Vehicle data prediction method based on GRU and differential privacy federated learning

Abstract

The invention belongs to the technical field of vehicle dynamics prediction, and discloses a vehicle data prediction method based on GRU and differential privacy federated learning, and the method comprises the following steps: a server side randomly initializes global GRU model parameters, and then carries out the training of a local GRU model; the client processes the trained local GRU model parameters through a differential privacy mechanism to generate a disturbance model; each client uploads the disturbance model parameters to a server side; carrying out global GRU model aggregation by adopting a weighted aggregation algorithm in federated learning; and the server side redistributes the aggregated global GRU model to each client side, and the global GRU model for accurately predicting the vehicle dynamics is obtained through multiple rounds of training and updating. According to the vehicle data prediction method based on GRU and differential privacy federated learning, through combination of a GRU network and a differential privacy mechanism, vehicle dynamics prediction in an Internet of Vehicles environment is realized, and privacy of a user is effectively protected.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle dynamics prediction, and in particular to a vehicle data prediction method based on GRU and differential privacy federated learning. Background Art

[0002] With the rapid development of intelligent transportation and Internet of Vehicles technologies, vehicle dynamics prediction has become an important technology to ensure traffic safety and optimize road resource management. In recent years, many researchers have proposed prediction methods based on deep learning and data-driven methods, but these methods still face challenges in terms of privacy protection and model performance.

[0003] A recurrent neural network (RNN) model based on real-time measurements is used for vehicle dynamics prediction. By combining open-loop and closed-loop training modes, the accuracy of long-term predictions is improved. A two-level data-driven vehicle speed prediction model first predicts the traffic speed of a road section through a neural network, and then uses hidden Markov models (HMMs) to capture the relationship between individual vehicle speed and traffic speed, taking into account the driver's random behavior. A deep belief network (DBN) model is used to predict vehicle speed and steering angle by analyzing real driving data and the state of surrounding vehicles. Using an adaptive neuro-fuzzy inference system, factors such as driver behavior, route, traffic, and weather are integrated into a big data-driven deep learning model to successfully predict vehicle speed in urban and highway scenarios, and the effect is better than traditional methods. In addition, the vehicle speed prediction model based on Markov chains, combined with road geometry and speed limits, effectively improves the prediction accuracy in real-time driving scenarios. Based on the two-stage adaptive Kalman filter algorithm, the acceleration, speed and position of the vehicle are predicted by integrating the autoregressive moving average model, and the traffic safety at the intersection is improved by quantifying the direction and lane information of the vehicle.

[0004] Although existing methods have made significant progress in improving the accuracy of vehicle dynamics prediction, most of these models rely on centralized data training and lack protection for user privacy. In the Internet of Vehicles environment, the local data of each vehicle often involves private information, and direct sharing poses a great security risk. Summary of the invention

[0005] The purpose of the present invention is to provide a vehicle data prediction method based on GRU and differential privacy federated learning, which is used for vehicle dynamics prediction in the Internet of Vehicles, while ensuring the accuracy of the model and protecting the privacy security of the vehicle; in order to protect user privacy and improve the accuracy of the prediction, the federated learning and differential privacy technologies are combined, and the local GRU model training is performed on each client using the GRU network to avoid direct data transmission; at the same time, the model parameters are denoised through the differential privacy mechanism to effectively protect the privacy of the user.

[0006] To achieve the above objectives, the present invention provides a vehicle data prediction method based on GRU and differential privacy federated learning, comprising the following steps:

[0007] Step S1: In the initial stage of training, the server randomly initializes the global gated recurrent unit (GRU) model parameters;

[0008] Step S2: Each client receives the initialized global GRU model and trains the local GRU model using the local vehicle dynamics data;

[0009] Step S3: After the local GRU model training is completed, in order to protect user privacy, the client processes the updated local GRU model parameters through the differential privacy mechanism to generate a perturbation model;

[0010] Step S4: After adding noise, each client uploads the processed disturbance model parameters to the server;

[0011] Step S5: Using the weighted aggregation algorithm in federated learning, the server integrates the uploaded perturbation model parameters and performs global GRU model aggregation;

[0012] Step S6: After the global GRU model aggregation is completed, the server redistributes the updated global GRU model to each client, and obtains a global GRU model that accurately predicts vehicle dynamics through multiple rounds of training and updating.

[0013] Preferably, in step S1, at the initial stage of training, the server randomly initializes the parameters of the global gated recurrent unit GRU model; the initialized global GRU model is distributed to each client, and each client uses the global GRU model as the initial state of its local GRU model to ensure that all clients have consistent models at the beginning of training.

