Network connection vehicle privacy protection communication optimization method based on locality sensitive hashing

Abstract

The invention provides a network connection vehicle privacy protection communication optimization method based on locality sensitive hashing, relates to the technical field of network connection vehicles, and realizes obvious reduction of communication cost in distributed cooperative training through efficient compression and transmission of model parameters. And meanwhile, the data privacy of each edge device in the collaborative awareness system is protected. Local sensitive hashing (LSH) technology is adopted to dynamically cluster local model parameters of clients in federated learning, communication is carried out by taking a clustering center (centroid) of each group of parameters as a representative, and only the clustered centroid is transmitted, so that the amount of data uploaded in the communication process is remarkably reduced. According to the method, a model structure or a training mechanism does not need to be changed, universality and high efficiency are achieved, the communication efficiency can be effectively optimized on the premise that the model performance and precision are kept, and the use cost of bandwidth resources is remarkably reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of connected vehicles, and specifically to an optimized method for privacy protection communication of connected vehicles based on locality-sensitive hashing. Background Art

[0002] Federated learning is a distributed machine learning technology aimed at jointly training a model by multiple participants without directly sharing the original data. In this process, each participant uses its own local dataset and completes joint learning by exchanging and integrating model parameters or updates. Federated learning can not only effectively protect data privacy but also fully exploit the massive information in distributed datasets to generate a model with higher accuracy and generalization ability. Currently, federated learning has been widely applied in many fields such as intelligent devices, healthcare, and financial services, successfully solving the problems of data privacy and security. However, as a distributed learning method, federated learning still faces challenges such as data heterogeneity, high communication overhead, and low model synchronization efficiency, which limit its performance and scalability in practical applications.

[0003] Locality-Sensitive Hashing (LSH) is an efficient approximate nearest neighbor search algorithm for high-dimensional data. By designing specific hash functions, LSH can map similar data points to the same or adjacent hash buckets with a high probability, thus accelerating approximate nearest neighbor search and significantly reducing the complexity of data storage and computation. In federated learning, the LSH technique can quickly cluster model parameters into several groups with high similarity, use the cluster centers of each group as representatives, and only transmit these centroids to reduce the communication data volume, thereby significantly reducing the overhead of all-to-all communication. After communication, by combining a residual compensation mechanism to restore the clustering error, the training efficiency can be greatly improved while ensuring the model accuracy. The core function of this method is to compress data to effectively reduce communication overhead, which is particularly suitable for distributed cooperation scenarios with limited communication resources.

[0004] An intelligent connected vehicle refers to a vehicle that has the capabilities of perception, communication, decision-making, and control, and can realize information sharing and collaborative operation through the Internet and real-time communication between vehicles. Relying on sensor technology, artificial intelligence algorithms, and advanced communication technology, intelligent connected vehicles can achieve functions such as autonomous driving, traffic flow optimization, and intelligent navigation, significantly improving driving safety, efficiency, and comfort. However, in the process of data exchange and sharing of intelligent connected vehicles, a large amount of sensitive information is transmitted, such as location information, driving behavior data, etc. If this data is obtained by criminals, it may lead to serious consequences such as privacy leakage and vehicle safety threats. Therefore, how to effectively protect the data privacy of intelligent connected vehicles while ensuring communication efficiency has become a key issue in current research and applications. Summary of the Invention

[0005] Aiming at the deficiencies of the prior art, the purpose of the present invention is to propose an optimized method for privacy protection communication of connected vehicles based on locality-sensitive hashing, including:

[0006] Step 1: Collect the vehicle data of K connected vehicles. The vehicle data includes the position information of the connected vehicle and the distances between all objects within a preset radius centered on the connected vehicle and the connected vehicle; preprocess the vehicle data to obtain a set of feature vectors corresponding to the vehicle data, that is, local data;

[0007] Step 2: Each connected vehicle downloads a neural network model from the server as the initial local model. The initial model parameters of all the initial local models of the connected vehicles are the same. Take the initial local model as the current local model, set the initial number of communications, and take the initial number of communications as the current number of communications;

