An information processing method, device, storage medium and vehicle-mounted electronic controller

Abstract

The application discloses an information processing method and device, a storage medium and a vehicle-mounted electronic controller; knowledge when a vehicle normally travels is combined into a generative adversarial network, so that a sample generated by a generator is necessarily an abnormal sample with attack, and the generated sample is a sample most likely to be wrongly detected by an existing machine learning model, so that the effect of training the existing machine learning model is improved, and the generalization capability is improved; accordingly, the application also discloses corresponding device, medium and other related product embodiments, and the same or corresponding technical solutions are adopted to improve the technical performance of the related products or devices.

Description

Technical Field

[0001] This invention relates to the field of intelligent vehicle technology, and in particular to an information processing method, device, storage medium, and vehicle electronic controller. Background Technology

[0002] As the level of electronic information technology in automobiles continues to improve, the resulting information security issues are becoming increasingly prominent. Intelligent connected vehicles have many more external communication ports compared to traditional vehicles; for example, GSM or 2G (Global System for Mobile Communication, 2nd Generation Mobile Communication Technology) up to 4G (4th Generation Mobile Communication Technology), 5G, Bluetooth Low Energy (BLE) communication systems, and Wi-Fi (Wireless-Fidelity) wireless local area networks. This effectively opens up many vulnerable ports for attack, exposing the already fragile in-vehicle communication network to hackers' attack range.

[0003] Current automotive gateway products use manually defined rules for network intrusion detection modules to detect anomalies. However, manually defined rules cannot cover all operating conditions, especially when the physical meaning of related signals is abnormal. Machine learning-based detection algorithms can compensate for this deficiency, allowing the algorithm to automatically learn the correlations between signals. However, the automotive field differs from the computer field in that the amount of anomaly data is relatively small, and due to human and vehicle security reasons, it is difficult to obtain such data through cyberattacks on real vehicles. Even normal vehicle driving data cannot cover all operating conditions because data collection involves real vehicles, which is costly. This results in insufficient training samples for machine learning models, leading to problems such as overfitting and weak generalization ability.

[0004] There are two main approaches in the automotive field to address the problem of insufficient training samples: The first approach involves having the generator learn the distribution of normal or abnormal samples, allowing it to generate more of these samples. The drawback of this method is that if the original data distribution is uneven, the generator can only generate unevenly distributed data, and the generated samples may not be those that the machine learning model misdetects. Adding these samples to the training dataset does not guarantee improved generalization ability. The second approach treats the discriminator in a generative adversarial network (GAN) as a detection module, allowing the generator to generate samples to deceive the detection module. In this way, the generator-generated samples are those most likely to be misdetected by the machine learning model. However, this approach lacks restrictions on the samples generated by the generator. These samples may contain those without network attack capabilities, or even normal samples. While adding these samples to the training dataset can improve anomaly detection, it leads to a higher false positive rate, misclassifying normal samples as anomalous. Summary of the Invention

[0005] This invention discloses an information processing method, apparatus, storage medium, and vehicle electronic controller; it combines knowledge of normal vehicle operation into a generative adversarial network, ensuring that the samples generated by the generator are always aggressive and abnormal samples, and that the generated samples are the most likely to be detected errors by existing machine learning models, thereby improving the effect of training existing machine learning models and enhancing generalization ability.

[0006] The information processing method includes an information collection step, a model training step, and a sample screening step; wherein, the information collection step obtains necessary data by acquiring vehicle driving information; the vehicle driving information includes normal samples collected during normal vehicle driving.

[0007] Furthermore, a normal sample set is formed by normal samples; at this point, if there is no first original abnormal sample set formed by abnormal samples, an abnormal sample set corresponding to the information intrusion process is constructed, and the vehicle dataset is initialized.

[0008] The vehicle dataset includes a first training set, a second validation set, and a third test set; the features of the vehicle dataset include a first functional feature and a second non-functional feature.

[0009] Furthermore, the model training step trains the machine learning model using both normal and abnormal samples.

[0010] Furthermore, the sample screening step selects samples that meet the second non-functional characteristics as the first screened samples; replaces the samples to be optimized in the first original abnormal sample set in the first training set as the first screened samples; repeats the model training step and the sample screening step to form a loop iterative process, which optimizes the data until it meets the preset convergence condition.

[0011] Furthermore, the sample selection step may also include a first training process, which uses normal samples and abnormal samples to train a first discriminator to simulate a machine learning model or uses a machine learning model as the first discriminator.

[0012] Furthermore, the sample screening step may also include a second training process, which uses anomalous samples and a first discriminator to train a first generator; the first generator generates only a second non-functional feature; the first functional feature corresponds to the sample features associated with the network intrusion mode.

