Security authentication method for in-vehicle wireless devices based on deep metric learning

Abstract

The present invention discloses a method for secure authentication of vehicle-mounted wireless devices based on deep metric learning, and belongs to the technical field of machine learning. The method comprises: pre-processing radio frequency signals of multiple vehicle-mounted wireless devices to generate a training data set; constructing an open class authentication model based on improved class anchor loss; removing a part of the training data set as improper signal samples, and sending the remaining samples to network training respectively to achieve network model optimization; and authenticating the signals of the vehicle-mounted wireless devices to be authenticated. The present invention achieves efficient identification of vehicle-mounted wireless devices and malicious attack devices while improving the feature clustering effect by constructing an open class recognition model with improved class anchor loss, and balances the recognition rates of the two. At the same time, sparsity is introduced by using a multi-scale feature extraction module, and features are extracted by using different convolution kernels between adjacent convolution layers, and the results of multiple layers of convolution are spliced ​​as the input of the next layer, thereby improving the authentication accuracy of the model without increasing the computational overhead.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and more specifically to a vehicle-mounted wireless device security authentication method based on deep metric learning. Background Art

[0002] In recent years, smart cars have made significant progress in functions such as autonomous driving and cockpit interaction. With the continuous innovation of technology, cars not only have basic driving functions, but also gradually realize highly intelligent user experience. The continuous evolution of electronic and electrical architecture has led to an increasing number of communication nodes in the car, and performance requirements have also increased accordingly. However, traditional wired communication technology is facing increasingly obvious bottlenecks. Specifically, the length of the wiring harness increases with the increase in the number of communication nodes in the car, resulting in a significant increase in overall cost, curb weight and labor installation costs. At the same time, the increase in connection points limits the flexibility of the installation position of components, resulting in a decrease in the freedom of vehicle design.

[0003] In view of the above problems, in order to effectively meet the needs of cost control, flexible deployment and weight reduction in the vehicle manufacturing process, automobile manufacturers have gradually shifted the focus of research on intelligent vehicle communication methods to wireless communication to replace some wired communications. This can improve production and assembly efficiency, significantly reduce the risk of wiring harness connector failure, and also optimize the flexibility of vehicle layout.

[0004] The new generation of in-vehicle wireless short-range communication technology can support ultra-low latency, ultra-high reliability, and precise synchronization of in-vehicle wireless services. Specifically, through wireless short-range communication technology, in-vehicle wireless devices can transmit data to the vehicle gateway in real time to interact with the roadside unit, thereby supporting vehicle-road collaborative application scenarios such as autonomous driving, smart traffic management, and smart parking systems. However, the scope of in-vehicle services supported by existing wireless communication technologies is mostly limited to cabin entertainment and information assistance services such as wireless active noise reduction, panoramic surround view, and lighting control. The data of vehicle-road collaborative services involving driving and safety is still dominated by traditional wiring harness transmission methods. This is because driving safety-related services have extremely high requirements for the security and reliability of data transmission, and malicious attack devices may send improper wireless signals to attack the vehicle gateway, thereby affecting the normal operation of the vehicle and even threatening driving safety. Therefore, it is urgent to introduce wireless gateway authentication technology to ensure that only authenticated devices can communicate with the vehicle gateway. This authentication mechanism will help prevent malicious attack devices from accessing the gateway, further protect the safety of the vehicle and the integrity of the data, to support safer and more reliable vehicle-road collaborative applications, and ensure that the intelligent vehicle system can operate efficiently in a safe environment.

