An etc vehicle-road cooperation device and system

Abstract

The application discloses a battery system federal state evaluation method and system, and relates to the technical field of state evaluation. The method comprises the following steps: collecting historical operation data of a battery system, and constructing an operation scene identification framework according to the historical operation data; collecting real-time operation data, and determining a real-time operation scene by using the operation scene identification framework; matching a local state evaluation logic for each battery node based on the scene, and generating a battery node state cognition set; constructing a consistency calibration model based on a federal learning algorithm, inputting the battery node state cognition set into the model, and outputting a battery node state cognition set after collaborative calibration; and fusing the final battery system federal state evaluation result based on the calibrated battery node state cognition set. The application realizes the evaluation of the federal state of the battery system through the battery system federal state evaluation result.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation and privacy computing technology, and more specifically, to an ETC vehicle-road cooperative device and system. Background Technology

[0002] In cross-provincial ETC (Electronic Toll Collection) operations, vehicles using counterfeit license plates often employ diverse methods to evade tolls. For example, they might pass through a gantry in Province A by closely following another vehicle, and then use power spoofing in Province B to evade tolls. Since the behavioral characteristics of these counterfeit vehicles are scattered across the data of operators in different regions, effectively identifying them requires cross-domain sharing and collaborative analysis of multi-dimensional behavioral characteristics, including following distance, OBU signal strength, passage timestamps, and license plate image features.

[0003] In existing technologies, single differential privacy methods protect data privacy by adding random noise. However, this method has the following drawbacks: First, the introduction of noise reduces the accuracy of anomaly detection, and as the level of privacy protection increases, noise accumulation leads to a significant decline in model performance. Second, differential privacy methods cannot support real-time collaborative computation in encrypted states, making it difficult to meet the needs of real-time detection of cross-domain cloned vehicles. Furthermore, the original vehicle behavior feature data contains sensitive privacy information such as location trajectories and device fingerprints, and operators are constrained by data protection regulations and cannot directly share the original data.

[0004] Therefore, how to achieve high-accuracy cross-domain collaborative identification of cloned vehicles while protecting data privacy has become an urgent technical challenge. Summary of the Invention

[0005] This invention provides an ETC vehicle-road cooperative device and system, which solves the technical problem in related technologies that it is difficult to balance privacy protection and detection accuracy, and that a single differential privacy method cannot support real-time cooperative computing in encrypted state.

[0006] This invention discloses an ETC vehicle-road cooperative method, applied to a distributed system environment containing multiple regional operator terminals and a federated learning coordination server, comprising the following steps: Step 1: Obtain local ETC transaction records and multi-source sensor data, extract multi-dimensional vehicle behavior features, and generate a multi-dimensional vehicle behavior feature dataset; Step 2: Preprocess the multidimensional feature data to generate standardized feature vectors; Step 3: Input the standardized feature vector into the local neural network classifier for training and calculate the model gradient vector; Step 4: Perform differential privacy processing on the gradient vector to generate a noisy gradient vector; Step 5: Encrypt the noisy gradient vector using a homomorphic encryption algorithm to generate a ciphertext gradient vector; Step 6: Upload the ciphertext gradient vector to the federated learning coordination server, perform gradient aggregation in the ciphertext domain, and generate a global ciphertext gradient vector. Step 7: Decrypt the global gradient vector and update the parameters of the local neural network classifier; Step 8: Obtain real-time features of cross-domain vehicles and generate encrypted feature vectors; Step 9: Input the ciphertext feature vector into the trained recognition model to perform ciphertext domain inference and output the encrypted anomaly score; Step 10: In encrypted form, compare the anomaly score with a preset threshold to generate an encrypted suspected clone identifier; Step 11: Decrypt the suspected clone identifier and trigger subsequent processing based on the identifier result.

[0007] Furthermore, the neural network classifier consists of an input layer, multiple hidden layers, and an output layer, with the input layer receiving a standardized feature vector. ,in The feature dimension is represented; the hidden layer adopts a fully connected layer structure, the first... The output of the layer is The output layer uses the softmax function to map the last hidden state to a binary classification probability distribution.

[0008] Furthermore, the differential privacy processing includes: calculating the gradient vector. Norm as a sensitivity parameter Based on preset privacy budget parameters Calculate noise scale parameters From the Laplace distribution Mid-sampling generates noise vector The noise vector is superimposed on the original gradient vector to output a noisy gradient vector. .