[0008] Step 3: In the current number of communications, for each connected vehicle, at the client of the connected vehicle, input the local data into the current local model to obtain the predicted positions of the objects within a preset radius centered on the connected vehicle. Then, according to the predicted positions of the objects within a preset radius centered on the connected vehicle and the labels representing the true positions of the objects, obtain the loss function. Update the model parameters by the gradient descent method according to the loss function. Input the local data into the model with the updated model parameters, and then update the model parameters again. Repeat the above operations until the number of times of model parameter update reaches the preset threshold to obtain the embedding vector θ of the final model parameters of each connected vehicle k ;

[0009] Step 4: According to the embedding vector of the final model parameters of each connected vehicle, allocate all the parameters in the embedding vector to multiple hash buckets. For each hash bucket, calculate the cluster center to obtain multiple cluster centers of each connected vehicle. Then, calculate the residuals of each parameter in the hash bucket to obtain a set of residuals, and further obtain the set of residuals of each connected vehicle;

[0010] Step 5: Upload all the cluster centers of each connected vehicle to the server through the client of the connected vehicle;

[0011] Step 6: The server receives the multiple cluster centers uploaded by each client and calculates the global cluster center C global ;

[0012] Step 7: The server sends the global cluster center C global to each connected vehicle;

[0013] Step 8: For each connected vehicle, the client of the connected vehicle receives the global cluster center C global, update the parameters according to the residuals of each parameter in the residual set, so as to obtain the embedding vectors of the updated model parameters of each connected vehicle;

[0014] Step 9: Increment the current communication count by one to obtain a new current communication count. Update the current local model according to the embedding vectors of the updated model parameters of each connected vehicle to obtain a new current local model. Determine whether the new current communication count is less than a preset threshold. If the new current communication count is greater than or equal to the preset threshold, use the new current local model as the final local model. If the new current communication count is less than the preset threshold, use the new current communication count as the current communication count, use the new current local model as the current local model, and return to execute Step 3.

[0015] Optionally, in Step 1, preprocess the vehicle data to obtain a set of feature vectors corresponding to the vehicle data, including:

[0016] Filter, denoise, and extract feature vectors from the vehicle data to obtain a set of feature vectors corresponding to the vehicle data.

[0017] Optionally, in Step 3, update the model parameters by the gradient descent method according to the loss function, which is specifically implemented by the following formula:

[0018]

[0019] where η represents the learning rate, represents the embedding vector of the model parameters of the k-th connected vehicle at the t-th iteration, represents the embedding vector of the model parameters of the k-th connected vehicle at the (t + 1)-th iteration, is the gradient of the loss function L of the k-th connected vehicle at the t-th iteration k at .

[0020] Optionally, Step 4 specifically includes:

[0021] Step 4.1: For the embedding vector θ k of the final model parameters of each connected vehicle, process each parameter θ k in the embedding vector θ k,i through L different locality-sensitive hashing algorithms to obtain L hash values, which are specifically implemented by the following formula:

[0022] h l = argmax i∈{±1,±2,...,±d} |R l θ k,i | i , l = 1, 2,..., L;

[0023] Among them, h l is the hash value calculated by the l-th locality-sensitive hashing algorithm, d represents the dimension, argmax is a function used to select the index that makes |R l θ k,i | i the largest, and R l is the random rotation matrix of the l-th locality-sensitive hashing algorithm;

[0024] The L hash values are concatenated to obtain a hash signature, which is specifically implemented through the following formula:

[0025] signature = (h1, h2,..., h L );

[0026] Furthermore, 2*d hash signatures of the embedding vector θ k are obtained, and the parameters with the same hash signature in the embedding vector are assigned to the same hash bucket;

[0027] Step 4.2: For the parameters in each hash bucket, calculate the cluster center of this hash bucket, which is specifically implemented through the following formula:

[0028]