[0005] RF fingerprint authentication technology achieves identity authentication by extracting tiny individual features hidden in the signal and storing them in a fingerprint feature library. These features are derived from the distortion characteristics of electronic components such as I / Q imbalance and power amplifier nonlinearity, providing a more stable and difficult-to-forge authentication method for vehicle gateways. However, in vehicle-mounted wireless communications, there are a large number of communication nodes in the vehicle, and the number of malicious attack devices may continue to increase over time, making it extremely difficult to establish a comprehensive fingerprint feature library for identity authentication. To this end, it is necessary to authenticate unknown signals and dynamically update the existing feature library. Summary of the invention

[0006] The present invention provides a vehicle-mounted wireless device security authentication method based on deep metric learning, which aims to cope with the increasing number of malicious attack device authentications by improving the open class recognition model (ORCA-Net) with class anchor loss, and effectively distinguish vehicle-mounted wireless devices from malicious attack devices through feature extraction and optimization model training, thereby improving authentication accuracy and robustness.

[0007] The present invention adopts the following technical solution:

[0008] The in-vehicle wireless device security authentication method based on deep metric learning includes the following steps:

[0009] S1. Acquire radio frequency signals corresponding to a plurality of vehicle-mounted wireless devices, and pre-process the radio frequency signals of the plurality of vehicle-mounted wireless devices to generate a training data set;

[0010] S2. Establish an open class authentication model based on improved class anchor loss, wherein the open class authentication model includes a dimension reorganization module, a multi-scale feature extraction module, a plurality of two-dimensional convolutional layers, a fully connected layer and an improved class anchor loss function, wherein the formula of the improved class anchor loss function is shown in the following formula (1):

[0011] (1)

[0012] in, It describes the distance of each type of sample from its preset center point. Represents the distance from the negative class sample to the center point of the non-preset category, is the boundary interval between positive and negative samples; C is the total number of categories of the sample; j is an index, traversing except y All categories except y is the label of each class of samples.

[0013] S3, randomly extracting some categories from the training data set of step S1 as improper signal samples, which do not participate in the training of the model, and sending the remaining samples as known vehicle-mounted wireless devices to the network training respectively, calculating the authentication error through the improved class anchor loss function, completing the learning process after multiple iterations, and realizing the optimization of the network model;

[0014] S4. Collect the signal of the vehicle-mounted wireless device to be authenticated, and use the same model to extract its radio frequency fingerprint feature, compare the feature with the record in the database, and confirm whether the signal participates in the establishment of the identity information database. If so, the identity of the vehicle-mounted wireless device is compliant, otherwise, the vehicle-mounted wireless device is considered to have an attempt to improperly access the gateway; further, the feature is compared with the preset category anchor vector through a classification measurement method for authentication.

[0015] Preferably, the above open class authentication model includes a dimension restructuring module, a multi-scale feature extraction module, four two-dimensional convolutional layers, a fully connected layer and an improved class anchor loss function.

[0016] Furthermore, the multi-scale feature extraction module sends the outputs of the two-layer convolutional layers with dimension reorganization into three paths into a three-way network module composed of four two-dimensional convolutional layers. The convolution kernel sizes of the two convolutional layers are set to 1 and 5 respectively, the step size is set to 1, and the padding size is 1; after the outputs of the three-way network module are merged, the features are spliced.

[0017] In a preferred implementation scheme, the above step S3 specifically includes the following process: the pre-processed radio frequency signal of the vehicle-mounted wireless device is subjected to channel expansion processing of two convolution layers in the dimension reorganization module, and then sent to the multi-scale feature extraction module, and the output result is sequentially passed through four two-dimensional convolution layers and then dimensionally reorganized to obtain a feature map result; then, the Euclidean distance between the feature representation of each category of data and the pre-set center anchor point of each category is calculated in the distance vector calculation layer to obtain the distance vector of the input data; at the same time, the feature map result is sent to the fully connected layer, and after the output is activated by the activation function, the inner product between the output and the distance vector is obtained as the distance metric; each layer of the neural network is followed by an instance regularization layer and a ReLU function; finally, the network is optimized by an improved class anchor loss function.

