Block chain-based distributed federated learning Internet of Vehicles knowledge sharing method

Abstract

The invention provides a block chain-based distributed federated learning Internet of Vehicles knowledge sharing method, and relates to the field of block chains and Internet of Vehicles. Comprising the following steps: firstly, adopting a neural factor decomposition machine (NFM) as a prediction network, and selecting a vehicle node most suitable for participating in learning in combination with a dynamic adaptive node selection algorithm; secondly, the vehicle node transmits the learning result as transaction data to a roadside unit, and the roadside unit generates candidate blocks and ensures transparency and non-tampering of the data through a consensus mechanism based on knowledge proof; thirdly, global model updating is carried out in the edge cloud service by utilizing a knowledge distillation technology, and the model performance is optimized; and finally, constructing a non-cooperative game excitation model of vehicle nodes and roadside units, performing reasonable excitation distribution by adopting a deep Q network (DQN), and encouraging the nodes to actively participate in learning and contribute. Through the block chain technology and the distributed federated learning mechanism, the efficiency and security of knowledge sharing in the Internet of Vehicles environment are significantly improved.

Description

Technical Field

[0001] The present invention belongs to the fields of blockchain and Internet of Vehicles, and specifically relates to a blockchain-based distributed federated learning method for knowledge sharing in the Internet of Vehicles, aiming to solve the problems of data privacy leakage and single-point failure in the Internet of Vehicles environment, and improve the training efficiency and security of machine learning models. Background Art

[0002] With the development of Internet of Vehicles technology, data sharing and collaborative learning among vehicles have become important means to improve the performance of intelligent transportation systems. Vehicles can improve driving safety, road traffic efficiency, and driving experience by transmitting and sharing information with each other. However, traditional centralized learning methods have problems of data privacy leakage and single-point failure. Since the centralized server stores a large amount of sensitive vehicle data, once the server is attacked or fails, the security and stability of the entire system will be seriously threatened. Therefore, it is difficult to meet the requirements of efficient and secure knowledge sharing in the Internet of Vehicles environment.

[0003] A knowledge sharing method for the Internet of Vehicles based on blockchain-based distributed federated learning has been proposed to solve these problems. Blockchain technology ensures the security and transparency of data through its decentralized and immutable characteristics. Federated learning allows vehicle nodes to perform model training locally and only upload model parameters instead of raw data, thereby protecting data privacy and enabling efficient and secure knowledge sharing in the Internet of Vehicles environment, improving the overall system performance. Summary of the Invention

[0004] Object of the Invention: The object of the present invention is to construct an efficient and secure knowledge sharing system for the Internet of Vehicles by introducing blockchain technology and distributed federated learning. Through distributed federated learning, vehicle nodes perform model training locally and only upload model parameters instead of raw data, thereby protecting the privacy of vehicles and users. Utilizing the decentralized and immutable characteristics of blockchain technology, ensure the security and transparency of the data transaction and model update process, and prevent data tampering and malicious attacks. The non-cooperative game incentive mechanism maximizes the interests of each participating node by optimizing the utility function, and encourages vehicle nodes and roadside units to actively participate in data sharing and model training, thereby improving the overall performance and stability of the system. By constructing a hierarchical blockchain framework and a lightweight consensus mechanism, optimize the data processing and model aggregation process, and further improve the overall performance and stability of the intelligent transportation system in the Internet of Vehicles environment.

[0005] Technical Solution: A blockchain-based distributed federated learning method for knowledge sharing in the Internet of Vehicles: includes the following steps:

[0006] S1: Adopt a Neural Factorization Machine (NFM) as the prediction network, and combine it with a dynamic adaptive node selection algorithm to select the most suitable vehicle nodes to participate in learning;

[0007] S2: Based on the selected vehicle nodes in S1, the vehicle nodes perform NFM model training locally and upload the training results as transaction data to the roadside unit;

[0008] S3: Based on the transaction data uploaded in S2, the roadside unit generates candidate blocks and ensures the transparency and immutability of the data through a knowledge - based proof consensus mechanism;

[0009] S4: Based on the candidate blocks generated in S3, the edge cloud service uses knowledge distillation technology to perform global model updates and optimize the model performance;

[0010] S5: Based on the globally optimized model in S4, construct a non - cooperative game incentive model for vehicle nodes and roadside units, and use the deep Q - network (DQN) for reasonable incentive distribution to encourage nodes to actively participate in learning and contribution.

