backup communication channel. All response operations are digitally signed by the Hardware Security Module (HSM) to ensure the authenticity and non-repudiation of the commands.

[0029] First, a meta-learning-based incremental update scheduler was developed, which can automatically adjust the update frequency (adjustable from 1 to 10 minutes) and batch size (8 to 32 samples) according to the real-time load of the onboard computing platform. Second, a spatiotemporal constraint mechanism for response actions was designed to limit the number of repeated operations of the same ECU within a 500ms time window to prevent over-response due to false alarms. Finally, deep integration with the vehicle diagnostic system was achieved, and the correlation analysis between safety events and vehicle status information was performed through the UDS protocol (ISO 14229) to improve the accuracy of threat determination.

[0030] By employing dynamic sparse LSTM and hierarchical optimization strategies, the response latency in high-concurrency attack scenarios is reduced; channel-selective pruning and dynamic accuracy adjustment reduce model computational resource consumption and meet the limitations of automotive hardware; through incremental learning and threat level adaptive mechanisms, the detection accuracy of new attacks is improved, comprehensively solving the technical bottlenecks pointed out in the background, such as poor adaptability of static pruning, low model inference efficiency, and response latency in high-frequency scenarios.

[0031] To verify the effectiveness and inventiveness of this invention, comparative tests were conducted in a real vehicle environment. The test data came from 100 hours of real CAN bus logs, including normal traffic and various injection attacks (such as DoS, fuzzing attacks, and replay attacks). This invention was compared with rule engines and traditional SVM methods, and the results are as follows: Detection rate: This invention achieves 98.5%, the rule engine 85.2%, and SVM 90.1%; False positive rate: 0.5% for this invention, 3.2% for the rule engine, and 2.1% for SVM; Average response latency: 12 milliseconds for this invention, 50 milliseconds for the rule engine, and 35 milliseconds for SVM; Resource consumption: This invention reduces memory consumption by 60% after INT8 quantization and inference latency to less than 10 milliseconds.

[0032] The above data shows that the present invention is significantly superior to traditional methods in terms of detection accuracy, real-time performance, and resource efficiency, fully demonstrating the necessity and synergistic effect of the technology combination.

[0033] The above-disclosed embodiments are merely one or more preferred embodiments of this application and should not be construed as limiting the scope of this application. Those skilled in the art can understand that all or part of the processes for implementing the above embodiments and equivalent changes made in accordance with the claims of this application still fall within the scope of this application.

Claims

1. A method for generating a dynamic test strategy for automotive network and data security, characterized in that The method comprises the following steps: Collecting vehicle CAN bus message data; Pretreating the CAN bus message data to generate a continuous feature vector; Inputting the continuous feature vector into a dual-modal deep neural network for processing to generate a test strategy; Real-time monitoring of the utilization rate of the computing resources of the dual-modal deep neural network, and performing dynamic optimization of the dual-modal deep neural network when the computing load exceeds a preset threshold.

2. The method for dynamic test strategy generation for automotive network and data security of claim 1, wherein, The vehicle CAN bus message data comprises CAN ID, timestamp and payload data.

3. The method for dynamic test strategy generation of automotive network and data security according to claim 2, characterized in that, The continuous feature vector is generated by pretreating the CAN bus message data, and the specific steps comprise: Segmenting the CAN bus message data using a sliding window mechanism, dynamically adjusting the window size according to the current network traffic, and performing semantic encoding on the CAN ID and payload data to generate a continuous feature vector, wherein the sliding window mechanism is implemented through FPGA hardware acceleration and comprises parallel window adjustment logic and message sorting buffer.

4. The method for dynamic test strategy generation of automotive network and data security according to claim 3, characterized in that, The continuous feature vector is input into the dual-modal deep neural network for processing to generate a test strategy, and the specific steps comprise: The dual-modal deep neural network adopts INT8 quantization precision in the inference stage and is implemented through the TensorRT tool, and the average inference delay on the embedded GPU platform is less than 10 milliseconds, and the detection accuracy decreases by no more than 1.5%, wherein the dual-modal deep neural network comprises a time series feature extraction branch and a topology feature extraction branch.

5. The method for dynamic test strategy generation of automotive network and data security according to claim 4, characterized in that, The continuous feature vector is input into the dual-modal deep neural network for processing to generate a test strategy, and the specific steps comprise: The time series feature extraction branch is used to process the long-term dependence relationship of the continuous feature vector, and the topology feature extraction branch is used to model the communication relationship between the vehicle network nodes, and the weights of each branch are dynamically adjusted according to the real-time threat level to generate a test strategy, wherein the time series feature extraction branch adopts a dynamic sparsification LSTM structure, and the sparsification threshold θ is calculated based on the statistical distribution of the gradient amplitude, and the formula is: θ = μ-kσ, wherein μ is the gradient mean, σ is the gradient standard deviation, and k is an adjustable parameter with a value range of 1.5 to 2.

0.

6. The method for dynamic test strategy generation of automotive network and data security according to claim 5, characterized in that, Real-time monitoring of the utilization rate of the computing resources of the dual-modal deep neural network, and performing dynamic optimization of the dual-modal deep neural network when the computing load exceeds a preset threshold, and the specific steps comprise: The dynamic optimization comprises reducing the calculation precision and selectively pruning non-critical feature channels, and preferentially executing high-threat-level test tasks.

7. The method for dynamic test strategy generation of automotive network and data security according to claim 6, characterized in that, Real-time monitoring of the utilization rate of the computing resources of the dual-modal deep neural network, and performing dynamic optimization of the dual-modal deep neural network when the computing load exceeds a preset threshold, and the steps further comprise: Updating the parameters of the dual-modal deep neural network in an incremental learning manner, optimizing only the neurons most affected by new data, and performing corresponding response operations according to the threat level when an attack behavior is detected.

Images