[0013] Furthermore, the sample screening step may also include a third search process, which uses an optimization algorithm to search for a second non-functional feature; the optimization algorithm of the third search process does not include generative adversarial networks; the third search process obtains the first anomalous sample through the optimization algorithm.

[0014] Furthermore, by introducing random noise into the first original abnormal sample set and sampling the first generator to obtain the second abnormal sample, the first abnormal sample and the second abnormal sample are used as the first screening samples to replace the first original abnormal sample in the first training set.

[0015] Furthermore, the information collection step collects information once at preset time intervals; from the current time to the historical time, the first time, the second time, and so on until the Tth time are obtained sequentially from the first sampling interval, where T is a natural number.

[0016] Furthermore, the samples of the vehicle dataset that can be spliced ​​are one-dimensional vectors, and one-hot encoding is used to process non-numerical features. At the same time, the data is normalized to the [0,1] interval using minimum-maximum normalization. The information collection step collects data through wired or wireless means / channels. The wireless means / channels include 4G, 5G, BLE, and Wi-Fi networks, and the wired means / channels include CAN and Ethernet.

[0017] The convergence conditions include the machine learning model's F1 score on the test set no longer increasing or the F1 score approaching a preset limit value according to a preset error threshold.

[0018] Specifically, the machine learning models here can be divided into two categories: the first category of gradient models and the second category of non-gradient models. The first category of gradient models includes neural network models and logistic regression models; the second category of non-gradient models includes decision tree models and random forest models.

[0019] Furthermore, the optimization algorithm for its third search process includes the adversarial machine learning algorithm FGSM and the genetic algorithm; its first sampling interval is a preset time length; this time length is not unique and can be set to 1 second; its information acquisition step can obtain vehicle driving information under preset working conditions; among which, vehicle driving information includes vehicle speed, driver torque demand information and gear information.

[0020] Specifically, its iterative process may include an exchange refresh process, which replaces the corresponding sample of the same type before selection in the vehicle driving information with the first selected sample after selection.

[0021] Furthermore, embodiments of the present invention also disclose a machine learning device, including an information acquisition unit, a model training unit, and a sample selection unit.

[0022] Specifically, the information collection unit acquires vehicle driving information to complete subsequent processing; this vehicle driving information includes normal samples collected during normal vehicle driving.

[0023] Furthermore, a normal sample set is formed by normal samples; if there is no first original abnormal sample set formed by abnormal samples, an abnormal sample set corresponding to the information intrusion process is constructed, and the vehicle dataset is initialized.

[0024] The vehicle dataset includes a first training set, a second validation set, and a third test set; the features of the vehicle dataset include a first functional feature and a second non-functional feature.

[0025] Furthermore, the model training unit trains the machine learning model using normal and abnormal samples.

[0026] Furthermore, the sample screening unit selects samples that meet the second non-functional characteristics as the first screened samples; replaces the samples to be optimized in the first original abnormal sample set in the first training set as the first screened samples; repeats the information processing process in the model training unit and the sample screening unit, and refreshes the data in a cyclic iterative process; the cyclic iterative process optimizes the data until it meets the preset convergence condition.

[0027] Specifically, the sample selection unit also performs a first training process, which uses normal samples and abnormal samples to train a first discriminator to simulate a machine learning model or uses a machine learning model as the first discriminator.

[0028] Furthermore, the sample screening unit also performs a second training process, which uses anomalous samples and a first discriminator to train a first generator; the first generator generates only a second non-functional feature; the first functional feature corresponds to the sample features associated with the network intrusion mode.

[0029] Furthermore, the sample screening unit also performs a third search process, which uses an optimization algorithm to search for a second non-functional feature; the optimization algorithm of the third search process does not include generative adversarial networks; the third search process obtains a first anomalous sample through the optimization algorithm.

[0030] Furthermore, by introducing random noise into the first original abnormal sample set and sampling the first generator to obtain the second abnormal sample, the first abnormal sample and the second abnormal sample are used as the first screening samples to replace the first original abnormal sample in the first training set.

[0031] Specifically, the information acquisition unit collects information once at preset time intervals; from the current time to the historical time, the first time, the second time, and so on until the Tth time are obtained sequentially from the first sampling interval, where T is a natural number.

[0032] Furthermore, the samples in the vehicle dataset can be spliced ​​as one-dimensional vectors, and one-hot encoding can be used to process non-numerical features. The data can be normalized to the [0,1] interval using minimum-maximum normalization.

[0033] The information collection step collects data through wired or wireless means / channels; the wireless means / channels include 4G, 5G, BLE, and Wi-Fi networks, and the wired means / channels include CAN and Ethernet.