[0018] In a preferred implementation scheme, the specific determination method for comparing the feature with the preset category anchor vector for authentication by the classification measurement method in the above step S4 is as follows: when the minimum value of the rejection score of a sample for all known vehicle-mounted wireless devices is less than the authentication threshold, it is determined that it belongs to a compliant vehicle-mounted wireless device, and the corresponding index value is assigned to the sample as the vehicle-mounted wireless device category, and the access gateway is allowed to perform RSU communication; otherwise, it is determined to be a malicious attack device and the access gateway is denied to participate in RSU communication.

[0019] It can be seen from the above description of the present invention that, compared with the prior art, the present invention has the following advantages:

[0020] 1. The present invention constructs an open class recognition model (ORCA-Net) with improved class anchor loss, which improves the feature clustering effect while achieving efficient recognition of vehicle-mounted wireless devices and malicious attack devices, and balances the recognition rates of the two. At the same time, sparsity is introduced by using a multi-scale feature extraction module. By using different convolution kernels between adjacent convolution layers to extract features and splicing multi-layer convolution results as input to the next layer, the authentication accuracy of the model is improved without increasing the computational overhead. Compared with the traditional CAC loss, the NCA loss proposed in the present invention has a higher recognition rate when the number of improper signals increases. When the number of malicious attack devices accounts for 50% in the test set, the recognition rate can reach 97.15%. And as the proportion of malicious attack devices increases, the average recognition rate of the algorithm of the present invention can still be higher than 90%, which can effectively realize the security authentication of vehicle-mounted wireless devices.

[0021] 2. The improved N-pair loss function of the present invention adds the absolute distance between the sample and the corresponding category anchor center to accelerate the feature clustering of the sample while minimizing the occurrence of outlier feature values; it also adds a margin parameter to add a boundary interval between positive and negative samples to better guide feature clustering and assist in improving the performance of the open class recognition algorithm.

[0022] 3. The multi-scale feature extraction module of the present invention adopts different convolution kernel sizes and convolution layer splicing to extract rich multi-scale features, enhance the model performance, improve the authentication accuracy without increasing the computational overhead. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is the basic framework diagram of the present invention.

[0024] Figure 2 This is the network structure diagram of ORCA-Net of the present invention.

[0025] Figure 3 This is a structural diagram of the multi-scale feature extraction module of the present invention.

[0026] Figure 4 It is a coordinate diagram of the convergence of the NCA loss function of an embodiment of the present invention.

[0027] Figure 5 This is a coordinate diagram of the ORCA-Net recognition performance when a malicious attack device is added according to an embodiment of the present invention. DETAILED DESCRIPTION

[0028] Refer to the following Figure 1The specific implementation of the present invention is described. In order to fully understand the present invention, many details are described below, but for those skilled in the art, the present invention can be implemented without these details. For well-known components, methods and processes, they are not described in detail below.

[0029] The in-vehicle wireless device security authentication method based on deep metric learning provided in this embodiment. Figure 1 As shown, firstly, the RF signal data sets of different vehicle-mounted wireless devices are preprocessed, and the representative features corresponding to each vehicle-mounted wireless device are extracted through the RF fingerprint authentication model based on deep metric learning, in order to form an identity information library of vehicle-mounted wireless devices. In the network structure design, we fully extracted the features of the RF signals of vehicle-mounted wireless devices at various scales, and introduced sparsity through the multi-scale feature extraction module, which improved the authentication accuracy without increasing the model calculation overhead. Assisted by the improved anchor-like loss function, the loss function is used to optimize the network model to realize the vehicle-mounted wireless device security authentication method under the increase of malicious attack devices. Specifically, it includes the following steps:

[0030] S1. In the preprocessing stage, different radio frequency signal data sets of vehicle-mounted wireless devices are collected for preprocessing to generate a training data set.

[0031] S2. Establish an open-set recognition model based on improved anchor loss (Open-Set Recognition NetworkBased on Anchor Loss, ORCA-Net).

