Highway vehicle-road cooperation accurate charging method considering information interaction privacy protection

Abstract

The invention relates to a precise charging method for information interaction privacy protection in a highway vehicle-road cooperation scene, and the method comprises the following steps: a roadside center distributes exclusive public keys and initialized weight information to vehicles entering an RSU coverage range; a vehicle end fits a trajectory model by using own vehicle information and a bidirectional long-short-term memory network model, judges an abnormal value, encrypts a model network weight by using a public key to form a private key and uploads the private key to a roadside center, and the roadside center decrypts a vehicle node network model parameter weight within a service range and performs average aggregation; clustering is carried out according to aggregation weights of all parties, a vehicle client is randomly selected from each class group, aggregation parameters are encrypted and issued by using an exclusive public key to return to a vehicle, the weight of the vehicle is updated until a global model converges, the vehicle calculates the track length by using a final model, and the cost is calculated in combination with related information such as the vehicle type. According to the method, the user privacy of vehicle-road information interaction is protected while the toll collection and clearing based on the precise path of the expressway is realized.

Description

Technical Field

[0001] The present invention belongs to the fields of vehicle coordination, vehicle safety and traffic management. Background Art

[0002] With the commissioning of the national highway network toll collection system, the construction and management departments are faced with a common problem: toll collection and clearing. After the highway is opened to traffic, vehicles that need to travel on the highway will be charged tolls according to the toll standards specified by the region. The highway operation and management departments can use the collected tolls to recover costs, repay loans and expenses spent in the highway construction phase, and maintain the daily maintenance and repair expenses of the highway. The current mainstream toll collection method is polysemy path recognition charging, that is, through DSRC technology, radio frequency technology, video license plate technology, vehicle information is collected at the intermediate identification station or intermediate identification equipment, and the vehicle discrete points are used to determine the maximum probability of the vehicle's driving trajectory. The shortest path algorithm is used between discrete points, but this method may have problems with missing labels and mislabeling. In recent years, many provinces have established highway polysemy path recognition systems based on license plate recognition or wireless communication technologies.

[0003] With the expansion of the network structure and the removal of provincial toll stations, the ambiguity of the travel path has increased significantly, which is likely to cause abnormalities in highway toll collection. The method of ambiguous path identification and charging cannot meet the needs of the development of large-scale highway networks, and the toll collection needs to be more rational. With the development of intelligent transportation and vehicle-road collaboration technology, roadside units have more dense, accurate and efficient vehicle position collection equipment, and the highway toll system is expected to change from traditional ambiguous path probability charging to more accurate segmented charging. Arranging vehicle identification devices at the fork sections of the highway and splitting the roads of the ambiguous paths into segments can construct the real path of the vehicle and solve the problem of ambiguous paths. However, when vehicles converge on the highway, there will be frequent lane changes and even illegal U-turns, which will cause a waste of highway resources. Therefore, a charging method based on accurate vehicle trajectories is needed.

[0004] There are differences in the built-in data of different vehicles. When modeling vehicle trajectories, there are limitations to the modeling of a single vehicle. The built-in information stored in the vehicle, the vehicle's running trajectory, the driver's gender, the driver's age, the non-driving related tasks performed, etc. are private data and are not suitable for centralized data modeling. While information is being exchanged, transmitting the original data back to the computing center will not only cause a lot of communication overhead, but also pose a risk of privacy leakage. In order to protect user privacy while rationalizing charges, it is necessary to consider establishing a joint vehicle trajectory modeling model with multi-party data security.

[0005] The present invention fully considers the privacy-preserving information interaction of group heterogeneity and explores precise charging strategies based on joint modeling of vehicle trajectories. The present invention is suitable for vehicle-road collaborative scenarios with vehicle-side fusion perception (generally L3 autonomous driving and above vehicles) and vehicle-road collaborative perception. Summary of the invention

[0006] Table 1 shows the original data for training the models of each vehicle client, which is stored in the vehicle client. The vehicle type and grade information is inherent information of the vehicle and will not change for differentiated charging management services. The remaining variables may change, and the vehicle obtains the built-in storage information. Table 1 Basic information fields stored (raw data for model training)

[0007] There may be correlation between data variables, and PCA dimensionality reduction processing is needed to screen the six principal components with the highest to lowest cumulative variance contribution rates. The purpose is to improve the convergence speed and accuracy of the model and reduce the total time cost.

[0008] Vehicle client training uses a bidirectional LSTM network model to predict trajectories, selects feature inputs, constructs training sets and test sets, and initializes model parameters. Independent pre-training of the client model can obtain multi-step prediction results.

[0009] In the present invention, the server and the client use cryptographic methods to encrypt and protect the client parameters before transmission.

[0010] The FL server randomly selects a 160-bit elliptic curve finite field that conforms to Abelian group operations, selects a base point, and calculates the order of the base point.

[0011] When the roadside server detects that a vehicle enters a restricted area, the server generates a random legal private key. Different clients correspond to different legal private keys. The public key is generated by the elliptic curve encryption algorithm and sent to the client vehicle.