[0034] Furthermore, the convergence conditions for its iterative loop include: causing the F1 score of the machine learning model on the test set to no longer increase or the F1 score to approach a preset limit value according to a preset error threshold.

[0035] Specifically, its machine learning models include a first type of gradient model and a second type of non-gradient model; the first type of gradient model includes a neural network model and a logistic regression model; the second type of non-gradient model includes a decision tree model and a random forest model; the optimization algorithms of its third search process include the adversarial machine learning algorithm FGSM and a genetic algorithm; its first sampling interval is a preset time length; its time length is not unique, and the preset time length can be 1 second; its information acquisition steps obtain vehicle driving information under preset working conditions.

[0036] The vehicle driving information includes vehicle speed, driver torque requirements, and gear information.

[0037] Furthermore, its iterative process includes an exchange refresh process, in which the first selected sample after optimization replaces the corresponding sample of the same type before optimization in the vehicle driving information.

[0038] Meanwhile, embodiments of the present invention also disclose a computer storage medium and an in-vehicle electronic controller that adopt the same inventive concept. Their information processing process is the same as or corresponding to the above-mentioned methods and devices, and will not be described again here.

[0039] It should be noted that the terms "first," "second," and similar terms used in this article are merely for describing the constituent elements of the technical solution and do not constitute a limitation on the technical solution, nor should they be interpreted as an indication or implication of the importance of the corresponding elements; elements with terms such as "first," "second," or similar terms indicate that at least one of the elements is included in the corresponding technical solution. Attached Figure Description

[0040] To more clearly illustrate the technical solution of the present invention and facilitate a further understanding of the technical effects, technical features, and objectives of the present invention, the present invention will be described in detail below with reference to the accompanying drawings. The accompanying drawings constitute an essential part of the specification and are used together with Embodiment 1 of the present invention to illustrate the technical solution of the present invention, but do not constitute a limitation on the present invention.

[0041] The same reference numerals in the attached diagrams represent the same parts, specifically:

[0042] Figure 1 This is a schematic diagram of the data structure of the first type or normal sample in an embodiment of the present invention;

[0043] Figure 2 This is a schematic diagram of the information processing method according to an embodiment of the present invention;

[0044] Figure 3 This is a schematic diagram of the first training process of the discriminator in an embodiment of the present invention;

[0045] Figure 4 This is a schematic diagram of the second training process of the generator in an embodiment of the present invention;

[0046] Figure 5 This is a schematic diagram of the process and information flow of an embodiment of the method of the present invention;

[0047] Figure 6 This is a schematic diagram of the structure and information flow of an embodiment of the device of the present invention;

[0048] Figure 7 This is a schematic diagram showing the layout of the device, medium, and vehicle-mounted electronic controller of the present invention.

[0049] in:

[0050] 001-Information Acquisition Unit, 003-Model Training Unit, 005-Sample Selection Unit;

[0051] 009 - Vehicle, 011 - Machine learning device, 013 - Storage medium, 015 - Onboard electronic controller.

[0052] 097 - Vehicle driving information, 099 - Sampling time point

[0053] 100 - First sampling point or first sampling time, 101 - Second sampling point or second sampling time.

[0054] 10T - the Tth sampling point or the Tth sampling time, where T is a natural number;

[0055] 110 - First parameter / vehicle speed V (unit: meters per second).

[0056] 120 - Second parameter / Driver torque requirement T (unit: N·m)

[0057] 1i0 is the i-th parameter / gear position, where i is a natural number;

[0058] 201 - Information Collection Steps, 203 - Model Training Steps, 205 - Sample Selection Steps

[0059] 215 - First training process, 235 - Second training process, 255 - Third search process.

[0060] 275 - Exchange and refresh process; 300 - Looping and iterating process;

[0061] 401 - Input data set, 413 - Intermediate data set

[0062] 430 - Output data set; 435 - Optimal data replacement;

[0063] 501 - First Discriminator

[0064] 601 - First functional feature; 602 - Second non-functional feature;

[0065] 701 - Random noise, 800 - First screening sample. Detailed Implementation

[0066] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Of course, the specific embodiments described below are merely illustrative of the technical solutions of the present invention, and not intended to limit the invention. Furthermore, the parts described in the embodiments or drawings are merely illustrative examples of relevant parts of the present invention, and not the entirety of the invention.

[0067] like Figure 5As shown, the information processing method disclosed in this embodiment of the invention includes an information acquisition step 201, a model training step 203, and a sample screening step 205.

[0068] Information collection step 201: Obtain vehicle driving information 097 and complete necessary initialization; wherein, vehicle driving information 097 includes normal samples collected during normal vehicle driving.