[0011] Preferably, in S1, the neural factorization machine (NFM) is used as the prediction network, combined with a dynamic adaptive node selection algorithm to select the most suitable vehicle nodes for participating in learning, specifically including:

[0012] The vehicle nodes collect their own feature data, including vehicle features, road features, environmental features, user features, and time features, and classify these feature vectors to distinguish continuous variables and embedding vectors;

[0013] Perform standardization and normalization processing on the collected data to ensure the consistency and usability of the data;

[0014] Design a dynamic adaptive node selection algorithm, aiming at minimizing a partial global loss function, to select the most suitable vehicle nodes for participating in learning and provide optimized node selection for subsequent steps.

[0015] Preferably, in S2, the vehicle nodes transfer the learning results as transaction data to the roadside unit and upload the consensus - reached blocks to the edge cloud through a knowledge - based proof consensus mechanism, specifically including:

[0016] Based on the selected vehicle nodes in S1, the vehicle nodes perform NFM model training locally and upload the training results to the roadside unit;

[0017] The roadside unit collects and verifies the local model parameters uploaded by the vehicle nodes, generates candidate blocks, and provides a data basis for the subsequent consensus mechanism;

[0018] Through a knowledge - based proof consensus mechanism, select the transactions with the highest accuracy for block generation to ensure the transparency and immutability of the data and provide reliable data support for subsequent model updates.

[0019] Upload the successfully consensus block to the edge cloud service to provide reliable data support for subsequent global model updates.

[0020] Preferably, in S4, the edge cloud service uses knowledge distillation technology for global model updates, specifically including:

[0021] Based on the candidate blocks generated in S3, the edge cloud service performs knowledge distillation;

[0022] Through knowledge distillation technology, transfer the knowledge of the optimized teacher model to the roadside unit to ensure the efficiency and consistency of the model;

[0023] The edge cloud service further aggregates and optimizes the model parameters to generate the final global model, providing an optimized model basis for subsequent incentive distribution.

[0024] Preferably, in S5, constructing a non - cooperative game incentive model for vehicle nodes and roadside units specifically includes:

[0025] Based on the optimized global model in S4, design an incentive function for vehicle nodes, and calculate their incentives based on the communication cost, computing cost, and data quality index of vehicle nodes;

[0026] Design an incentive function for roadside units, and calculate their incentives based on the contribution degree and participation frequency of vehicle nodes;

[0027] Use a deep Q - network (DQN) to solve the Stackelberg game model, optimize the incentive distribution strategy, encourage vehicle nodes and roadside units to actively participate in learning and contribution, and ensure the continuous and efficient operation of the system.

[0028] Preferably, the dynamic adaptive node selection algorithm specifically includes:

[0029] According to the computing cost, communication cost, current location, current speed, direction angle, and data quality index of vehicle nodes, select the most suitable vehicle nodes to participate in learning;

[0030] Solve the problem of minimizing a partial global loss function through gradient descent combined with constraint optimization, ensuring that the selected nodes meet the communication range and time resource limit constraints, and providing an optimized node selection for subsequent steps.

[0031] Preferably, the knowledge - based proof consensus mechanism specifically includes:

[0032] The roadside unit collects and verifies the learning transactions uploaded by vehicle nodes to ensure their effectiveness and accuracy;

[0033] The roadside unit packs the verified transactions into candidate blocks and selects the transaction with the highest accuracy according to the learning results for block generation;

[0034] By calculating the PoK scores of each candidate block, the optimal candidate block is selected for broadcasting to ensure data transparency and immutability, providing reliable data support for subsequent model updates.

[0035] Preferably, the knowledge distillation technology specifically includes:

[0036] By using the cross-entropy loss function and KL divergence, the knowledge of complex models is transferred to simpler models to achieve model compression and knowledge sharing;

[0037] Through the knowledge distillation technology, the model parameters are optimized to improve the learning efficiency of the system and the generalization ability of the model, providing an optimized model basis for subsequent incentive allocation.

[0038] Preferably, the solution of the Stackelberg game model by the deep Q-network (DQN) specifically includes:

[0039] Define the state space and action space, including the specific actions of the roadside unit adjustment incentive strategy and the vehicle node adjustment strategy and incentives;

[0040] Estimate the expected total reward obtained after taking action α in state s through the Q-network model to optimize the incentive allocation strategy;

[0041] Through continuous iteration and training, the agent gradually learns the optimal incentive strategy and the strategy of vehicle nodes, maximizes the cumulative reward and satisfies the constraint conditions to ensure the continuous and efficient operation of the system.