[0032] As an example, Figure 2 As shown in the figure, ORCA-Net adopts a multi-layer two-dimensional convolutional neural network model as a whole, including a dimension reorganization module, a multi-scale feature extraction module, four two-dimensional convolutional layers, a fully connected layer and an improved anchor-like loss function.

[0033] As an example, Figure 3 As shown in the figure, the implementation of the multi-scale feature extraction module is to divide the output of the two convolution layers in the dimension reorganization module into three inputs. The three-way network module consists of four two-dimensional convolution layers. The convolution kernel sizes of the two convolution layers are set to 1 and 5 respectively, the step size is set to 1, and the padding size is 1. After the output of the three-way network module is merged, the features are spliced ​​to increase the number of channels in the overall data dimension.

[0034] First, the preprocessed radio frequency signal of the on-board wireless device will be reorganized in data dimension to adapt to the input of the two-dimensional convolution layer. After the input signal is processed by the channel expansion of the two convolution layers in the dimension reorganization module, it is sent to the multi-scale feature extraction module. The output results are sequentially passed through four two-dimensional convolution layers and then dimensionally reorganized to obtain the feature map results. Then, the Euclidean distance between the feature representation of each category of data and the pre-set center anchor points of each category is calculated in the distance vector calculation layer to obtain the distance vector of the input data. At the same time, the feature map result is sent to the fully connected layer. After the output is activated by the activation function, the inner product between it and the distance vector is calculated as the distance metric. Each layer of the neural network is followed by the instance normalization layer (IN) and the ReLU function. Finally, the verification error is calculated by the improved class anchor loss. After multiple iterations, the learning process is completed to achieve network model optimization. The purpose of this method is to add the absolute distance between the sample and the corresponding category anchor center to accelerate the feature clustering of the sample, while minimizing the appearance of outlier feature values.

[0035] S3. Randomly extract some categories from the training data set of step S1 as improper signal samples, and do not participate in the training of the model. The remaining samples are sent to the network training as known vehicle-mounted wireless devices. The authentication error is calculated by the improved class anchor loss. After multiple iterations, the learning process is completed to achieve network model optimization.

[0036] The improved class anchor loss (New Class Anchor Loss, NCA Loss) minimizes the distance from the known vehicle wireless device data to its corresponding category anchor point and maximizes the distance to other category anchor points. In the improved class anchor loss (NewClass Anchor Loss, NCA Loss), the anchor loss in the traditional class anchor clustering loss method (ClassAnchorClustering Loss, CAC Loss) is retained, the N-pair loss is modified, and the network is optimized according to the final improved NCA loss function.

[0037] The following formula (1) is the modified N-pair loss function formula:

[0038] (1)

[0039] in, C is the total number of categories of the sample; j is an index, traversing except y All categories except y is the label of each class of samples. describes the distance of each type of sample from its preset center point, and Represents the distance between the negative class sample and the center point of the non-preset class. The last part of the loss function is based on the triplet loss, and a margin parameter is added to add a boundary interval between the positive and negative class samples to better guide feature clustering and assist in improving the performance of the open class recognition algorithm. The final improved NCA loss of the present invention is shown in formula (2), where is the retained anchor loss:

[0040] (2)

[0041] S4. During the authentication phase, the system will collect the signal of the vehicle wireless device to be authenticated, and use the same model to extract its radio frequency fingerprint feature. First, the feature is compared with the record in the database to confirm whether the signal participates in the establishment of the identity information database. If it participates, the device identity is compliant, otherwise the vehicle wireless device is considered to have an attempt to improperly access the gateway. Further, the feature is compared with the preset category anchor vector for authentication through a classification measurement method. When the minimum value of the rejection score of a sample for all known vehicle wireless devices is less than the authentication threshold, it is determined to be a compliant vehicle wireless device, and the corresponding index value is assigned to the sample as the vehicle wireless device category, allowing access to the gateway for RSU communication; otherwise, it is determined to be a malicious attack device and denied access to the gateway for RSU communication.