[0069] like Figure 4 The normal sample set is composed of normal samples; if there is no first original abnormal sample set 600 composed of abnormal samples, then an abnormal sample set corresponding to the information intrusion process is constructed and the vehicle dataset is initialized; wherein, the vehicle dataset includes a first training set, a second validation set, and a third test set; the features of the vehicle dataset include a first functional feature 601 and a second non-functional feature 602.

[0070] Furthermore, such as Figure 2 Model training step 203 trains the machine learning model using normal and abnormal samples.

[0071] Further, in the sample screening step 205, samples that meet the second non-functional characteristics are selected as the first screened samples 800; the samples to be optimized in the first original abnormal sample set 600 in the first training set are replaced as the first screened samples 800; the model training step 203 and the sample screening step 205 are repeated to form a loop iteration process 300, and the loop iteration process 300 optimizes the data until it meets the preset convergence conditions.

[0072] Furthermore, such as Figure 2 , Figure 3 The sample selection step 205 also includes a first training process 215, which uses normal samples and abnormal samples to train a first discriminator 501 to simulate a machine learning model or uses a machine learning model as the first discriminator 501.

[0073] Furthermore, such as Figure 2 , Figure 4 The sample screening step 205 also includes a second training process 235, which uses abnormal samples and a first discriminator 501 to train a first generator 503; the first generator 503 generates only a second non-functional feature 602; the first functional feature corresponds to the sample features associated with the network intrusion method.

[0074] Furthermore, such as Figure 2 The sample screening step 205 also includes a third search process 255, which uses an optimization algorithm to search for second non-functional features; wherein, the optimization algorithm of the third search process 255 does not include the method of generative adversarial networks; the third search process obtains the first abnormal sample through the optimization algorithm.

[0075] Furthermore, such as Figure 4 Random noise 701 is introduced into the first original abnormal sample set and the first generator 503 is sampled to obtain the second abnormal sample. The first abnormal sample and the second abnormal sample are used as the first screened sample 800 to replace the first original abnormal sample 600 in the first training set.

[0076] Furthermore, such as Figure 1 Information collection step 201 collects information once every preset time interval of 0.99; from the current time to the historical time, the first time interval 100, the second time interval 101, and so on until the Tth time interval 10T are obtained in sequence, where T is a natural number.

[0077] Furthermore, the samples of the spliced ​​vehicle dataset are one-dimensional vectors, and one-hot encoding is used to process the non-numerical features. Then, the data is normalized to the [0,1] interval using minimum-maximum normalization.

[0078] The information collection step 201 collects data through wired or wireless means / channels; the wireless means / channels include 4G, 5G, BLE, and Wi-Fi networks, and the wired means / channels include CAN and Ethernet.

[0079] Specifically, the aforementioned convergence conditions include the machine learning model no longer improving its F1 score on the test set or the F1 score approaching a preset limit value according to a preset error threshold.

[0080] Furthermore, its machine learning model can adopt either a first-class gradient model or a second-class non-gradient model; wherein, the first-class gradient model includes neural network models and logistic regression models; the second-class non-gradient model includes decision tree models and random forest models; and the optimization algorithm of its third search process 255 includes adversarial machine learning algorithm FGSM and genetic algorithm.

[0081] Specifically, such as Figure 1 , Figure 2 The first sampling interval is a preset time length; its length is not unique, and the preset time length can be 1 second; the information collection step 201 obtains vehicle driving information 097 under preset working conditions; the vehicle driving information 097 includes vehicle speed 110, driver torque demand 120, and gear information 1i0; its iterative process 300 includes an exchange refresh process 275; the exchange refresh process 275 replaces the corresponding sample of the same type before optimization in the vehicle driving information 097 with the first screened sample 800 after optimization.

[0082] Among them, normal driving data under various operating conditions is collected to obtain normal samples; the collection interval can be 1 second, such as... Figure 1The sliding window duration is from the current time 0 seconds to the history - T seconds; Figure 1 This is the data structure of a normal sample.

[0083] Furthermore, vehicle knowledge can be used to identify the actual features of samples that play a role in each type of network intrusion, i.e., functional features, and an initial set of anomalous samples can be constructed.

[0084] For example, the velocity signal at the current moment The signal tampering attack method; by using the vehicle's longitudinal dynamics model and historical data such as vehicle speed - 1 second, driver torque demand, and gear position signal, it is possible to estimate... The value of .

[0085] Among them, the definition and A threshold for deviation is used to determine whether signal tampering has occurred.

[0086] For example when This is signal tampering;

[0087] This leads to the range of values ​​for the tampered signal. Randomly select values ​​within this range. Then use Replace the original data This creates an abnormal sample.

[0088] The functional characteristics of this anomalous sample include the vehicle speed (historical value - 1 second), driver torque demand, gear position signal, and tampering. These ensure that the sample is a genuine anomalous sample; while the other signals are non-functional features.