[0009] Furthermore, the sensitivity parameter By applying gradient contributions to individual training samples The gradient for each sample is obtained by norm clipping. Calculate its If the norm exceeds the preset clipping threshold Then scale the gradient to .

[0010] Furthermore, the homomorphic encryption algorithm employs the Paillier homomorphic encryption algorithm, applying it to each element of the noisy gradient vector. Perform encryption operations ,in For public key, It is a random number.

[0011] Furthermore, the ciphertext domain gradient aggregation utilizes the additive homomorphism of Paillier homomorphic encryption to calculate... ,in For the first The aggregate weight of each operator, The number of operators participating in the aggregation.

[0012] Furthermore, the aggregation weight According to the The number of local training samples for each operator is determined, and the calculation is performed. ,in For the first The number of local samples for each operator.

[0013] Furthermore, during the ciphertext domain inference process, the ReLU function is approximated as a quadratic polynomial using a polynomial approximation method for the nonlinear activation function. The polynomial value is calculated in the ciphertext state using addition and constant multiplication operations supported by homomorphic encryption.

[0014] Furthermore, the comparison operation in the ciphertext state is implemented through a secure comparison protocol to calculate the ciphertext difference. The sign bit is extracted and used for determination through a threshold decryption protocol.

[0015] This invention also discloses an ETC vehicle-road cooperative system, comprising: multiple regional operator terminals and a federated learning coordination server; the regional operator terminals are equipped with data acquisition devices such as roadside units, lidar, and license plate recognition systems, used to collect vehicle behavior feature data, perform local model training, and encrypted domain inference; the federated learning coordination server is used to receive encrypted gradient vectors uploaded by each operator, perform weighted aggregation calculations in the encrypted domain, and distribute the global encrypted gradient vector to each operator.

[0016] This invention solves the technical problem of balancing privacy protection and detection accuracy in cross-domain vehicle clone recognition by introducing a dual privacy protection mechanism of differential privacy and homomorphic encryption, achieving the following technical effects: Differential privacy mechanisms are applied at the gradient vector level. They protect the statistical characteristics of the gradient through controllable Laplacian noise, preventing attackers from inferring the original vehicle behavior data from the gradient information and protecting the privacy at the data statistical characteristic level. Homomorphic encryption mechanisms are used at the gradient transmission and feature inference levels, enabling the federated learning coordination server to complete gradient aggregation calculations in the ciphertext domain, and allowing each operator to perform real-time feature inference and comparison in the ciphertext state, thus protecting the confidentiality of data during the calculation process. Differential privacy and homomorphic encryption mechanisms protect data security at different levels, and their synergistic effect enables operators to achieve cross-domain collaborative training and real-time detection without exposing original sensitive data. While protecting sensitive privacy information such as vehicle location trajectory and device fingerprint, it achieves high-accuracy cross-domain collaborative identification of cloned vehicles. Attached Figure Description

[0017] Figure 1 This is a flowchart of the federated learning and training phase of the ETC vehicle-road cooperative method of the present invention; Figure 2 This is a flowchart of the real-time detection and inference stage of the ETC vehicle-road cooperative method of the present invention. Detailed Implementation Detailed Implementation Method 1 In cross-provincial ETC (Electronic Toll Collection) operations, vehicles using counterfeit license plates often employ diverse methods to evade tolls. For example, they might pass through a gantry in Province A by closely following another vehicle, and then use power spoofing in Province B to evade tolls. Since the behavioral characteristics of these counterfeit vehicles are scattered across the data of operators in different regions, effectively identifying them requires cross-domain sharing and collaborative analysis of multi-dimensional behavioral characteristics, including following distance, OBU signal strength, passage timestamps, and license plate image features.

[0019] However, the aforementioned raw vehicle behavior data contains sensitive privacy information such as location trajectories and device fingerprints, and operators are constrained by data protection regulations and cannot directly share the raw data. In existing technologies, single differential privacy methods protect data privacy by adding random noise, but the introduction of noise reduces the accuracy of anomaly detection, and differential privacy methods cannot support real-time collaborative computation in encrypted states. Therefore, achieving high-accuracy cross-domain collaborative identification of cloned vehicles while protecting data privacy has become an urgent technical challenge.

[0020] According to an embodiment of this implementation, this implementation provides an ETC vehicle-road cooperative method, which is applied to a distributed system environment containing multiple regional operator terminals and a federated learning coordination server. Each regional operator terminal is equipped with data acquisition devices such as roadside units (RSU), lidar, and license plate recognition systems.