[0014] Preferably, in step S2, after each client receives the initial global GRU model, it uses local vehicle dynamics data to train the local GRU model, and the client updates the weight parameters of the local GRU model according to its own driving record, speed, and acceleration sequence data.

[0015] Preferably, a GRU network is used for vehicle dynamics prediction, wherein the GRU network comprises a GRU layer for capturing the input sequence to generate a hidden state; a fully connected layer for mapping the hidden state to an output;

[0016] The relevant driving information is captured locally in the vehicle, and a training sample sequence database is generated by sliding step by step according to the time series, and the generated training sample sequence database is used as input;

[0017] The input data and the hidden state of the previous time step are passed to the reset gate, where the reset gate is calculated as follows:

[0018] r t =σ(W r ·[h t-1 , x t ]+b r );

[0019] Among them, r t is the value of the reset gate, h t-1 is the hidden state of the previous time step, x t is the input of the current time step, σ is the sigmoid activation function, W r is the weight matrix of the reset gate, which contains the weights related to the input and the hidden state of the previous time step, b r is the bias vector of the reset gate, which is used to perform bias correction on the linear transformation result;

[0020] The input data and the hidden state of the previous time step are simultaneously passed into the update gate, where the update gate is calculated as follows:

[0021] z t =σ(W z ·[h t-1 , x t ]+b z );

[0022] Among them, z t is the value of the update gate, W z is the weight matrix of the update gate, which weights the linear combination of the previous hidden state and the current input, b z It is the bias matrix of the update gate, which is used to make bias correction to the result of the linear combination;

[0023] Under the action of the reset gate, the candidate hidden state is calculated by combining the input data and the adjusted previous hidden state, where the candidate hidden state is calculated as follows:

[0024]

[0025] in, is a candidate hidden state, which is a new hidden state candidate value generated at the current time step t, W h is the weight matrix of the candidate hidden state, which performs weighted calculation on the previous hidden state and the current input, b h is the bias vector of the candidate hidden state, which is used to adjust the result of the weighted sum calculation. Tanh represents the hyperbolic tangent activation function, which is used to compress the value of the candidate hidden state into the range of (-1, 1);

[0026] Based on the output of the update gate, the updated hidden state is passed to the next time step, where the hidden state is updated as follows:

[0027]

[0028] Among them, h t is the updated hidden state of the current time step, which is the hidden state that is finally passed to the next time step;

[0029] After the GRU network processes all time steps, it outputs the hidden state of the last time step, which is the final output after sequence processing; this hidden state is then passed to the fully connected layer to map it to the final output value.

[0030] Preferably, in step S3, after the local GRU model training is completed, in order to further protect user privacy, the client processes the updated local GRU model parameters through a differential privacy mechanism;

[0031] Differential privacy noise adding methods include: Laplace noise or Gaussian noise. The specific process is as follows:

[0032] (1) The description of Gaussian noise is as follows:

[0033] Add Gaussian noise to the local GRU model of the client, that is, the Gaussian noise added to the local GRU model satisfies the standard normal distribution, where the perturbation model is as follows:

[0034] P=g+N(0,σ 2 );

[0035] Among them, P is the perturbation model after adding Gaussian noise, g is the local GRU model trained locally on the client, N(0,σ 2 ) is Gaussian noise, σ satisfies:

[0036]

[0037] Among them, Δf is the sensitivity of the function, which indicates the maximum impact of the change of a single sample in the data set on the function output, and Δf is related to the difference between the training sets; ε is the privacy budget to measure the level of noise used; δ is the privacy failure probability, which indicates the probability that differential privacy protection may fail;

[0038] (2) Laplace noise is described as follows:

[0039] Add Laplace noise to the local GRU model of the client, that is, the Laplace noise added to the local GRU model satisfies the standard normal distribution, where the perturbation model is as follows:

[0040] P' = g + Noise;

[0041] Among them, P' is the disturbance model after adding Gaussian noise, and Noise satisfies the Laplace distribution, as shown below:

[0042]

[0043] Among them, ∈ is the privacy budget in differential privacy, which controls the intensity of adding noise; Laplace represents the Laplace distribution, as shown below:

[0044]

[0045] Where f is the probability density function of the Laplace distribution, which represents the probability distribution of the random variable x given the mean μ and scale parameter b, x is the random variable, μ is the mean, and b is the variance.