[0089] For each abnormal sample, record the positions of its functional and non-functional features in the feature vector to facilitate subsequent differentiation.

[0090] Typically, the dataset can be divided into a first training set, a second validation set, and a third test set; if necessary, the data can be preprocessed: first, a sample can be concatenated into a one-dimensional vector.

[0091] For non-numerical features, one-hot encoding can be used; then, minimum-maximum normalization can be used to unify the data to the [0,1] interval.

[0092] Furthermore, machine learning models are trained using normal and abnormal samples.

[0093] In this sample, the label is 0 for normal samples and 1 for abnormal samples.

[0094] Specifically, machine learning models can be any model, and they are divided into two categories according to whether the gradient of the output with respect to the input can be obtained. The category with obtainable gradients includes neural networks and logistic regression, while the category without obtainable gradients includes decision trees and random forests.

[0095] Furthermore, the discriminator can be trained using normal and abnormal samples to simulate the detection behavior of the machine learning model, or the machine learning model can be directly regarded as the discriminator. When the machine learning model can obtain the gradient, it can be directly regarded as the discriminator for subsequent training processes, and this step can be skipped.

[0096] When the gradient cannot be obtained, a discriminator based on a fully connected network can be trained separately to simulate the detection behavior of a machine learning model.

[0097] Note that the sample labels in this dataset are no longer the labels from the original dataset, but rather the classification categories given by the machine learning model.

[0098] Its loss function is:

[0099] ,

[0100] in Expressing expectations, Indicates a sample, This represents the set of samples that are judged as normal by the machine learning model.

[0101] This represents the set of samples that are judged as abnormal by the machine learning model. Indicates the discriminator,

[0102] Indicates in and The sample set sampled between,

[0103] Right now .

[0104] The loss function uses the form from Wasserstein generative adversarial networks, so the logarithmic form of cross-entropy is removed, and a gradient penalty term is added, which helps to stabilize the training process.

[0105] Furthermore, the sigmoid activation function was removed from the last layer of the discriminator network. Gradient descent was then used for training.

[0106] Discriminator network structure example:

[0107] 1) The input layer is a one-dimensional vector with a length equal to the length of the feature vector of the sample;

[0108] 2) Linear layer, with 512 neurons and Leaky ReLU activation function;

[0109] 3) Linear layer, with 256 neurons and Leaky ReLU activation function;

[0110] 4) Linear layer, with 1 neuron and no activation function.

[0111] Furthermore, such as Figure 4 The generator is trained using anomalous samples and a discriminator, and the generator is restricted to generating only non-functional features to ensure that the samples are offensive.

[0112] The generator is a fully connected network, and its input is a concatenation of the feature vector of the original abnormal sample and the random noise vector.

[0113] Specifically, the activation function of the last layer of the generator can be the Sigmoid activation function, and the output is the feature vector of the new sample.

[0114] To preserve some of the original data's value norms, such as integers, the generated feature vectors can be restored to the original data before minimum-maximum normalization, then rounded, and then normalized again. This makes the generated samples more realistic.

[0115] Then, non-functional features are extracted from the feature vector of this new sample. These non-functional features are then concatenated with the functional features in the feature vector input to the generator. The position of the features during concatenation must be the same as the original position, and finally a generated anomalous sample is formed. This complete anomalous sample is then fed into the discriminator for detection.

[0116] The loss function of the generator is ,

[0117] in Represents the original abnormal sample vector. Represents the noise vector.

[0118] This represents the functional features in the feature vector of anomaly samples.

[0119] Represents a generator. This represents the original set of abnormal samples.

[0120] The generator's loss function also adopts the form in Wasserstein generative adversarial networks, that is, the logarithmic form of cross-entropy is removed.

[0121] Furthermore, gradient descent is used for training.

[0122] Generator network structures are as follows:

[0123] 1) The input layer is a one-dimensional vector, and its length is equal to the length of the feature vector of the sample plus the length of the random noise vector;

[0124] 2) Linear layer, with 512 neurons and Leaky ReLU activation function;

[0125] 3) Linear layer, with 256 neurons and Leaky ReLU activation function;

[0126] 4) Linear layer: The number of neurons equals the length of the feature vector of the sample, and the activation function is Sigmoid.

[0127] Then, non-functional features need to be extracted from the output layer based on the record of the functional features of the input samples.

[0128] Furthermore, other gradient-based or non-gradient optimization algorithms besides generative adversarial networks can be used to search for non-functional features to identify anomalous samples that the discriminator is most likely to misjudge.