[0021] Step 1: Obtain local ETC transaction records and multi-source sensor data, extract multi-dimensional vehicle behavior features, and generate a multi-dimensional vehicle behavior feature dataset.

[0022] Operators in each region obtain vehicle transaction records from local ETC gantry systems, calculate vehicle following distance using LiDAR point cloud data, extract OBU signal strength and arrival timestamps from roadside unit logs, and generate license plate image feature vectors by feature encoding of license plate images using a license plate recognition system. These multi-source data are then integrated to form a multi-dimensional vehicle behavior feature dataset containing following distance, signal strength, timestamps, and license plate features.

[0023] It should be noted that the above following distance refers to the spatial distance between two adjacent vehicles passing through the gantry, calculated using lidar point cloud data; the above OBU signal strength refers to the power value of the signal transmitted by the vehicle-mounted unit received by the roadside unit, which can be used to detect power camouflage behavior.

[0024] In a collaborative identification system involving three operators in East China (Jiangsu), South China (Guangdong), and North China (Hebei), on January 15, 2024, a vehicle with license plate number Su A12345 was recorded as exhibiting abnormal traffic behavior in the gantry systems of the three regions. In East China, the gantry on the G2 Beijing-Shanghai Expressway (Suzhou section) detected the vehicle closely following another vehicle, with a following distance of only 2.1 meters. In South China, the gantry on the G4 Beijing-Hong Kong-Macau Expressway (Guangzhou section) detected an abnormally low OBU signal strength of -78dBm, suspected of power spoofing. In North China, the gantry on the G1 Beijing-Harbin Expressway (Tangshan section) recorded a logical contradiction between the vehicle's passage time and the timestamp of the preceding gantry. The three operators extracted multi-dimensional behavioral feature data of this vehicle and other historical vehicles from their local ETC systems. Table 1 shows some samples of the original feature data collected by each operator.

[0025] Table 1. Samples of multidimensional vehicle behavior feature data collected by various operators; ; Step 2: Preprocess the multidimensional feature data to generate standardized feature vectors.

[0026] By using feature normalization methods to map feature values ​​of different dimensions to a unified numerical range, and by using encoding methods to numerically represent categorical features, standardized feature vectors suitable for neural network inputs are generated.

[0027] Step 3: Input the standardized feature vector into the local neural network classifier for training and calculate the model gradient vector.

[0028] Each operator inputs its locally standardized feature vectors as training samples into a locally deployed neural network classifier, using the identifiers of vehicles using counterfeit license plates as labels for supervised learning training. The gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm, and the model gradient vector is output. .

[0029] Furthermore, the identification of the cloned vehicle is obtained from a database of cloned vehicle cases verified manually by each operator in the past. Samples of vehicles that have been manually confirmed as cloned in the historical records are marked as follows: For vehicles passing through normally, the sample is labeled as .

[0030] The aforementioned neural network classifier consists of an input layer, multiple hidden layers, and an output layer. The input layer receives the standardized feature vector generated in step 2. ,in The feature dimension is represented; the hidden layer adopts a fully connected layer structure, the first... The layer output is:

[0031] in, For the first The hidden state vector of the layer, and The first Layer weight matrix and bias vector, The ReLU activation function is used; the output layer uses the softmax function to map the last hidden state to a binary classification probability distribution.

[0032] in, This is the output two-dimensional probability vector. Indicates transpose. and These represent the probabilities of a vehicle being a legitimate vehicle and a vehicle using a counterfeit license plate, respectively. This represents the total number of hidden layers.

[0033] Furthermore, the feature dimension This is equal to the number of features extracted in step 1, including the sum of the dimensions of following distance, OBU signal strength, arrival timestamp, and license plate image feature vectors; the total number of hidden layers. The value is determined based on the data size, typically ranging from 2 to 4 layers; the aforementioned... Layer weight matrix ,in For the first The number of neurons in the layer For the first The number of neurons in a layer is typically set to 1 to 3 times the input dimension.

[0034] The aforementioned neural network classifier is trained using supervised learning, and the cross-entropy loss function is used to measure the difference between the model's predicted probability distribution and the true label.

[0035] in, The number of training samples. For sample index, For the first The true label of each sample and The model pairs are respectively the first and second. The probability of a normal vehicle and the probability of a vehicle with a cloned license plate are output for each sample. This represents the complete set of parameters for a neural network classifier.