[0029] Among them, C j is the cluster center of the j-th hash bucket, n j is the number of parameters in the j-th hash bucket, and θ k,j,m is the m-th parameter in the j-th hash bucket;

[0030] In this way, multiple cluster centers of each connected vehicle are obtained;

[0031] Step 4.3: For each parameter in the same hash bucket, calculate the residual between the parameter and the cluster center. The residuals of the embedding vectors of all model parameters form a residual set, which is specifically implemented through the following formula:

[0032] Δ k,j,m = θ k,j,m - C j ;

[0033] Δcluster j = {Δ k,j,,m};

[0034] Among them, Δ k,j,m is the residual between the m-th parameter in the j-th hash bucket and the cluster center, and Δcluster j is the residual set of the j-th hash bucket;

[0035] In this way, multiple residual sets of each connected vehicle are obtained.

[0036] Optionally, step 6 specifically includes:

[0037] Step 6.1: The server receives multiple cluster centers uploaded by each client and determines the weight of each cluster center.

[0038] Step 6.2: Aggregate the cluster centers by the weighted average method according to the weights of the cluster centers to obtain the global cluster center C global .

[0039] Optionally, step 6.2 is specifically implemented by the following formula:

[0040]

[0041] where C k,j is the j-th cluster center of the k-th client, W k,j is the weight of C k,j and B k is the total number of cluster centers uploaded by the k-th client.

[0042] Optionally, in step 8, the parameters are updated according to the residuals of each parameter in the residual set, and it is specifically implemented by the following formula:

[0043] θ' k,j,m = C global + Δ k,j,m ;

[0044] where Δ k,j,m is the residual of the m-th parameter in the j-th hash bucket of the k-th connected vehicle in the residual set, and θ' k,j,m is the updated parameter of the m-th parameter in the j-th hash bucket of the k-th connected vehicle.

[0045] The beneficial effects of adopting the above technical solutions are as follows:

[0046] The present invention proposes an optimized method for connected vehicle communication based on locality-sensitive hashing. By efficiently compressing and transmitting model parameters, it realizes a significant reduction in communication costs in distributed collaborative training, while protecting the data privacy of each edge device in the collaborative perception system. The locality-sensitive hashing LSH technology is used to dynamically cluster the local model parameters of each client in federated learning, and communicate with the representative of the cluster center (centroid) of each group of parameters, only transmitting the centroids after clustering, thus significantly reducing the amount of data uploaded during the communication process. This method does not require modification of the model structure or training mechanism, has universality and high efficiency, can effectively optimize communication efficiency while maintaining the performance and accuracy of the model, and significantly reduces the usage cost of bandwidth resources. Description of the Drawings

[0047] Figure 1Schematic flow chart of an optimized method for privacy protection communication of connected vehicles based on locality - sensitive hashing in an embodiment of the present invention. Specific embodiments

[0048] The following further describes the specific embodiments of the present invention in detail with reference to the accompanying drawings and embodiments. The following embodiments are used to illustrate the present invention, but are not used to limit the scope of the present invention.

[0049] In view of the problems existing in the prior art, the present invention provides an optimized method for privacy protection communication of connected vehicles based on locality - sensitive hashing. Combining Figure 1 , it may include the following steps:

[0050] Step 1: Collect vehicle data of K connected vehicles. The vehicle data includes the location information of the connected vehicle and the distances between all objects within a preset radius centered on the connected vehicle and the connected vehicle. The objects can be other vehicles, traffic lights, road obstacles, etc. Among them, the vehicle data can be collected by various sensors such as cameras, lidars, and ultrasonic sensors configured on each connected vehicle. Pre - process the vehicle data. Specifically, filter, denoise, and extract feature vectors from the vehicle data to obtain a set of feature vectors corresponding to the vehicle data, that is, local data. Pre - processing the vehicle data is used to improve the quality of the data and generate a set of feature vectors for neural network training.