[0046] Preferably, in step S4, after the noise is added, each client uploads the processed disturbance model parameters to the server; wherein the model uploaded by the client only contains the protected parameter information, and the original data is still retained locally.

[0047] Preferably, in step S5, after receiving the perturbation models from each client, the server adopts a weighted aggregation algorithm in federated learning to integrate the perturbation model parameters uploaded by all clients to generate an updated global GRU model;

[0048] During aggregation, the server assigns different weights based on the amount of local data of each client and the quality factor of the global GRU model update to ensure that the global GRU model can fully absorb the feature information of each client.

[0049] Preferably, in step S6, after the global GRU model aggregation is completed, the server redistributes the updated global GRU model to each client, and the client uses the global GRU model as the starting point for the next round of training;

[0050] This cycle is repeated continuously, and through multiple rounds of training and updating, a global GRU model that accurately predicts vehicle dynamics is obtained.

[0051] Therefore, the present invention adopts the above-mentioned vehicle data prediction method based on GRU and differential privacy federated learning, and the beneficial effects are as follows:

[0052] (1) The present invention realizes vehicle dynamics prediction in the Internet of Vehicles environment and effectively protects the privacy of users by combining the GRU network and the differential privacy mechanism.

[0053] (2) The present invention uses the GRU network to process the sequence data in vehicle dynamics, which can fully capture the time series characteristics such as vehicle speed and acceleration, thereby improving the accuracy of prediction.

[0054] (3) The present invention introduces a differential privacy mechanism to add an appropriate amount of noise without leaking the user's original data, thereby ensuring privacy security during data uploading and model aggregation.

[0055] (4) Compared with traditional centralized data processing methods, the present invention not only reduces the risk of data leakage, but also realizes distributed and efficient training through a federated learning framework, thereby significantly improving the scalability and security of the system.

[0056] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is an overall flow chart of a vehicle data prediction method based on GRU and differential privacy federated learning of the present invention;

[0058] Figure 2 It is a schematic diagram of the GRU structure of the present invention;

[0059] Figure 3 The vehicle speed prediction result of the present invention;

[0060] Figure 4 The vehicle heading prediction result of the present invention. DETAILED DESCRIPTION

[0061] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0062] like Figure 1 As shown, the present invention provides a vehicle data prediction method based on GRU and differential privacy federated learning, comprising the following steps:

[0063] Step S1: Initialize the global GRU model.

[0064] In the initial stage of training, the server randomly initializes the parameters of the global GRU (Gated Recurrent Unit) model. The GRU network is selected for the vehicle dynamics prediction task because of its high efficiency and low computational overhead when processing sequence data.

[0065] The initialized global GRU model is then distributed to each client, and each client uses the global GRU model as the initial state of its local GRU model to ensure that all clients have consistent models at the beginning of training.

[0066] Step S2: local GRU model training.

[0067] After receiving the initialized global GRU model, each client uses the local vehicle dynamics data to train the local GRU model. During this process, the client updates the weight parameters of the local GRU model based on its own driving record, speed, acceleration and other sequence data. Since each client's local data is collected independently and not shared with other clients, different vehicle dynamics characteristics will be accumulated during the model training process of each client. This distributed training method ensures the protection of data privacy.

[0068] In order to predict vehicle dynamics, such as Figure 2 As shown, a GRU network is used, which includes a GRU layer for capturing the input sequence to generate a hidden state; a fully connected layer maps the hidden state to the output.

[0069] Among them, relevant driving information is captured locally in the vehicle, and a training sample sequence database is generated by sliding step by step according to the time series, and the generated training sample sequence database is used as input.

[0070] The input data and the hidden state of the previous time step are passed to the reset gate together. The function of the reset gate is to control how much information in the hidden state of the previous time step is retained. If the reset gate output is close to 0, it means that most of the historical information is forgotten; if it is close to 1, more historical information is retained. The reset gate is calculated as follows:

[0071] r t =σ(W r ·[h t-1 , x t ]+b r );

[0072] Among them, r t is the value of the reset gate, h t-1 is the hidden state of the previous time step, x t is the input of the current time step, σ is the sigmoid activation function, W ris the weight matrix of the reset gate, which contains the weights related to the input and the hidden state of the previous time step, b r is the bias vector of the reset gate, which is used to perform bias correction on the linear transformation result.