[0036] The aforementioned neural network classifier uses the Adam optimization algorithm to update model parameters and calculates the loss function through backpropagation. The output gradient vector is the gradient relative to the gradients of the parameters of each layer. .

[0037] In the fifth round of federated training iterations, the three operators trained their neural networks on their respective local sample sets. The East China region had 8500 training samples, the South China region had 6200 training samples, and the North China region had 5300 training samples. All operators used the same network structure: the input layer dimension was... (Including following distance, OBU signal strength, arrival timestamp encoding, and license plate image feature vector in 4 dimensions), the hidden layer is set to 3 layers with 14, 10, and 6 neurons respectively, and the output layer dimension is 2. After local training, the loss function values ​​and some gradient vector elements calculated by each operator are shown in Table 2.

[0038] Table 2 Local training results for each operator; ; Step 4: Perform differential privacy processing on the gradient vector to generate a noisy gradient vector.

[0039] Calculate the gradient vector of Norm as a sensitivity parameter Based on preset privacy budget parameters The noise scale parameters are calculated using the Laplace mechanism:

[0040] in, Let be the scale parameter of the Laplace distribution. For the gradient vector Norm sensitivity For privacy budget parameters, The smaller the value, the higher the level of privacy protection.

[0041] Furthermore, the sensitivity parameter By applying gradient contributions to individual training samples Norm clipping yields, specifically, the gradient for each sample. Calculate its If the norm exceeds the preset clipping threshold Then scale the gradient to Thus ensuring .

[0042] Furthermore, the cropping threshold By applying unpruned gradients during the early stages of training The norm distribution is determined through statistical analysis; specifically, the gradient of all samples is calculated in the first few batches of training. Norm, selection The 95th percentile of the norm distribution is used as the clipping threshold. This protects privacy while avoiding information loss due to excessive cropping.

[0043] Furthermore, the privacy budget parameter The range of values ​​is In practical applications, when It offers strong privacy protection but is noisy. It provides moderate privacy protection while maintaining good model accuracy.

[0044] Based on the above scale parameters From the Laplace distribution Mid-sampling generates a noise vector with the same dimension as the gradient vector. The noise vector is superimposed on the original gradient vector to output a noisy gradient vector. .

[0045] It should be noted that the differential privacy processing described above prevents attackers from inferring the original behavioral characteristics of a single vehicle from the shared gradient information by injecting controllable random noise into the gradient vector, thereby protecting the statistical characteristics of the gradient from being leaked.

[0046] The three operators each perform differential privacy processing on the locally computed gradient vectors. The system has preset privacy budget parameters. Gradient clipping thresholds are determined through statistical analysis. Gradient in East China The norm is 2.834, which does not exceed the clipping threshold and therefore no clipping is needed; gradient in the South China region. The norm is 3.167, which also does not exceed the threshold; the gradient in the North China region The norm is 3.452, which exceeds the pruning threshold, and it needs to be scaled down to 3.2. Subsequently, each operator calculates the noise scale parameters. From the Laplace distribution Mid-sampling generates a noise vector, which is then superimposed on the original gradient or the clipped gradient. Table 3 shows a comparison of the gradient vectors before and after differential privacy processing.

[0047] Table 3. Comparison of gradient vectors before and after differential privacy processing; ; Step 5: Encrypt the noisy gradient vector using a homomorphic encryption algorithm to generate a ciphertext gradient vector.

[0048] Each operator uses the Paillier homomorphic encryption algorithm to add noise to the gradient vector. Encryption is performed to generate a ciphertext gradient vector that satisfies additive homomorphism. .

[0049] The aforementioned Paillier homomorphic encryption algorithm receives a noisy gradient vector. and public key As input, for each element in the gradient vector Perform encryption operations ,in Output the ciphertext gradient vector using random numbers. ,in These are the first, second, and m-th elements of the noisy gradient vector, respectively; the Paillier homomorphic encryption algorithm satisfies the additive homomorphic property. ,in Let be any two plaintext values ​​such that addition can be performed directly in the ciphertext state.

[0050] Furthermore, the public key Generated in the following manner: The federated learning coordination server selects two large prime numbers. and ,calculate Select generator satisfy , public key Distributed to various operators for encryption, with the private key kept on the coordinating server. Used for decryption, where prime numbers and The bit length is usually chosen to be 1024 bits or 2048 bits to ensure the strength of encryption security.