[0012] After the customer vehicle receives the public key, it hashes the sent weight (SHA 256) and regards it as the private key, encrypts it using the public key pair, and sends the encrypted ciphertext to the FL server.

[0013] After receiving the ciphertext, the server uses the private key to decrypt the ciphertext to form plaintext, and then uses hash calculation to output the client weight information.

[0014] All vehicle clients within the service range of the server are checked into the FL server. Each device downloads the initialized neural network weights from the server and performs a round of learning to obtain the neural network model weights of each client and give them to the server.

[0015] The server aggregates the client neural network weights to obtain the global neural network model weights, and calculates the performance of the global model at the server test set data level. Due to differences in client data and models, the global model may diverge after multiple global trainings. It is necessary to reward or punish the performance of each client model on the server, calculate the MAE error function of each client model on the server data, map the global MAE data from 1 to 0 to score the client, and give the client weight bonus during federation aggregation. The federation aggregation initialization weight is 1 / N, where N is the number of clients.

[0016] After receiving the model uploaded by the vehicle, the server updates the model backup on the server, calculates the Manhattan similarity matrix of the model weights of each vehicle client, performs hierarchical clustering on the matrix, classifies the heterogeneous clients and gives corresponding classification labels to form K clusters.

[0017] A random sampling of 1 is performed on each type of client cluster in the clustering data, and a total of K vehicle clients are extracted.

[0018] The server sends the global neural network model weights to all clients. The encryption and decryption process is the reverse process of the server collecting weight information.

[0019] The server converts the weight after global aggregation into curve points and regards it as a private key, encrypts it with the public key pair, and sends the encrypted ciphertext to the client. After receiving the ciphertext, the client uses the private key to decrypt the ciphertext to form plaintext, and then uses the curve point conversion to output the weight information after global aggregation.

[0020] The client stores the weight backup and replaces the global weight. The server sends data requests to the selected clients according to the results of cluster sampling. The invited clients perform a round of training with the new model. The trained model weight information will be sent to the server again, and the client neural network model weight data of the new round of training stored on the server will be updated. The clustering, federated aggregation process, random selection, and local training process are repeated until the global neural network model converges. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a charging trajectory prediction model for the vehicle client of the present invention.

[0022] Figure 2This is a charging trajectory calculation model based on federated learning on the server side of the present invention.

[0023] Figure 3 This is the charging system of the present invention based on the federal trajectory model.

[0024] Figure 4 It is the vehicle-road information elliptic curve encryption algorithm of the present invention.

[0025] Figure 5 Functional module of the present invention DETAILED DESCRIPTION

[0026] The data stored in the client has n indicators (corresponding to n factors), m evaluation objects (corresponding to m records), and the principal component analysis (PCA) method is used for data preprocessing in order to reduce the dimension and communication overhead.

[0027] First, the data is normalized and mapped to the range [0, 1].

[0028] For m variables, calculate the covariance matrix to form a covariance matrix with m rows and m columns. Solve for the eigenvalues ​​and eigenvectors of the covariance matrix.

[0029] Select the eigenvectors corresponding to k eigenvalues ​​to form the matrix W n·k , and then perform matrix calculation Z m·k =X m·n W n·k .

[0030] Determine the coefficients of the linear combination of the indicators in each component β is the element corresponding to each principal component variable, and α is the eigenvalue of the corresponding principal component variable.

[0031] The corresponding coefficient matrix can be obtained, namely

[0032] The variance contribution rate of the k principal components in the linear model is calculated by using the variance contribution rate of the principal components. The comprehensive score coefficient is Select N principal components with high comprehensive score coefficients.

[0033] During client training, the vehicle reads the locally stored trajectory data set and divides the trajectory data into a training set and a test set with a time step of Δt.

[0034] The training set and test set are input into the bidirectional LSTM model for training. Here is an explanation of the specific working process of the two-layer LSTM:

[0035] The dual-layer LSTM architecture of the present invention includes a forward layer and a backward layer. The forward layer uses the input iteration from time t' to t'-1, and the backward layer uses the reverse input calculation. The output of the two layers is calculated using the following update equation: t =σ g (W f x t +U f h t-1 +b f ),i t =σ g (W i x t +U i h t-1 +b i ),o t =σ g (W o x t +U o h t-1 +b o ),

[0036] W f , W i , W o is the weight matrix that maps the hidden layer to the three gates and the input unit state, U f , U i , U o is the weight matrix that maps the state of the migration unit to the three gates and the input unit, b f , b i , b o , b c is the deviation vector, σ g It is the gate activation function, uses the average function activation, and uses the tanh or sigmoid function.

[0037] The unit output status is The layer output calculation formula is h t =o t tanh(C t ).

[0038] The previous layer generates the output vector After the front and back layers are passed, the output vectors of the front and back layers are averaged and aggregated, and the result can be expressed as Y T =[y T-1 , ..., y T-1 ].

[0039] The global and local loss functions use the MSE loss function

[0040] When updating weights, a truncated time-based back-propagation algorithm can be used. The client weight sequence is The loss function of the local federated training model is based on the different data tables contained in the client (the number of tables is Q c ) is the sum of the losses. In order to prevent overfitting, the loss function is regularized. The loss function is

[0041] The roadside server randomly generates parameters a, b, p, B for each vehicle client, randomly selects an integer K from 20 to 30, and selects an elliptic curve for y 2 ≡x 3 +ax+b(mod p), these will be sent to each client as public.