[0073] The input data and the hidden state of the previous time step are passed to the update gate at the same time. The update gate determines how much the hidden state needs to be updated, that is, it makes a trade-off between the new input data and the past memory. The update gate controls whether to update the hidden state to a new candidate state or keep the original memory. The calculation method of the update gate is as follows:

[0074] z t =σ(W z ·[h t-1 , x t ]+b z );

[0075] Among them, z t is the value of the update gate, W z is the weight matrix of the update gate, which weights the linear combination of the previous hidden state and the current input, b z It is the bias matrix of the update gate, which is used to perform bias correction on the result of the linear combination.

[0076] Under the action of the reset gate, the candidate hidden state is calculated by combining the input data and the adjusted previous hidden state. The candidate hidden state contains important information about the current time step for the update gate to decide whether to use it to update the hidden state. The candidate hidden state is calculated as follows:

[0077]

[0078] in, is a candidate hidden state, which is a new hidden state candidate value generated at the current time step t, W h is the weight matrix of the candidate hidden state, which performs weighted calculation on the previous hidden state and the current input, b h is the bias vector of the candidate hidden state, which is used to adjust the result of the weighted sum calculation. Tanh represents the hyperbolic tangent activation function, which is used to compress the value of the candidate hidden state into the range of (-1, 1).

[0079] According to the output of the update gate, the hidden state of the current time step will be updated to a new hidden state, or partially keep the state of the previous time step. The updated hidden state will be passed to the next time step. Among them, the hidden state is updated as follows:

[0080]

[0081] Among them, ht is the updated hidden state for the current time step, which is the hidden state that is ultimately passed to the next time step.

[0082] After the GRU network processes all time steps, it outputs the hidden state of the last time step. This is the final output after sequence processing. This hidden state is then passed to the fully connected layer (Linear Layer) to map it to the final output value.

[0083] Step S3: Differential privacy noise addition.

[0084] After the local GRU model training is completed, in order to further protect user privacy, the client processes the updated local GRU model parameters through the differential privacy mechanism. Differential privacy adds noise to the local GRU model parameters to ensure that the original data cannot be directly inferred from the parameters during the upload process. The intensity of the noise is controlled by a pre-set privacy budget. Of course, the client can choose different noise intensities according to the required privacy level to balance privacy protection and model accuracy.

[0085] Differential privacy noise adding methods include: Laplace noise or Gaussian noise. The specific process is as follows:

[0086] (1) The description of Gaussian noise is as follows:

[0087] Add Gaussian noise to the local GRU model of the client, that is, the Gaussian noise added to the local GRU model satisfies the standard normal distribution, where the perturbation model is as follows:

[0088] P=g+N(0,σ 2 );

[0089] Among them, P is the perturbation model after adding Gaussian noise, g is the local GRU model trained locally on the client, N(0,σ 2 ) is Gaussian noise, σ satisfies:

[0090]

[0091] Among them, Δf is the sensitivity of the function, which indicates the maximum impact of the change of a single sample in the data set on the function output, and Δf is related to the difference between the training sets; ε is the privacy budget that measures the level of noise used; δ is the privacy failure probability, which indicates the probability that differential privacy protection may fail; in the present invention, it is customized for the DQN network.

[0092] (2) Laplace noise is described as follows:

[0093] Add Laplace noise to the local GRU model of the client, that is, the Laplace noise added to the local GRU model satisfies the standard normal distribution, where the perturbation model is as follows:

[0094] P' = g + Noise;

[0095] Among them, P' is the disturbance model after adding Gaussian noise, and Noise satisfies the Laplace distribution, as shown below:

[0096]

[0097] Among them, ∈ is the privacy budget in differential privacy, which controls the intensity of adding noise; Laplace represents the Laplace distribution, as shown below:

[0098]

[0099] Where f is the probability density function of the Laplace distribution, which represents the probability distribution of the random variable x given the mean μ and scale parameter b, x is the random variable, μ is the mean, and b is the variance.

[0100] Step S4: After adding noise, each client uploads the processed disturbance model parameters to the server. In this process, the model uploaded by the client only contains the protected parameter information, and the original data is still retained locally.

[0101] Step S5: Global GRU model aggregation.

[0102] After receiving the perturbation models from each client, the server uses the weighted aggregation algorithm in federated learning to integrate the perturbation model parameters uploaded by all clients to generate an updated global GRU model. During aggregation, the server will assign different weights based on factors such as the amount of local data of each client and the quality of the global GRU model update to ensure that the global GRU model can fully absorb the feature information of each client. The aggregated global GRU model can better adapt to the global characteristics of vehicle dynamics and has stronger generalization capabilities.

[0103] Step S6: Global GRU model distribution.