[0051] Furthermore, the random number The range of values ​​is ,Right now Less than And with Coprime positive integers, where The modulus in the Paillier encryption scheme is determined by uniformly and randomly selecting values ​​from a range. To ensure the semantic security of the ciphertext.

[0052] Step 6: Upload the ciphertext gradient vector to the federated learning coordination server, perform gradient aggregation in the ciphertext domain, and generate a global ciphertext gradient vector.

[0053] Each operator will generate locally-printed ciphertext gradient vectors Uploaded to the federated learning coordination server, the coordination server utilizes the additive homomorphism of Paillier homomorphic encryption to perform a process in the ciphertext field involving data from... The gradients of each operator are weighted and aggregated for calculation:

[0054] in, This is the global ciphertext gradient vector. For operators' indexing, For the first The encrypted gradient vector uploaded by each operator. For the first The aggregate weight of each operator, The number of operators participating in the aggregation.

[0055] Furthermore, the aggregate weight According to the The number of local training samples for each operator is determined, specifically, calculated... ,in For operators' indexing, For the first The number of local samples for each operator, and the aggregation weights satisfy the normalization constraint. and .

[0056] The coordination server distributes the aggregated global ciphertext gradient vector to each operator.

[0057] In the embodiments of this application, during the aforementioned ciphertext domain gradient aggregation process, the coordination server performs calculations entirely in ciphertext, and cannot obtain any operator's gradient plaintext information, thereby protecting the data privacy of all parties while achieving multi-party collaborative training.

[0058] After receiving the encrypted gradient vectors uploaded by the three operators, the federated learning coordination server first calculates the aggregation weights based on the number of samples from each operator. The number of samples in East China is 8500, in South China it is 6200, and in North China it is 5300, for a total of 20000 samples. Therefore, the aggregation weights are as follows: , , The coordinating server utilizes the additive homomorphic property of Paillier homomorphic encryption to compute weighted aggregations in the ciphertext domain. The global ciphertext gradient vector is generated and distributed to each operator. Table 4 shows the weight allocation process for ciphertext domain gradient aggregation.

[0059] Table 4. Ciphertext domain gradient aggregation process; ; Step 7: Decrypt the global gradient vector and update the parameters of the local neural network classifier.

[0060] Each operator uses its locally held private key to decrypt the received global ciphertext gradient vector, obtaining the plaintext global gradient vector. The model parameters of the local neural network classifier are then updated using this global gradient vector.

[0061] Furthermore, the model parameters are updated using the gradient descent method, and the update formula is as follows: ,in For the first Model parameters for the next iteration The learning rate parameter has a value range of 1. In practical applications, it is usually set To balance convergence speed and stability.

[0062] The three operators respectively processed the received global ciphertext gradient vector After decryption, all operators obtain the same plaintext global gradient vector, whose first five elements are [-0.0095, 0.0013, -0.0026, 0.0612, 0.0012]. Subsequently, each operator uses this global gradient vector and applies a learning rate... Update the model parameters of the local neural network classifier .

[0063] Repeat steps 3 through 7 until the model loss function converges or the preset number of iterations is reached to complete the federated training of the neural network classifier.

[0064] Furthermore, the convergence criterion for the model loss function is: calculating continuous... The next iteration ( The loss function value is typically taken as 3 to 5. ,in The first The next iteration, the... The iteration continues to the... The loss function value of the next iteration, if it satisfies and If , then it is determined to be convergent, where The preset convergence threshold is typically set to a value of [value to be filled in]. to .

[0065] Based on steps 1 to 7 above, the following steps for real-time detection are also included: Step 8: Obtain real-time features of cross-domain vehicles and generate encrypted feature vectors.

[0066] When a cross-domain vehicle passes through the gantry of the current operator's jurisdiction, real-time behavioral features such as the following distance, OBU signal strength, arrival timestamp, and license plate image features of the vehicle are obtained. The real-time features are preprocessed using the same feature normalization method as in step 2 to eliminate dimensional differences. The preprocessed real-time features are combined into a standardized feature vector, and then the Paillier homomorphic encryption algorithm is used to encrypt the feature vector to generate a ciphertext feature vector.

[0067] Step 9: Input the ciphertext feature vector into the trained recognition model to perform ciphertext domain inference and output the encrypted anomaly score.