[0042] As an example, Figure 4 The improved loss function and its various parts change with the increase of training times. The data in the figure show that ORCA-Net reaches convergence when the number of iterations is 300. Although the loss function fluctuates slightly when the number of iterations continues to increase, it is basically stable at around 0.8, and the anchor loss penalty term eventually tends to 0, indicating that after training, ORCA-Net can continuously gather features toward the expected anchor point.

[0043] As an example, Table 1 shows the results of the impact of different structural components on network recognition performance, analyzes the contribution of components in the network structure to the overall performance of the model, marks the multi-scale feature extraction module in ORCA-Net as MsF, and marks the concatenated convolutional layers other than the MsF layer as CNNet for experiments. Randomly extract 5 categories of signals from the two data sets as malicious attack device samples for testing, and the remaining samples are used as known vehicle-mounted wireless device samples and the data sets are divided. The data in the table are all accuracy results. Among them, "√" and "×" represent whether the module is added during this training process, with a total of 4 rows of data, marked as "1-4" from top to bottom. First, the results of rows 1, 3 and 2, 4 show that the multi-scale feature extraction module MsF can extract sample features with high granularity and enhance the degree of distinction between samples, thereby achieving a significant improvement in recognition performance. Secondly, the data in rows 1 and 2 show that the performance of CAC loss is better than NCA loss when MsF is not used; after adding this module, as shown in rows 3 and 4, the network performance under the guidance of NCA loss is greatly improved. Because MsF improves the feature extraction effect, it also increases the difficulty of clustering detailed features. The improved NCA adds the measurement of absolute distance between similar samples on the basis of the original loss. Under the premise of ensuring the clustering of known vehicle wireless device samples, the distance from malicious attack devices to other types of centers is increased as much as possible, thereby obtaining recognition results that are better than CAC loss. The fourth row of data is the final result of ORCA-Net.

[0044] Table 1 The impact of different structural components on network recognition performance (%)

[0045]

[0046] As an example, Figure 4 This is the change in recognition performance when the number of malicious attack devices increases while the number of vehicle-mounted wireless device sample categories in the dataset is fixed. Figure 5 The unknown class represents the category of malicious attack devices, and the known class represents the category of vehicle-mounted wireless device signals; the dark blue bar graph represents the recognition rate of known vehicle-mounted wireless devices, the light blue represents the recognition rate of malicious attack devices, and the red five-pointed star-marked dotted line represents the average recognition rate in each case. The results show that as the number of malicious attack devices increases, the overall network recognition rate gradually decreases. This is because there are certain differences between different malicious attack devices, but when performing malicious attack device identification tasks, all malicious attack devices need to be rejected as one category. When the number of malicious attack devices increases, their feature distribution also begins to become widespread. It is difficult to find a decision distance that can completely distinguish these two types of samples, so the recognition rate also decreases.

[0047] In summary, the present invention proposes a method for secure authentication of in-vehicle wireless devices based on deep metric learning. By constructing an open class recognition model (ORCA-Net) with improved class anchor loss, it improves the feature clustering effect while realizing efficient recognition of in-vehicle wireless devices and malicious attack devices, and balances the recognition rates of the two. At the same time, sparsity is introduced by using a multi-scale feature extraction module. By using different convolution kernels between adjacent convolution layers to extract features, and splicing multi-layer convolution results as the input of the next layer, the authentication accuracy of the model is improved without increasing the computational overhead. Compared with the traditional CAC loss, the NCA loss proposed in this embodiment has a higher recognition rate when the number of improper signals increases. When the number of malicious attack devices accounts for 50% in the test set, the recognition rate can reach 97.15%. And as the proportion of the number of malicious attack devices increases, the average recognition rate of the algorithm proposed in this embodiment can still be higher than 90%, which can effectively realize the secure authentication of in-vehicle wireless devices.

[0048] The above is only a specific implementation of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial changes to the present invention using this concept shall be deemed as an infringement of the protection scope of the present invention.