[0051] Among them, each intelligent connected vehicle communicates with the central server through the client. In the traditional method, it is usually necessary to upload the local data collected by each client to the server. This approach not only may make it difficult to effectively guarantee the privacy and security of the uploaded data, but also significantly increase the communication cost due to the large amount of local data. However, in this embodiment, the local data of the K intelligent connected vehicles is always stored in the client and does not need to be uploaded to the server, while still being able to ensure accurate and efficient communication between the client and the server.

[0052] Step 2: Each connected vehicle downloads a neural network model from the server as the initial local model. The initial model parameters of the initial local models of all connected vehicles are the same. Take the initial local model as the current local model, set the initial number of communications, and take the initial number of communications as the current number of communications;

[0053] Among them, the neural network model is determined according to the actual situation. For example, a convolutional neural network (CNN) model.

[0054] Step 3: In the current communication times, for each connected vehicle, at the client of the connected vehicle, input the local data into the current local model to obtain the predicted positions of the objects within the preset radius centered on the connected vehicle, and then, based on the predicted positions of the objects within the preset radius centered on the connected vehicle and the labels representing the true positions of the objects, obtain the loss function, which is specifically represented by the following formula:

[0055]

[0056] where N is the total number of training samples, and the training samples include local data and labels representing the true positions of the objects, x r is the r-th local data, θ k represents the embedding vector of the model parameters of the k-th connected vehicle, f(x r ; θ k ) is the predicted position of the object within the preset radius centered on the connected vehicle, y r is the label representing the true position of the object corresponding to the r-th local data, is a function used to measure the difference between the predicted position of the object within the preset radius centered on the connected vehicle and the label representing the true position of the object. For example, mean squared error, cross entropy, etc. The specific function used depends on the model and task used;

[0057] It should be noted that in the present invention, the local data includes the position information of the connected vehicle and the distances between all the objects within the preset radius centered on the connected vehicle and the connected vehicle. The true label is the label representing the true position of the object corresponding to the local data. In actual use, data can be collected, labels can be set, and the type of neural network can be set according to actual usage requirements.

[0058] Update the model parameters according to the loss function by the gradient descent method, which is specifically implemented by the following formula:

[0059]

[0060] where η represents the learning rate, represents the embedding vector of the model parameters of the k-th connected vehicle at the t-th iteration, represents the embedding vector of the model parameters of the k-th connected vehicle at the (t + 1)-th iteration, is the gradient of the loss function L k at for the k-th connected vehicle at the t-th iteration.

[0061] Input the local data into the model with updated model parameters, and then update the model parameters again. Repeat the above operations until the number of iterations reaches the preset threshold to obtain the final embedding vector θ k ; of the model parameters of each connected vehicle

[0062] That is to say, after obtaining the updated model parameters, the local data is input into the model with updated parameters to obtain the predicted positions output by the model, and then the loss function is calculated. Based on the loss function, the model parameters are updated, and the process of inputting local data into the model and updating the parameters is repeated until the number of times of updating the model parameters reaches the preset number of times, that is, the number of training iterations reaches the preset number of local training rounds.

[0063] It should be noted that multiple local data are obtained in step 1. Correspondingly, there are also multiple true labels corresponding to the local data, which are used to implement the training of the neural network model.

[0064] Step 4: According to the embedding vectors of the final model parameters of each connected vehicle, all the parameters in the embedding vectors are assigned to multiple hash buckets. For each hash bucket, the clustering center is calculated to obtain multiple clustering centers of each connected vehicle. Then, the residuals of each parameter in the hash bucket are calculated to obtain a set of residuals, and further a set of residuals of each connected vehicle is obtained;

[0065] Step 4.1: For the embedding vector θ of the final model parameters of each connected vehicle k , through L different locality-sensitive hashing algorithms LSH, for example, the Cross-Polytope Hashing function, each parameter θ k in the embedding vector θ k,i is processed to obtain L hash values, which are specifically implemented through the following formula:

[0066] h l = argmax i∈{±1,±2,...,±d} |R l θ k,i | i , l = 1, 2,..., L;