[0042] Beneficial effects: The method for sharing vehicle networking knowledge based on blockchain-based distributed federated learning of the present invention, firstly, by combining blockchain technology and distributed federated learning mechanisms, effectively solves the problems of data privacy leakage and single-point failure existing in traditional centralized learning methods, and significantly improves the training efficiency and security of machine learning models in the vehicle networking environment. Secondly, by using the neural factorization machine (NFM) as the prediction network, it can accurately estimate the task objectives of each vehicle node, and design a dynamic adaptive node selection method to ensure the selection of the most suitable vehicle nodes for participating in learning, optimizing the resource allocation and utilization efficiency. In addition, by using the knowledge distillation technology, the learning results are transmitted as transactions to the roadside units, combined with the proof-of-knowledge-based consensus mechanism, ensuring data transparency and immutability, and enhancing the reliability and trust of the system. Finally, by constructing the utility function of the vehicle nodes participating in learning, reasonable incentive allocation is achieved, encouraging vehicle nodes to actively participate and contribute, promoting cooperation and knowledge sharing among vehicles, and further improving the overall performance and stability of the intelligent transportation system. These improvement measures not only enhance the robustness and security of the system, but also improve the efficiency and effectiveness of machine learning in vehicle networking, and have significant practical application value.

[0043] The present invention aims to solve the problems of data privacy protection and low model training efficiency in the blockchain-based vehicle networking system. Aiming at the problems that a large amount of scattered data in vehicle networking applications is difficult to centrally process and the traditional methods have single-point failure problems, this paper proposes a knowledge sharing method based on distributed federated learning and blockchain technology. Through this method, the problems of data privacy leakage and insufficient optimization of node selection can be solved, and the efficiency and security of model training can be improved by designing a dynamic adaptive node selection method and a consensus mechanism based on knowledge distillation. Therefore, the present invention aims to design a distributed federated learning knowledge sharing method applicable to the blockchain-based vehicle networking architecture, so as to improve the cooperation efficiency, data privacy protection and overall performance of the vehicle networking system. Description of the Drawings

[0044] Figure 1 Vehicle networking knowledge sharing model diagram of distributed federated learning based on blockchain.

[0045] Figure 2 Flowchart of vehicle networking knowledge sharing of distributed federated learning based on blockchain. Detailed Implementation Modes

[0046] The technical method of the present invention will be further described in detail below with reference to the drawings in the specification.

[0047] As shown in Figure (1), this method establishes a vehicle networking knowledge sharing model of distributed federated learning based on blockchain. The model mainly includes vehicle nodes (Vehicle), roadside unit nodes (RSU), and edge cloud services.

[0048] Vehicle nodes: The main functions of vehicle nodes include data collection and preprocessing, local model training and uploading, participation in preliminary aggregation and consensus, and model optimization and knowledge sharing using knowledge distillation technology. Specifically, vehicle nodes first collect and preprocess their own feature data, then train the NFM model locally, and upload the training results to the RSU. Then, vehicle nodes participate in the aggregation of the preliminary model, select the best candidate block through the consensus mechanism, and finally use the global model for knowledge distillation to optimize the local model and achieve efficient knowledge sharing and collaborative learning.

[0049] Roadside units: As lightweight nodes of the blockchain, they collect and verify the local model parameters uploaded by vehicle nodes, perform aggregation, and select the best candidate block through the knowledge proof consensus mechanism. In addition, the RSU is also responsible for distributing the optimized model back to the vehicle nodes to ensure the effectiveness and accuracy of knowledge sharing, thereby improving the learning efficiency and model performance of the entire system.

[0050] As a full node of the blockchain, the edge cloud service collects the preliminarily aggregated model parameters from the roadside nodes, performs further global aggregation and optimization, and generates the final global model. Through knowledge distillation technology, the edge cloud service transfers the knowledge of the optimized teacher model to the roadside units to ensure the efficiency and consistency of the model. In addition, the edge cloud service is also responsible for coordinating data processing and model updates in the entire system, providing computing resources and storage support, thereby enhancing the overall performance and security of the system.

[0051] In step (2), a federated learning algorithm and a lightweight consensus mechanism based on the blockchain are constructed.

[0052] Step 2-1: Establish a Neural Factorization Machine (NFM) model by combining vehicle features, road features, environmental features, user features, time features, and other features. All participating nodes use this model for local training and updating. First, each vehicle node collects its own feature data, including vehicle features, road features, environmental features, user features, and time features, and classifies these feature vectors to distinguish continuous variables and embedding vectors; then, the collected data is standardized and normalized to ensure the consistency and usability of the data.