[0068] The ciphertext feature vector generated in step 8 is input into the neural network classifier trained in step 7. Utilizing the ciphertext domain computation capability supported by homomorphic encryption, model inference calculations are completed without decrypting the feature data, and an anomaly score of the encrypted state is output. .

[0069] Furthermore, the ciphertext domain inference process is implemented in the following manner: For the neural network... Layer calculation First, the linear transformation of the ciphertext state is calculated using the additive homomorphism of Paillier homomorphic encryption. Specifically, for matrix multiplication, where For the index of hidden layer neurons, calculate For addition operations, calculate Then for nonlinear activation functions The ReLU function is approximated as a quadratic or cubic polynomial using a polynomial approximation method. By utilizing the addition and constant multiplication operations supported by homomorphic encryption, polynomial values ​​are calculated in the ciphertext state, thereby obtaining the hidden layer output of the ciphertext state. Finally, the output layer is calculated in the same way to obtain the encrypted anomaly score. .

[0070] Furthermore, the polynomial approximation coefficients The least squares method is used to find the effective input range of the ReLU function. The above fitting is used to obtain the ReLU function value by uniformly sampling several points within the effective input interval. ,in Using the sampling point index, construct a least squares optimization problem. The approximate coefficients are obtained by solving the problem, and are usually taken as... , , This ensures that the approximation error within the effective input interval is less than 0.1.

[0071] Step 10: In encrypted state, compare the anomaly score with a preset threshold to generate an encrypted suspected clone identifier.

[0072] Using a secure comparison protocol, anomaly scoring is performed in encrypted form. The algorithm compares the data with a preset suspected clone threshold to generate an encrypted suspected clone identifier. The value of the identifier is true, which indicates that the vehicle is suspected of using a fake license plate, and the value of the identifier is false, which indicates that the vehicle has passed the inspection.

[0073] The aforementioned secure comparison protocol receives encrypted anomaly scores. and preset threshold First, the ciphertext difference is calculated as input. Then, the sign of the difference ciphertext is determined. If the difference is greater than zero, the encrypted identifier value is output. Otherwise output .

[0074] Furthermore, the symbol determination is achieved through the following secure multi-party computation protocol: the ciphertext difference is... Decryption permissions are secretly shared among multiple participants, each holding a share of the private key. They collectively decrypt the sign bit of the difference using a threshold decryption protocol without revealing the actual value of the difference. Specifically, bit decomposition is used to represent the difference in binary form, and the highest bit is extracted as the sign bit. If the sign bit is 0, the difference is positive, and the result is output. If the sign bit is 1, the difference is negative, and the output is... Throughout the process, all participants only receive symbolic information and cannot obtain the original score value.

[0075] Furthermore, the preset threshold Based on the classification probability distribution of the model trained in step 7 on the validation set, specifically, by analyzing the abnormal rating distributions of known cloned and legitimate vehicles in the validation set, the quantile that minimizes the sum of the false positive and false negative rates is selected as the threshold. The threshold value range is [insert range here]. The anomaly score is represented by the model prediction probability value output in step 9.

[0076] Furthermore, the validation set is obtained by dividing the historical labeled data of each operator in chronological order. Specifically, each operator's local dataset is divided into a training set and a validation set in an 8:2 ratio, where the training set is used for model training in steps 3 to 7, and the validation set is used for thresholding. The determination of the threshold and the evaluation of the model performance should ensure that the ratio of cloned vehicle samples to normal vehicle samples in the verification set is consistent with the ratio in the actual application scenario, so as to ensure the rationality of the threshold setting.

[0077] After 50 rounds of federated training, the model converged on the validation set. The threshold for suspected clone was determined by analyzing the abnormal rating distribution on the validation set. On January 20, 2024, the South China regional gantry system detected a vehicle with license plate number Su A12345 passing through the Shenzhen section of the Guangzhou-Shenzhen Expressway gantry again. The system obtained the vehicle's real-time behavioral characteristics: following distance of 2.8 meters, OBU signal strength of -76dBm, passage timestamp of 2024-01-20 16:32:18, and license plate image feature vector of [0.83, 0.74, 0.91, 0.68]. The South China regional operator preprocessed the real-time features according to the normalization method in step 2: the following distance was normalized to 0.14, the OBU signal strength to 0.22, and the timestamp encoding to 0.69, generating a standardized feature vector. The feature vector is then encrypted using the Paillier homomorphic encryption algorithm to generate a ciphertext feature vector. .