[0104] After the global GRU model is aggregated, the server redistributes the updated global GRU model to each client, and the client uses the global GRU model as the starting point for the next round of training. This cycle is repeated continuously, and through multiple rounds of training and updating, a high-precision global GRU model that accurately predicts vehicle dynamics is finally obtained.

[0105] Example

[0106] Based on the vehicle datasets of Helsinki, the Netherlands and Shanghai traffic datasets, experiments were conducted using multi-dimensional information such as vehicle location information, historical speed information, historical heading information, vehicle type, and vehicle driving route information for joint prediction. Two experiments were conducted on vehicle speed prediction and vehicle heading prediction, and comparative experiments were conducted under different numbers of participating vehicles. The experimental results are as follows:

[0107] In this embodiment, a total of 20 rounds of communication are performed, and a different number of vehicles are selected for training in each round. Group-10 means that 10 vehicles are selected for training in each round; Group-20 means that 20 vehicles are selected for training in each round; and Group-30 means that 30 vehicles are selected for training in each round. In the vehicle speed prediction task, the vehicle speed is used as the prediction target; in the vehicle heading prediction task, the vehicle heading angle is used as the prediction target.

[0108] like Figure 3 As shown in the figure, the models obtained in the vehicle speed prediction task all show an obvious convergence trend. In addition, good convergence effects are obtained in different groups, and the MAE error values ​​are all at a relatively small position, indicating that the method proposed in the present invention is feasible and applicable. With the increase in the number of participating vehicles and the increase in the number of training rounds, the present invention also shows better performance in prediction accuracy.

[0109] like Figure 4 As shown in the figure, the models obtained in the vehicle heading prediction task all show an obvious convergence trend. In addition, good convergence effects are obtained in different groups, and the MAE error values ​​are all at a relatively small position, indicating that the method proposed in the present invention is feasible and applicable. With the increase of participating vehicles and the increase of training rounds, the present invention also shows better performance in prediction accuracy.

[0110] Therefore, the present invention adopts the above-mentioned vehicle data prediction method based on GRU and differential privacy federated learning for vehicle dynamics prediction in the Internet of Vehicles, which protects the privacy and security of the vehicle while ensuring the accuracy of the model; in order to protect user privacy and improve the accuracy of the prediction, the present invention combines federated learning and differential privacy technology, and uses the GRU network on each client to perform local GRU model training to avoid direct data transmission; at the same time, the model parameters are denoised through the differential privacy mechanism to effectively protect the privacy of the user.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A vehicle data prediction method based on GRU and differential privacy federated learning, characterized in that: The following steps are involved: Step S1: In the initial stage of training, the server randomly initializes the global gated recurrent unit (GRU) model parameters; Step S2: Each client receives the initialized global GRU model and trains the local GRU model using the local vehicle dynamics data; Step S3: After the local GRU model training is completed, in order to protect user privacy, the client processes the updated local GRU model parameters through the differential privacy mechanism to generate a perturbation model; Step S4: After adding noise, each client uploads the processed disturbance model parameters to the server; Step S5: Using the weighted aggregation algorithm in federated learning, the server integrates the uploaded perturbation model parameters and performs global GRU model aggregation; Step S6: After the global GRU model aggregation is completed, the server redistributes the updated global GRU model to each client, and obtains a global GRU model that accurately predicts vehicle dynamics through multiple rounds of training and updating.

2. The vehicle data prediction method based on GRU and differential privacy federated learning according to claim 1, characterized in that: In step S1, at the initial stage of training, the server randomly initializes the parameters of the global gated recurrent unit GRU model; the initialized global GRU model is distributed to each client, and each client uses the global GRU model as the initial state of its local GRU model to ensure that all clients have consistent models at the beginning of training.

3. The vehicle data prediction method based on GRU and differential privacy federated learning according to claim 2, characterized in that: In step S2, each client receives the initialized global GRU model and trains the local GRU model using the local vehicle dynamics data. The client updates the weight parameters of the local GRU model based on its own driving record, speed, and acceleration sequence data.