[0067] where h l is the hash value calculated by the l-th locality-sensitive hashing algorithm, d represents the dimension, argmax is a function used to select the index that makes |R l θ k,i | i the largest, and R l is the random rotation matrix of the l-th locality-sensitive hashing algorithm;

[0068] The L hash values are concatenated to obtain a hash signature, which is specifically implemented through the following formula:

[0069] signature = (h1, h2,..., h L );

[0070] Thus, the embedding vector θ is obtainedk For the 2*d hash signatures, the parameters with the same hash signature in the embedding vectors are assigned to the same hash bucket;

[0071] That is to say, for each parameter in the embedding vector of each connected vehicle, L hash values are calculated, and a hash signature is obtained by concatenation. Furthermore, each parameter has a hash signature, and the parameters with the same hash signature are divided into one hash bucket.

[0072] Among them, the present invention can select a variety of LSH algorithms. When selecting a specific LSH algorithm, the most suitable hash method can be selected according to the distribution characteristics of the model parameters and the application scenario to ensure that the hashed parameters can effectively reflect the characteristics of the original parameters.

[0073] Step 4.2: Calculate the clustering center of each hash bucket for the parameters in it, which is specifically implemented by the following formula:

[0074]

[0075] Among them, C j is the clustering center of the j-th hash bucket, n j is the number of parameters in the j-th hash bucket, and θ k,j,m is the m-th parameter in the j-th hash bucket;

[0076] Thus, multiple clustering centers of each connected vehicle are obtained, and the number of clustering centers is the same as the number of hash buckets. For example, if 4 hash buckets are divided in Step 4.1, 4 clustering centers are obtained at this time.

[0077] Step 4.3: Calculate the residual between the parameter and the clustering center for each parameter in the same hash bucket. The residuals of the embedding vectors of all model parameters form a residual set, which is specifically implemented by the following formula:

[0078] Δ k,j,m = θ k,j,m - C j ;

[0079] Δcluster j = {Δ k,j,,m};

[0080] Among them, Δ k,j,m is the residual between the m-th parameter in the j-th hash bucket and the clustering center, and Δcluster j is the residual set of the j-th hash bucket;

[0081] Thus, multiple residual sets of each connected vehicle are obtained.

[0082] Step 5: Upload all the clustering centers of each connected vehicle to the server through the connected vehicle client;

[0083] After completing the LSH clustering, the client of each connected vehicle only uploads all the calculated clustering centers {C j} to the server, without the need to upload all the model parameters. This can significantly reduce the communication overhead because the number of clustering centers is much less than the number of original parameters.

[0084] Step 6: The server receives multiple clustering centers uploaded by each client and calculates the global clustering center C global ;

[0085] Step 6.1: The server receives multiple clustering centers uploaded by each client and determines the weight of each clustering center;

[0086] Among them, the weight of each clustering center can be determined according to the number of data points in the cluster or the importance of the client.

[0087] Step 6.2: Aggregate the clustering centers by the weighted average method according to the weights of the clustering centers to obtain the global clustering center C global , which is specifically implemented by the following formula:

[0088]

[0089] Among them, C k,j is the j-th clustering center of the k-th client, W k,j is the weight of C k,j , and B k is the total number of clustering centers uploaded by the k-th client.

[0090] Among them, in the present invention, calculating the global clustering center is to aggregate the clustering centers based on the weighted average method, and in the actual use process, it can be dynamically adjusted according to factors such as the model update frequency and data quality of each vehicle.

[0091] Step 7: The server sends the global clustering center C global to each connected vehicle;

[0092] Step 8: For each connected vehicle, the client of the connected vehicle receives the global clustering center C global , and updates the parameters according to the residuals of each parameter in the residual set, which is specifically implemented by the following formula:

[0093] θ' k,j,m = C global + Δ k,j,m ;

[0094] Among them, Δ k,j,m is the residual of the m-th parameter in the j-th hash bucket of the k-th connected vehicle in the residual set, and θ' k,j,mis the updated parameter of the m-th parameter in the j-th hash bucket of the k-th connected vehicle.