[0129] These samples are complementary to those found by the generative adversarial network, especially the non-gradient optimization algorithm; because the gradient descent algorithm used when training the generative adversarial network is gradient-based and prone to getting trapped in local optima.

[0130] Gradient-based optimization algorithms, such as the adversarial machine learning algorithm FGSM (Fast Gradient SignMethod).

[0131] The perturbation only affects non-functional features.

[0132] Right now ,

[0133] This represents the non-functional features in the feature vector of anomaly samples.

[0134] This represents a perturbation of non-functional features in the sample. To represent a small constant, The symbol function is represented; ultimately, for each anomalous sample, a corresponding adversarial sample can be found using the FGSM algorithm.

[0135] Among them, non-gradient optimization algorithms, such as genetic algorithms, require the optimization of non-functional features in the original abnormal samples.

[0136] For each anomalous sample, non-functional features are searched. When calculating the fitness function, the searched non-functional features are concatenated with the functional features in the original anomalous sample. The position of the features during concatenation should be the same as the original position. Then, the features are substituted into the discriminator to calculate the fitness.

[0137] Therefore, the optimization objective is .

[0138] Ultimately, by optimizing each anomalous sample, a corresponding optimized anomalous sample can be found.

[0139] Specifically, new anomalous samples are generated by sampling from the generator using the original anomalous sample set and random noise. These new anomalous samples are then used together with other new anomalous samples found by gradient-based or non-gradient optimization algorithms to replace the anomalous samples in the training set. This process is repeated to train the machine learning model, thus forming a loop. The above steps are repeated until the convergence condition is met.

[0140] The above step preserves the original normal samples in the training dataset; the label of the normal samples is 0, and the label of the new abnormal samples is 1; the convergence condition is that the f1 score of the machine learning model on the test set no longer improves.

[0141] like Figure 6 As shown in the figure, this embodiment of the invention also discloses a machine learning device, including an information acquisition unit 001, a model training unit 003, and a sample selection unit 005.

[0142] Among them, such as Figure 1 , Figure 4 The information acquisition unit 001 acquires vehicle driving information 097, which includes normal samples collected during normal vehicle driving; the normal samples constitute a normal sample set; if there is no first original abnormal sample set 600 composed of abnormal samples, then an abnormal sample set corresponding to the information intrusion process is constructed, and the vehicle dataset is initialized; the vehicle dataset is divided into a first training set, a second verification set, and a third test set; the features of the vehicle dataset are divided into a first functional feature 601 and a second non-functional feature 602.

[0143] Furthermore, the model training unit 003 trains the machine learning model using normal and abnormal samples.

[0144] Furthermore, the sample screening unit 005 selects samples that meet the second non-functional characteristics as the first screened samples 800; replaces the samples to be optimized in the first original abnormal sample set 600 in the first training set as the first screened samples 800; repeats the information processing process in the model training unit 003 and the sample screening unit 005, and refreshes the data with a loop iteration process 300; then the loop iteration process 300 optimizes the data until it meets the preset convergence conditions.

[0145] Furthermore, the sample selection unit 005 also performs a first training process 215, which uses normal samples and abnormal samples to train a first discriminator 501 to simulate a machine learning model or uses a machine learning model as the first discriminator 501.

[0146] Furthermore, the sample screening unit 005 also performs a second training process 235, which uses abnormal samples and the first discriminator 501 to train the first generator 503; the first generator 503 generates only the second non-functional feature 602; the first functional feature corresponds to the sample feature associated with the network intrusion mode.

[0147] Furthermore, the sample screening unit 005 also performs a third search process 255, which uses an optimization algorithm to search for a second non-functional feature; the optimization algorithm of the third search process 255 does not include the method of generative adversarial networks; the third search process obtains the first abnormal sample through the optimization algorithm.

[0148] Furthermore, random noise 701 is introduced into the first original abnormal sample set and the first generator 503 is sampled to obtain a second abnormal sample. The first abnormal sample and the second abnormal sample are used as the first screened sample 800 to replace the first original abnormal sample 600 in the first training set.

[0149] Specifically, the information acquisition unit 001 acquires information once every preset time interval 099; from the current time to the historical time, the first sampling interval is sequentially obtained as the first time 100, the second time 101, and so on up to the Tth time 10T, where T is a natural number; the samples of the spliced ​​vehicle dataset are one-dimensional vectors, and one-hot encoding is used to process the non-numerical features, and then the data is normalized to the [0,1] interval using minimum-maximum normalization; wherein, the information acquisition step 201 acquires data through wired or wireless means / channels; the wireless means / channels include 4G, 5G, BLE, and Wi-Fi networks, and the wired means / channels include CAN and Ethernet.

[0150] Furthermore, its convergence conditions include: causing the F1 score of the machine learning model on the test set to no longer increase or the F1 score to approach a preset limit value according to a preset error threshold.