4. The vehicle data prediction method based on GRU and differential privacy federated learning according to claim 3 is characterized by: The GRU network is used for vehicle dynamics prediction. The GRU network consists of a GRU layer that captures the input sequence to generate hidden states; a fully connected layer maps the hidden states to outputs; The relevant driving information is captured locally in the vehicle, and a training sample sequence database is generated by sliding step by step according to the time series, and the generated training sample sequence database is used as input; The input data and the hidden state of the previous time step are passed to the reset gate, where the reset gate is calculated as follows: r t =σ(W r ·[h t-1 ,x t ]+b r ); Among them, r t is the value of the reset gate, h t-1 is the hidden state of the previous time step, x t is the input of the current time step, σ is the sigmoid activation function, W r is the weight matrix of the reset gate, which contains the weights related to the input and the hidden state of the previous time step, b r is the bias vector of the reset gate, which is used to perform bias correction on the linear transformation result; The input data and the hidden state of the previous time step are simultaneously passed into the update gate, where the update gate is calculated as follows: z t =σ(W z ·[h t-1 ,x t ]+b z ); Among them, z t is the value of the update gate, W z is the weight matrix of the update gate, which weights the linear combination of the previous hidden state and the current input, b z It is the bias matrix of the update gate, which is used to make bias correction to the result of the linear combination; Under the action of the reset gate, the candidate hidden state is calculated by combining the input data and the adjusted previous hidden state, where the candidate hidden state is calculated as follows: in, is a candidate hidden state, which is a new hidden state candidate value generated at the current time step t, W h is the weight matrix of the candidate hidden state, which performs weighted calculation on the previous hidden state and the current input, b h is the bias vector of the candidate hidden state, which is used to adjust the result of the weighted sum calculation. Tanh represents the hyperbolic tangent activation function, which is used to compress the value of the candidate hidden state into the range of (-1,1); Based on the output of the update gate, the updated hidden state is passed to the next time step, where the hidden state is updated as follows: Among them, h t is the updated hidden state of the current time step, which is the hidden state that is finally passed to the next time step; After the GRU network processes all time steps, it outputs the hidden state of the last time step, which is the final output after sequence processing; this hidden state is then passed to the fully connected layer to map it to the final output value.

5. The vehicle data prediction method based on GRU and differential privacy federated learning according to claim 4, characterized in that: In step S3, after the local GRU model training is completed, in order to further protect user privacy, the client processes the updated local GRU model parameters through the differential privacy mechanism; Differential privacy noise can be added in two ways: Laplace noise or Gaussian noise. The specific process is as follows: (1) The description of Gaussian noise is as follows: Add Gaussian noise to the local GRU model of the client, that is, the Gaussian noise added to the local GRU model satisfies the standard normal distribution, where the perturbation model is as follows: P=g+N(0,σ 2 ); Among them, P is the perturbation model after adding Gaussian noise, g is the local GRU model trained locally on the client, N(0,σ 2 ) is Gaussian noise, σ satisfies: Among them, Δf is the sensitivity of the function, which indicates the maximum impact of the change of a single sample in the data set on the function output, and Δf is related to the difference between the training sets; ε is the privacy budget to measure the level of noise used; δ is the privacy failure probability, which indicates the probability that differential privacy protection may fail; (2) Laplace noise is described as follows: Add Laplace noise to the local GRU model of the client, that is, the Laplace noise added to the local GRU model satisfies the standard normal distribution, where the perturbation model is as follows: P' = g + Noise; Among them, P' is the disturbance model after adding Gaussian noise, and Noise satisfies the Laplace distribution, as shown below: Among them, ∈ is the privacy budget in differential privacy, which controls the intensity of adding noise; Laplace represents the Laplace distribution, as shown below: Where f is the probability density function of the Laplace distribution, which represents the probability distribution of the random variable x given the mean μ and scale parameter b. x is a random variable representing the value of the noise, μ is the mean, and b is the variance.

6. The vehicle data prediction method based on GRU and differential privacy federated learning according to claim 5, characterized in that: In step S4, after adding noise, each client uploads the processed disturbance model parameters to the server; wherein, the model uploaded by the client only contains the protected parameter information, and the original data is still retained locally.

7. The vehicle data prediction method based on GRU and differential privacy federated learning according to claim 6, characterized in that: In step S5, after receiving the perturbation models from each client, the server uses the weighted aggregation algorithm in federated learning to integrate the perturbation model parameters uploaded by all clients to generate an updated global GRU model; During aggregation, the server assigns different weights based on the amount of local data of each client and the quality factor of the global GRU model update to ensure that the global GRU model can fully absorb the feature information of each client.

8. The vehicle data prediction method based on GRU and differential privacy federated learning according to claim 7, characterized in that: In step S6, after the global GRU model is aggregated, the server redistributes the updated global GRU model to each client, and the client uses the global GRU model as the starting point for the next round of training; This cycle is repeated continuously, and through multiple rounds of training and updating, a global GRU model that accurately predicts vehicle dynamics is obtained.

Images