[0053] According to the collected data feature categories, establish a Neural Factorization Machine (NFM) model to predict the learning individual task objectives:

[0054]

[0055] where w T x is the linear part, representing the inner product of the feature vector x and the weight vector w. is the factorization part, used to capture the second-order interaction between features.

[0056] In the above formula, s represents the elements in the embedding vector, used to capture the interaction relationship between features. s i,f represents the embedding vector of feature i in the factor f dimension. x i represents the value of feature i. K represents the number of dimensions of the embedding vector, and n represents the number of features.

[0057] represents the square of the weighted sum of all features in the factor f dimension.

[0058] represents the weighted sum of squares of each feature i in the factor f.

[0059] Local loss function: where τ i respectively represent the number of samples and training parameters of node i. μ is the learning rate, is the gradient of the loss function.

[0060] Step 2-2: To improve the learning efficiency and accuracy, a dynamic adaptive learning node optimization selection algorithm is designed. With the goal of minimizing the local loss function and considering the heterogeneity of nodes and network conditions, the most suitable vehicle nodes for participating in learning are selected.

[0061] There is a set of vehicle nodes {v1, v2, …, v M}, and each node v i has a set of metric values {c i , m i , p i , V i , θ i , f i} representing its computing cost, communication cost, current location, current speed, direction angle, and data quality index.

[0062] According to Shannon's theorem, the communication rate of the i-th vehicle node at the t-th time slot can be expressed as:

[0063]

[0064] where B is the channel bandwidth between the vehicle and the roadside unit. S represents the average power of the transmitted signal in the channel, and N represents the power of Gaussian noise in the channel. Assuming that the communication of the vehicle network is carried out in an open space, the signal-to-noise ratio can be expressed as:

[0065]

[0066] where is the distance between the vehicle and the roadside unit, and η0 is the propagation factor.

[0067] The communication cost m i of node v i consists of the communication (uploading and downloading) time cost of each global round.

[0068]

[0069] where ψ represents the knowledge sharing size, and represent the download and upload communication rates respectively.

[0070] The computing cost c i of node v i is related to the current resource performance and the number of training samples and the local training round number ξi Related to

[0071]

[0072] Let the vehicle node v i The current position p i =(x i , y i ) Establish a motion trajectory model of the vehicle at time slot t:

[0073]

[0074] Calculate the Euclidean distance between the position of the vehicle at time slot t and the position of the RSU

[0075]

[0076] If (Communication radius of the roadside unit, usually taken between 300 meters and 1000 meters), then the vehicle is within the communication range of the roadside unit.

[0077] Let the local data quality index

[0078] Partial global loss function: Where M represents the number of nodes participating in learning; Represents the global training parameter; Γ represents the total number of samples of all nodes participating in learning.

[0079] This method aims to minimize the loss function and establish the following function:

[0080]

[0081] Constraint conditions:

[0082]

[0083] Among them, T max Respectively represent the maximum acceptable global training time, and T represents the total number of time slots included within the maximum acceptable global training time.

[0084] Use the gradient descent method combined with constrained optimization to solve the problem of minimizing the partial global loss function.

[0085] First, use the initialized global training parameter and learning rate μ.

[0086] Secondly, calculate the global loss function And its gradient

[0087]

[0088] Next, update the parameters using the gradient descent method:

[0089]

[0090] Calculate the distance of each node at each time slot t And detect and check whether the communication range constraint and the time resource limit constraint are satisfied. If the constraint conditions are violated, perform constraint projection to ensure that the parameters meet the constraint conditions. Repeat the above steps until the loss function converges.

[0091] Step 2-3: Construct a lightweight proof-of-knowledge (LPoK) consensus mechanism.

[0092] First, in the transaction verification phase, the roadside unit collects and verifies the learning transactions uploaded by the vehicle nodes to ensure their validity and accuracy. The learning result θ i generated by the vehicle node v i and its corresponding accuracy ACC i are uploaded to the roadside unit through the transaction, and the roadside unit verifies the validity and accuracy of each transaction.

[0093] Let the transaction set T = {(θ i , ACC i )|i = 1, 2, …, i, …, M}, where M is the number of vehicle nodes.

[0094] Secondly, in the block generation phase, after the roadside unit receives and verifies the transaction data of each node, it packs these transaction data into candidate blocks. The data structure of the candidate block is Candidate block B RSU is broadcast among multiple roadside units on the same roadside chain.

[0095]

[0096] After each vehicle node v i completes model training locally, it calculates its proof-of-knowledge score (PoK score).