[0151] Furthermore, the machine learning model can be divided into a first type of gradient model and a second type of non-gradient model; the first type of gradient model includes neural network model and logistic regression model; the second type of non-gradient model includes decision tree model and random forest model; the optimization algorithm of its third search process 255 includes adversarial machine learning algorithm FGSM and genetic algorithm; its first sampling interval is a preset time length; its length is not unique, and the preset time length can be 1 second; its information acquisition step 201 obtains vehicle driving information 097 under preset working conditions.

[0152] Specifically, such as Figure 2 The vehicle driving information 097 includes vehicle speed 110, driver torque requirement 120, and gear information 1i0; the iterative process 300 includes an exchange refresh process 275, in which the first selected sample after optimization is replaced with the same type of sample before optimization in the vehicle driving information 097.

[0153] like Figure 7 The present invention also discloses a computer storage medium and a corresponding vehicle electronic controller, which adopts the same technical solution as the method and device of the present invention, solves the same technical problem, and has the same implementation method or process. The specific implementation process will not be repeated here.

[0154] It should be noted that the above embodiments are only for more clearly illustrating the technical solution of the present invention. Those skilled in the art will understand that the implementation of the present invention is not limited to the above content. Any obvious changes, substitutions or replacements made based on the above content do not exceed the scope of the technical solution of the present invention. Other implementations will also fall within the scope of the present invention without departing from the concept of the present invention.

Claims

1. An information processing method characterized by comprising: include: Information collection step (201), model training step (203), sample selection step (205), The information collection step (201) obtains vehicle driving information (097), which includes normal samples collected during normal vehicle driving. The normal samples constitute a normal sample set; if there is no first original abnormal sample set (600) composed of abnormal samples, then the abnormal sample set corresponding to the information intrusion process is constructed: the sample features that play a real role under the network intrusion method are given by vehicle knowledge, that is, functional features, and an initial abnormal sample set is constructed: the normal value of the signal is estimated according to the functional features, and the value range of the tampered signal is determined according to the threshold of the deviation between the signal and the normal value; a value is randomly selected within the value range of the tampered signal, and the corresponding signal value in the original data is replaced, thereby constructing the original abnormal sample. Initialize the vehicle dataset; the vehicle dataset includes a first training set, a second validation set, and a third test set; the features of the vehicle dataset include a first functional feature (601) and a second non-functional feature (602); The model training step (203) trains the machine learning model using the normal samples and the abnormal samples; The sample screening step (205) selects samples that meet the second non-functional characteristics as the first screened samples (800); replaces the samples to be optimized in the first original abnormal sample set (600) in the first training set as the first screened samples (800); repeats the model training step (203) and the sample screening step (205) to form a loop iteration process (300), and the loop iteration process (300) optimizes the data until it meets the preset convergence conditions; The sample screening step (205) further includes a first training process (215), which uses the normal samples and the abnormal samples to train a first discriminator (501) to simulate the machine learning model or uses the machine learning model as the first discriminator (501). The sample screening step (205) further includes a second training process (235), which uses the abnormal samples and the first discriminator (501) to train a first generator (503); the first generator (503) generates only the second non-functional feature (602); the first functional feature corresponds to the sample feature associated with the network intrusion mode; The sample screening step (205) further includes a third search process (255), which uses an optimization algorithm to search for the second non-functional feature; the optimization algorithm of the third search process (255) does not include the method of generative adversarial networks; the third search process obtains a first anomalous sample by searching with the optimization algorithm: the second non-functional feature obtained by the optimization algorithm is concatenated with the functional feature in the original anomalous sample to generate the first anomalous sample; Random noise (701) is introduced into the first original abnormal sample set and the first generator (503) is sampled to obtain the second abnormal sample: the second non-functional feature output by the generator is concatenated with the functional feature in the feature vector when input to the generator to generate the second abnormal sample; The first abnormal sample and the second abnormal sample are used as the first screened samples (800) to replace the first original abnormal sample (600) in the first training set.

2. The information processing method as described in claim 1, wherein: The information collection step (201) collects information once every preset time interval (099); from the current time to the historical time, the first time interval (100), the second time interval (101), and up to the Tth time interval (10T) are obtained in sequence, where T is a natural number; The samples of the vehicle dataset are concatenated into a one-dimensional vector. One-hot encoding is used to process non-numerical features, and the data is normalized to the [0, 1] interval using minimum-maximum normalization. The information collection step (201) collects data through wired or wireless means / channels; the wireless means / channels include 4G, 5G, BLE, and Wi-Fi networks, and the wired means / channels include CAN and Ethernet.