[0042] The vehicle client randomly selects the base B, encodes the information m through the hash SHA3-256, and then converts it into the B base (usually 16), which can convert the information into a fixed 64-bit character placeholder, that is, m = m0 + m1B + m2B 2 +…+m N (B N+1 ≤n), the code values ​​a, b, c, d, e, f correspond to the values ​​10, 11, 12, 13, 14, 15.

[0043] The vehicle client rewrites the plaintext information using m′=mK+j so that m 3 +am+b is a perfect square number, satisfying the Legendre symbol At this time, the plain text information can correspond to The probability of failure is only 2 -K , where (m+1)K<p.

[0044] The vehicle client calculates y 2 ≡x 3 +ax+b(mod p) to get the curve code point (x, y).

[0045] The roadside service end randomly selects a point on the elliptic curve finite field as the base point G and calculates the finite field order n (n is a prime number) of G.

[0046] The roadside server randomly selects Q c integer k c (k c <n), each client corresponds to one, and uses the elliptic curve encryption algorithm to output the public key K, where K c =k c G.

[0047] The roadside server sends the public key K to the corresponding client. c , E p (a, b), G information.

[0048] The vehicle client accepts the server's public key K c , E p (a, b), G information, weight information Perform SHA256 hash encoding to obtain As E p M on (a, b) c point.

[0049] The vehicle client selects an integer r less than n c As the private key, calculate Get the ciphertext The vehicle client sends the ciphertext to the server.

[0050] After receiving the ciphertext sent by the client, the roadside server calculates:

[0051] The server only needs to calculate Combined with hash decoding, the points on the curve can be mapped to plaintext.

[0052] At initialization, all clients in the federated framework participate in training. After e epochs of local training, the gradient descent function in client c is The learning rate preset by client c is η, so after e epochs, the weight of the local node is

[0053] The federated learning framework will perform client selection, and the single selected client set is The size is C, the global loss function Further simplification

[0054] The aggregation model of the federation framework is Passing parameters from the server The aggregation weight is taken as, W = z_score(W).

[0055] After the aggregation operation, the server will The model is sent to all clients, and the model parameters are cryptographically protected. During the second client local aggregation, only the selected clients need to be trained. After continuous iterative training, the global model converges.

[0056] The converged global model is imported into each client. The client collects real-time trajectory data. The trajectory may be missing. The data n steps before the last data update time t′ before the data missing is used as the input of the trained bidirectional LSTM model to predict the trajectory coordinates at time t′+1.

[0057] The client calculates the precise fee based on the completed vehicle trajectory, vehicle type, discount type and other information, encrypts the fee information and sends it to the server, which performs a background check on the charging information.

Claims

1. The present invention proposes a highway vehicle-road collaborative accurate charging method that takes into account information interaction privacy protection, which is characterized by: The roadside server randomly generates a dedicated public key and initializes weight information; The vehicle uses its own vehicle information to train the model using trajectories, fit trajectories, and determine outliers; The vehicle segment encryption model network weights form a private key and upload it to the roadside center; The roadside service end decrypts the information uploaded by vehicles within the service range; The roadside center clusters the weights of all parties, extracts clients using unbiased sampling, sets a reward function based on the training error for weighted aggregation, and updates the global model parameters; The vehicle updates its weights, repeats training and uploads weight information until the global model converges; The vehicle uses the convergence model to calculate the trajectory length and abnormal conditions, packages the vehicle type, encrypts it and uploads it to the roadside center, which calculates the driving costs.

2. The weighted aggregation method based on client model clustering sampling proposed in claim 1 is characterized by: The roadside service calculates the Manhattan distance of the model weights of each vehicle client to form a similarity matrix, performs hierarchical clustering on the matrix, classifies heterogeneous clients and gives corresponding classification labels. For each cluster, the number of samples of the vehicle client is linearly mapped to the sampling probability, and the vehicle client of each cluster is unbiasedly sampled with data of 1 according to the sampling probability. The model training error of each client is linearly mapped to 0-1 interval data as the coefficient of global model aggregation. The roadside service updates the parameters according to the aggregated model.

3. The vehicle client model training method according to claim 1 is characterized in that: The bidirectional LSTM model consists of two LSTM networks, one forward and one reverse, which use both forward and backward information to model the sequence data, thereby capturing richer semantic dependencies. They process the forward and reverse order of the input sequence respectively, and concatenate their outputs as the final output.

4. The vehicle information proposed in claim 1 is characterized in that: Contains relevant information about the vehicle and neighboring vehicles, obtained by the vehicle perception system. The vehicle information includes the vehicle's longitudinal position, vehicle's lateral position, vehicle's longitudinal position, vehicle's lateral position, vehicle's longitudinal speed, and vehicle's lateral speed; neighboring vehicle information includes: the following distance between the vehicle and the front and rear vehicles, and the license plate ID of the adjacent vehicle.

Images