Claims

1. A vehicle-mounted wireless device security authentication method based on deep metric learning, characterized in that: The following steps are involved: S1. Acquire radio frequency signals corresponding to a plurality of vehicle-mounted wireless devices, and pre-process the radio frequency signals of the plurality of vehicle-mounted wireless devices to generate a training data set; S2. Establish an open class authentication model based on improved class anchor loss, wherein the open class authentication model includes a dimension reorganization module, a multi-scale feature extraction module, a plurality of two-dimensional convolutional layers, a fully connected layer and an improved class anchor loss function, wherein the formula of the improved class anchor loss function is shown in the following formula (1): (1) in, It describes the distance of each type of sample from its preset center point. Represents the distance from the negative class sample to the center point of the non-preset category, is the boundary interval between positive and negative samples; C is the total number of categories of the sample; j is an index, traversing except y All categories except y is the label of each class of samples; S3, randomly extracting some categories from the training data set of step S1 as improper signal samples, which do not participate in the training of the model, and sending the remaining samples as known vehicle-mounted wireless devices to the network training respectively, calculating the authentication error through the improved class anchor loss function, completing the learning process after multiple iterations, and realizing the optimization of the network model; S4. Collect the signal of the vehicle-mounted wireless device to be authenticated, and use the same model to extract its radio frequency fingerprint feature, compare the feature with the record in the database, and confirm whether the signal participates in the establishment of the identity information database. If so, the identity of the vehicle-mounted wireless device is compliant, otherwise, the vehicle-mounted wireless device is considered to have an attempt to improperly access the gateway; further, the feature is compared with the preset category anchor vector through a classification measurement method for authentication.

2. The in-vehicle wireless device security authentication method based on deep metric learning as claimed in claim 1, characterized in that: The open class authentication model includes a dimension restructuring module, a multi-scale feature extraction module, four 2D convolutional layers, a fully connected layer and an improved class anchor loss function.

3. The in-vehicle wireless device security authentication method based on deep metric learning according to claim 1 or 2, characterized in that: The multi-scale feature extraction module sends the outputs of the two-layer convolutional layers with dimension reorganization into three paths into a three-way network module composed of four two-dimensional convolutional layers. The convolution kernel sizes of the two convolutional layers are set to 1 and 5 respectively, the step size is set to 1, and the padding size is 1; after the outputs of the three-way network module are merged, the features are spliced.

4. The in-vehicle wireless device security authentication method based on deep metric learning as claimed in claim 2, characterized in that: The step S3 specifically The invention comprises the following processes: the pre-processed radio frequency signal of the on-board wireless device is processed by the channel expansion of the two convolution layers in the dimension reorganization module, and then sent to the multi-scale feature extraction module, and the output result is sequentially passed through four two-dimensional convolution layers and then dimensionally reorganized to obtain the feature map result; then, the Euclidean distance between the feature representation of each category of data and the pre-set center anchor point of each category is calculated in the distance vector calculation layer to obtain the distance vector of the input data; at the same time, the feature map result is sent to the full connection layer, and after the output is activated by the activation function, the inner product between the output and the distance vector is calculated as the distance metric; each layer of the neural network is followed by an instance regularization layer and a ReLU function; finally, the network is optimized by an improved class anchor loss function.

5. The in-vehicle wireless device security authentication method based on deep metric learning as claimed in claim 1, characterized in that: The specific determination method for comparing the feature with the preset category anchor vector for authentication by the classification measurement method in step S4 is as follows: when the minimum value of the rejection score of a sample for all known vehicle-mounted wireless devices is less than the authentication threshold, it is determined that it belongs to a compliant vehicle-mounted wireless device, and the corresponding index value is assigned to the sample as the vehicle-mounted wireless device category, and the access gateway is allowed to perform RSU communication; otherwise, it is determined to be a malicious attack device and the access gateway is denied to participate in RSU communication.

Images