3. The information processing method as described in claim 2, wherein: The convergence conditions include: the F1 score of the machine learning model on the test set no longer increases or the F1 score approaches a preset limit value according to a preset error threshold; The machine learning model includes a first type of gradient model and a second type of non-gradient model; the first type of gradient model includes a neural network model and a logistic regression model; the second type of non-gradient model includes a decision tree model and a random forest model; the optimization algorithm of the third search process (255) includes the adversarial machine learning algorithm FGSM and a genetic algorithm; The first sampling interval is a preset time length; the preset time length is not unique, and the preset time length includes 1 second; the information collection step (201) acquires the vehicle driving information (097) under preset working conditions; The vehicle driving information (097) includes vehicle speed (110), driver torque requirement (120), and gear information (1i0); The iterative process (300) includes an exchange refresh process (275), in which the exchange refresh process (275) replaces the corresponding sample of the same type before selection in the vehicle driving information (097) with the first selected sample (800) after selection.

4. A machine learning device, comprising: Information acquisition unit (001), model training unit (003), sample screening unit (005); The information collection unit (001) acquires vehicle driving information (097), which includes normal samples collected during normal vehicle driving. The normal samples constitute a normal sample set; if there is no first original abnormal sample set (600) composed of abnormal samples, then the abnormal sample set corresponding to the information intrusion process is constructed, and the sample features that play a real role under the network intrusion method are given through vehicle knowledge, that is, functional features, and an initial abnormal sample set is constructed: the normal value of the signal is estimated according to the functional features, and the value range of the tampered signal is determined according to the threshold of the deviation between the signal and the normal value; a value is randomly selected within the value range of the tampered signal, and the corresponding signal value in the original data is replaced, thereby constructing the original abnormal sample; and the vehicle dataset is initialized; the vehicle dataset includes a first training set, a second validation set, and a third test set; the features of the vehicle dataset include a first functional feature (601) and a second non-functional feature (602). The model training unit (003) trains a machine learning model using the normal samples and the abnormal samples; The sample screening unit (005) selects samples that meet the second non-functional characteristics as the first screened samples (800); replaces the samples to be optimized in the first original abnormal sample set (600) in the first training set as the first screened samples (800); repeats the information processing process in the model training unit (003) and the sample screening unit (005), and refreshes the data in a loop iteration process (300); The iterative process (300) optimizes the data until it meets the preset convergence conditions; The sample screening unit (005) also performs a first training process (215), which uses the normal samples and the abnormal samples to train a first discriminator (501) to simulate the machine learning model or uses the machine learning model as the first discriminator (501). The sample screening unit (005) also performs a second training process (235), which uses the abnormal samples and the first discriminator (501) to train a first generator (503); the first generator (503) generates only the second non-functional feature (602); the first functional feature corresponds to the sample feature associated with the network intrusion mode; The sample screening unit (005) further executes a third search process (255), which uses an optimization algorithm to search for the second non-functional feature; the optimization algorithm of the third search process (255) does not include the method of generative adversarial networks; the third search process obtains a first anomalous sample through the optimization algorithm; the second non-functional feature obtained by the optimization algorithm is concatenated with the functional feature in the original anomalous sample to generate the first anomalous sample; Random noise (701) is introduced into the first original abnormal sample set and the first generator (503) is sampled to obtain the second abnormal sample. The second non-functional feature output by the generator is concatenated with the functional feature in the feature vector input to the generator to generate the second abnormal sample. The first abnormal sample and the second abnormal sample are used as the first screened samples (800) to replace the first original abnormal sample (600) in the first training set.

5. The machine learning apparatus of claim 4, wherein: The information acquisition unit (001) acquires information once every preset time interval (099); from the current time to the historical time, the first time interval (100), the second time interval (101), and up to the Tth time interval (10T) are obtained in sequence, where T is a natural number.

6. The machine learning apparatus of claim 5, wherein: The machine learning model includes a first type of gradient model and a second type of non-gradient model; the first type of gradient model includes a neural network model and a logistic regression model; the second type of non-gradient model includes a decision tree model and a random forest model; the optimization algorithm of the third search process (255) includes the adversarial machine learning algorithm FGSM and a genetic algorithm; the first sampling interval is a preset time length; the preset time length is not unique, and the preset time length includes 1 second; the information collection step (201) obtains the vehicle driving information (097) under preset working conditions.

7. A computer storage medium, comprising: The storage medium itself used to store computer programs; When the computer program is executed by the microprocessor, it implements any of the information processing methods described in claims 1 to 3.

8. An on-board electronic controller, comprising: The machine learning apparatus as described in any of claims 4 to 6, or the machine learning apparatus as described in any of claims 4 to 6 and the computer storage medium as described in claim 7.

Images