[0095] Furthermore, the embedding vectors of the updated model parameters of each connected vehicle are obtained;

[0096] In this way, the client can restore the complete parameters of its local model without retransmitting all parameters, thus further reducing the communication overhead.

[0097] Step 9: Increment the current communication count by one to obtain a new current communication count. Update the current local model according to the embedding vectors of the updated model parameters of each connected vehicle to obtain a new current local model. Determine whether the new current communication count is less than a preset threshold. If the new current communication count is greater than or equal to the preset threshold, use the new current local model as the final local model. If the new current communication count is less than the preset threshold, use the new current communication count as the current communication count and the new current local model as the current local model, and return to execute Step 3.

[0098] Among them, in order to reduce communication latency and bandwidth occupancy, the system can adopt communication optimization strategies such as batch upload and differential update.

[0099] For the process of uploading all cluster centers of each connected vehicle to the server through the connected vehicle client, and the process of the server sending the global cluster center C global to each connected vehicle, encryption technology can be used to encrypt the hash parameters and mappings to prevent man-in-the-middle attacks and data leakage, and further ensure the security of the system. For example, use AES (Advanced Encryption Standard) to perform symmetric encryption on the uploaded data, or use RSA for asymmetric encryption.

[0100] The above description is only the preferred embodiment of the present disclosure and the description of the applied technical principles. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present disclosure is not limited to the technical solution formed by the specific combination of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the technical solutions formed by mutually replacing the above features with the (but not limited to) technical features with similar functions disclosed in the embodiments of the present disclosure.

Claims

1. A privacy protection communication optimization method for connected vehicles based on local sensitive hashing, characterized in that: include: Step 1: Collect vehicle data of K connected vehicles, where the vehicle data includes the location information of the connected vehicles and the distances between all objects and the connected vehicles within a preset radius around the connected vehicles; pre-process the vehicle data to obtain a set of feature vectors corresponding to the vehicle data, i.e., local data; Step 2: Each connected vehicle downloads a neural network model from the server as an initial local model. The initial model parameters of the initial local models of all connected vehicles are the same. The initial local model is used as the current local model, and the initial communication times are set, and the initial communication times are used as the current communication times. Step 3: In the current number of communications, for each connected vehicle, at the client of the connected vehicle, the local data is input into the current local model to obtain the predicted position of the object within the preset radius with the connected vehicle as the center, and then the loss function is obtained based on the predicted position of the object within the preset radius with the connected vehicle as the center and the label representing the true position of the object. The model parameters are updated by the gradient descent method according to the loss function, and the local data is input into the model after the model parameters are updated, and then the model parameters are updated again. The above operation is repeated until the number of model parameter updates reaches the preset threshold, and the final model parameter embedding vector θ of each connected vehicle is obtained. k ; Step 4: According to the embedding vector of the final model parameters of each connected vehicle, all parameters in the embedding vector are distributed into multiple hash buckets. For each hash bucket, the cluster center is calculated to obtain multiple cluster centers of each connected vehicle. Then, the residual of each parameter in the hash bucket is calculated to obtain a residual set, and then the residual set of each connected vehicle is obtained. Step 5: Upload all cluster centers of each connected vehicle to the server through the connected vehicle client; Step 6: The server receives multiple cluster centers uploaded by each client and calculates the global cluster center C global ; Step 7: The server will global cluster center C global Sent to each connected vehicle; Step 8: For each connected vehicle, the client of the connected vehicle receives the global cluster center C global , according to the residual of each parameter in the residual set, the parameters are updated, and then the embedded vector of the updated model parameters of each connected vehicle is obtained; Step 9: Add one to the current communication number to obtain a new current communication number. Update the current local model according to the embedded vector of the updated model parameters of each connected vehicle to obtain a new current local model. Determine whether the new current communication number is less than the preset threshold. If the new current communication number is greater than or equal to the preset threshold, use the new current local model as the final local model. If the new current communication number is less than the preset threshold, use the new current communication number as the current communication number and the new current local model as the current local model, and return to step 3.