[0078] The system inputs the encrypted feature vector into a trained neural network classifier for encrypted domain inference. For the calculation of the first hidden layer, the first row of the weight matrix... , bias is ,in For feature dimension indexing, the homomorphic properties of multiplication and addition in homomorphic encryption are used to compute... For the ReLU activation function, a quadratic polynomial approximation is used. Calculated in encrypted state The ciphertext field calculations are performed sequentially across the three hidden layers and the output layer, ultimately yielding the encrypted anomaly score. .

[0079] Calculate in encrypted state using a secure comparison protocol. Extract the sign bit using a threshold decryption protocol, determine if the difference is positive, and output an encrypted suspected clone identifier. Table 5 shows the feature data and inference process of the vehicle's real-time detection.

[0080] Table 5. Real-time detection feature data and inference process; ; Step 11: Decrypt the suspected clone identifier and trigger subsequent processing based on the identifier result.

[0081] The encrypted suspected cloned vehicle identifier is decrypted. When the decrypted identifier value is true, a manual review process is triggered, and the characteristics of the suspected vehicle are shared with the relevant operators in encrypted form for cross-domain joint verification.

[0082] The South China regional operator has encrypted suspected clone identifiers. Decryption yielded an identifier value of 1, indicating that the license plate Su A12345 was suspected of being a counterfeit plate. The system immediately triggered a manual review process, pushing the vehicle's passage records and on-site captured images to the inspectors' workstations. Simultaneously, the South China region shared the vehicle's encrypted feature data with the operators in the East China and North China regions. The East China region retrieved the vehicle's historical records from January 15th and found an abnormal behavior of a following distance of 2.1 meters at the Suzhou gantry. The North China region retrieved the historical records from January 15th and found that the vehicle's passage time at the Tangshan gantry was 08:15:33, while its passage time at the Guangzhou gantry was 14:56:23, a time span of 6 hours and 41 minutes. However, the actual distance between the two locations is approximately 2100 kilometers, and the normal travel time would be over 20 hours, presenting a temporal and spatial logical contradiction. After comprehensive analysis by the three parties' inspectors, it was confirmed that the vehicle was a counterfeit plate vehicle, blacklisted, and reported to the national ETC system. Table 6 shows the summary results of the cross-regional joint verification.

[0083] Table 6 Summary of cross-domain joint verification results; ; This implementation overcomes the contradiction between privacy protection and detection accuracy that leads to the challenge of cross-domain vehicle clone recognition by introducing a dual privacy protection mechanism that combines differential privacy and homomorphic encryption.

[0084] Specifically, since a single differential privacy method can only protect the statistical features of data by adding noise, the accumulation of noise will reduce the accuracy of anomaly detection, and differential privacy methods cannot support real-time computation in the encrypted state. In this embodiment, the differential privacy mechanism is applied at the gradient vector level, protecting the statistical features of the gradient through controllable Laplacian noise to prevent attackers from inferring the original vehicle behavior data from the gradient information; the homomorphic encryption mechanism is applied at the gradient transmission and feature inference level, enabling the federated learning coordination server to complete gradient aggregation calculation in the encrypted domain, and enabling each operator to complete real-time feature inference and comparison in the encrypted state.

[0085] Therefore, differential privacy and homomorphic encryption protect data security at different levels. Differential privacy protects privacy at the statistical feature level, while homomorphic encryption protects data confidentiality during the computation process. The synergistic effect of the two mechanisms enables operators to achieve cross-domain collaborative training and real-time detection without exposing original sensitive data, thereby solving the problem of balancing privacy protection and detection accuracy in cross-domain collaborative identification of cloned vehicles.

Claims

1. An ETC vehicle-road cooperative method, characterized in that, This system, applied to a distributed system comprising multiple regional operator terminals and a federated learning coordination server, includes the following steps: Acquiring local ETC transaction records and sensor data, extracting multi-dimensional vehicle behavior features, generating a multi-dimensional vehicle behavior feature dataset, and performing standardized preprocessing; Inputting the standardized feature vector into a local neural network classifier for training, and calculating the model gradient vector; Performing differential privacy processing on the gradient vector, calculating a noise scale parameter based on the gradient vector's sensitivity parameter and privacy budget parameter, sampling from a Laplace distribution to generate a noise vector, and superimposing it onto the gradient vector to generate a noisy gradient vector; Encrypting the noisy gradient vector using a homomorphic encryption algorithm to generate a ciphertext gradient vector; Uploading the ciphertext gradient vector to the federated learning coordination server, and using the additive homomorphism of homomorphic encryption in the ciphertext domain to perform weighted aggregation calculations on the gradients from each operator to generate a global ciphertext gradient vector; Each operator decrypting the global ciphertext gradient vector and using the global gradient vector to update the parameters of its local neural network classifier; Wherein, the noise scale parameter is the ratio of the sensitivity parameter to the privacy budget parameter, and the sensitivity parameter is obtained by norm clipping the gradient contribution of a single training sample.

2. The method according to claim 1, characterized in that, The multidimensional vehicle behavior features include: the following distance between vehicles calculated using lidar point cloud data, the signal strength of the vehicle-mounted unit received by the roadside unit, the arrival timestamp of the vehicle passing through the gantry, and the license plate image feature vector generated by the license plate recognition system through feature encoding of the license plate image.

3. The method according to claim 1, characterized in that, The differential privacy processing of the gradient vector includes: calculating the norm of the gradient of each sample; if it exceeds a preset clipping threshold, scaling the gradient to the clipping threshold. The clipping threshold is determined by statistically analyzing the norm distribution of the un-clipped gradient in the early stage of training and selecting a preset percentile of the norm distribution.

4. The method according to claim 1, characterized in that, The homomorphic encryption algorithm is the Paillier homomorphic encryption algorithm, which performs encryption operations on each element in the noisy gradient vector and outputs a ciphertext gradient vector that satisfies the additive homomorphism, so that addition operations can be performed directly in the ciphertext state.

5. The method according to claim 1, characterized in that, In the weighted aggregation calculation, the aggregation weight is determined based on the number of local training samples of each operator, specifically the ratio of the number of local samples of each operator to the total number of samples of all operators. The aggregation weight satisfies the normalization constraint.

6. The method according to claim 1, characterized in that, Repeat the steps of calculating gradient vectors, differential privacy processing, homomorphic encryption, ciphertext domain aggregation, and parameter updating until the model loss function converges or reaches the preset number of iterations, thus completing federated training; the convergence criterion for the loss function is that the change in the loss function value over several consecutive iterations is less than the preset convergence threshold.

7. The method according to claim 1, characterized in that, It also includes a real-time detection step: acquiring real-time features of cross-domain vehicles and performing standardized preprocessing; using a homomorphic encryption algorithm to encrypt the feature vectors to generate ciphertext feature vectors; inputting the ciphertext feature vectors into a trained neural network classifier for ciphertext domain inference; and outputting an encrypted anomaly score.

8. The method according to claim 7, characterized in that, The ciphertext domain reasoning process includes: calculating the linear transformation of the ciphertext state using the additive homomorphism of homomorphic encryption; approximating the activation function as a polynomial using a polynomial approximation method for nonlinear activation functions; and calculating the polynomial value in the ciphertext state using the addition and constant multiplication operations supported by homomorphic encryption.

9. An ETC vehicle-road cooperative device, characterized in that, The system includes a memory and a processor. The memory stores an ETC vehicle-road cooperative program. When the ETC vehicle-road cooperative program is executed by the processor, it implements the steps of the ETC vehicle-road cooperative method as claimed in any one of claims 1 to 8.

10. An ETC vehicle-road cooperative system, used to execute the ETC vehicle-road cooperative method according to any one of claims 1-8, characterized in that, include: The feature acquisition module is used to acquire local ETC transaction records and sensor data, extract multi-dimensional vehicle behavior features, generate a multi-dimensional vehicle behavior feature dataset, and perform standardized preprocessing. The local training module is used to input standardized feature vectors into a local neural network classifier for training and to calculate the model gradient vector; the differential privacy processing module is used to calculate the noise scale parameter based on the sensitivity parameter and the privacy budget parameter, sample noise vectors from the Laplace distribution and superimpose them on the gradient vector to generate a noisy gradient vector. The homomorphic encryption module is used to encrypt the noisy gradient vector and generate a ciphertext gradient vector. The federated aggregation module, deployed on the coordination server, is used to perform weighted aggregation calculations on gradients from various operators in the ciphertext domain using the additive homomorphism of homomorphic encryption, generating a global ciphertext gradient vector. The parameter update module is used to decrypt the global ciphertext gradient vector and update the parameters of the local neural network classifier using the global gradient vector.

Images