2. According to claim 1, a privacy protection communication optimization method for connected vehicles based on local sensitive hashing is characterized in that: In step 1, the vehicle data is preprocessed to obtain a set of feature vectors corresponding to the vehicle data, including: The vehicle data is filtered, denoised and feature vectors are extracted to obtain a feature vector set corresponding to the vehicle data.

3. According to claim 1, a privacy protection communication optimization method for connected vehicles based on local sensitive hashing is characterized in that: In step 3, the model parameters are updated by the gradient descent method according to the loss function, which is specifically implemented by the following formula: Among them, η represents the learning rate, represents the embedding vector of the model parameters of the t-th iteration of the k-th connected vehicle, represents the embedding vector of the model parameters of the kth connected car at the t+1th iteration, is the loss function L of the k-th connected vehicle at the t-th iteration k exist The gradient at .

4. According to claim 1, a privacy protection communication optimization method for connected vehicles based on local sensitive hashing is characterized in that: Step 4 specifically includes: Step 4.1: Embedding vector θ of the final model parameters for each connected car k , through L different locality sensitive hashing algorithms, the embedding vector θ k Each parameter θ in k,i Processing is performed to obtain L hash values, which are specifically implemented through the following formula: h l =argmax i∈{±1,±2,...,±d} |R l i k,i | i ,l=1,2,...,L; Among them, h l is the hash value calculated by the l-th local sensitive hashing algorithm, d represents the dimension, and argmax is used to select the hash value that makes |R L θ k,i | i The largest index function, R l is the random rotation matrix of the lth local sensitive hashing algorithm; The L hash values ​​are concatenated to obtain a hash signature, which is implemented using the following formula: signature=(h1,h2,...,h L ); Then we get the embedding vector θ k 2*d hash signatures are assigned to the same hash bucket for the parameters with the same hash signature in the embedded vector; Step 4.2: For each parameter in the hash bucket, calculate the cluster center of the hash bucket, which is implemented by the following formula: Among them, C j is the cluster center of the jth hash bucket, n j is the number of parameters in the jth hash bucket, θ k,j,m is the mth parameter in the jth hash bucket; Thus, multiple cluster centers of each connected vehicle are obtained; Step 4.3: For each parameter of the same hash bucket, calculate the residual between the parameter and the cluster center. The residuals of the embedding vectors of all model parameters constitute the residual set, which is implemented by the following formula: D k,j,m =θ k,j,m -C j ; Δcluster j ={Δ k,j,,m }; Among them, Δ k,j,m is the residual between the mth parameter of the jth hash bucket and the cluster center, Δcluster j is the residual set of the jth hash bucket; Thus, multiple residual sets of each connected vehicle are obtained.

5. According to claim 1, a privacy protection communication optimization method for connected vehicles based on local sensitive hashing is characterized in that: Step 6 specifically includes: Step 6.1: The server receives multiple cluster centers uploaded by each client and determines the weight of each cluster center; Step 6.2: According to the weight of the cluster center, aggregate the cluster centers by weighted average method to obtain the global cluster center C global .

6. The privacy protection communication optimization method for connected vehicles based on local sensitive hashing according to claim 5 is characterized in that: Step 6.2 is specifically implemented by the following formula: Among them, C k,j is the jth cluster center of the kth client, W k,j C k,j The weight of B k is the total number of cluster centers uploaded by the kth client.

7. The privacy protection communication optimization method for connected vehicles based on local sensitive hashing according to claim 1 is characterized in that: In step 8, the parameters are updated according to the residual of each parameter in the residual set, which is specifically implemented by the following formula: θ′ k,j,m =C global +D k,j,m ; Among them, Δ k,j,m is the residual of the mth parameter in the jth hash bucket of the kth connected vehicle in the residual set, θ' k,j,m The updated parameter of the mth parameter in the jth hash bucket of the kth connected vehicle.

Images