[0097]

[0098] where ω1, ω2, ω3 are the weight coefficients of each dimension and

[0099] According to the historical performance of the node and the system requirements, dynamically adjust the coefficients of each dimension to ensure the fairness of the score. Let the initial weights be ω1, ω2, ω3, and adjust according to the historical performance of the node:

[0100]

[0101] where α is the adjustment coefficient, and ΔP i is the incremental performance of the node in the current learning round, and P i is the historical draw performance of the node.

[0102] For each candidate block the roadside unit calculates its comprehensive score

[0103]

[0104] where represents the knowledge proof score of the j-th transaction in the candidate block , and represents the frequency of participation in learning of the vehicle node where the j-th transaction in the candidate block is located.

[0105] After receiving multiple candidate blocks, the consensus node selects the block B with the highest comprehensive score * as the optimal candidate block, and then calculates the global loss function according to the training parameters in the transaction. Consensus is reached if and only if the global loss function obtained by validating all the training parameters by more than 1 / 3 of the roadside unit nodes is less than the threshold δ.

[0106]

[0107] Step 2-4: Design a blockchain-based federated learning algorithm.

[0108] First, for roadside-layer federated learning, vehicle nodes train a neural factorization machine (NFM) model on local data and upload the learning results to the roadside unit. The roadside unit performs preliminary aggregation on the collected learning results to form a preliminary global model.

[0109]

[0110] Secondly, the roadside unit uploads the generated block to the edge cloud service, and the cloud edge service performs knowledge distillation.

[0111] Loss function of knowledge distillation:

[0112] L distill = γL CE (Z s , y)+(1 - γ)L KL (Z s , Z T ) (21)

[0113] where, LCE is the cross - entropy loss function, I KL is the KL divergence, γ is the weight coefficient, Z s is the output of the roadside layer, Z T is the output of the edge layer, and y is the true label.

[0114] The optimized model parameters can be expressed as:

[0115]

[0116] where η is the learning rate, is the gradient of the distillation loss with respect to the parameters of the preliminary aggregation model.

[0117] Finally, the edge - cloud service will further aggregate according to the model parameters to form the final global model θ EC ,

[0118]

[0119] In step (3), design an incentive mechanism for non - cooperative games and use the deep Q - network (DQN) to solve the Stackelberg game model. Through fair and effective incentive strategies, encourage nodes to actively participate and contribute, thereby improving the efficiency, performance, and security of the distributed system.

[0120] Step 3 - 1, design the incentive function for vehicle nodes. The total communication cost of each vehicle node v i can be expressed as the sum of the communication time costs of multiple global rounds in the local training learning task. Assuming there are a total of T global rounds, according to formula (4), the total communication cost of vehicle node v i is:

[0121]

[0122] According to formula (5), the calculation cost of vehicle node v i is

[0123]

[0124] The incentive function of the vehicle node

[0125]

[0126] Step 3 - 2, design the incentive function for roadside nodes

[0127]

[0128] where Γ represents the total number of samples of all nodes participating in the learning, and log(Γ + 1) is used to smooth and control the growth rate of the incentive.

[0129] Step 3-3: Design a non-cooperative game model and use a deep Q-network (DQN) to solve the Stackelberg game model.

[0130] First, the goal of the roadside unit is to maximize its incentive function:

[0131]

[0132] Second, the goal of the vehicle node is to maximize its incentive function:

[0133]

[0134] Constraints:

[0135]

[0136] where T max , C max , I max represent the upper limits of communication resources, computing resources, and incentives respectively, and max{ψ i} and min{ψ i} represent the maximum and minimum values of the scale of allowed shared knowledge respectively.

[0137] Third, define the state space and action space, including the specific actions of the roadside unit to adjust the incentive strategy and the vehicle node to adjust the strategy and incentives. Combine the penalty function to handle the constraints such as communication resources, computing resources, and incentive budgets.

[0138] The state space S, that is, the various parameters in the game are used as states. The incentive strategy of the roadside unit The size of the shared knowledge ψ of the vehicle node i , the incentive of the vehicle node

[0139]

[0140] The action space A, the actions that the roadside unit and the vehicle node can take. The roadside unit can actively participate in consensus by generating blocks to adjust its incentive strategy The vehicle node can adjust the scale of knowledge sharing ψ i and improve the learning accuracy to adjust

[0141]

[0142] Define the Q function Q(s,α) to estimate the expected total reward obtained after taking action α in state s.

[0143]

[0144] where \(s\) is the current state, \(a\) is the current action, \(\gamma\) is the discount factor, usually between 0 and 1, which is used to balance future incentives and current incentives; is the incentive obtained at time step \(t\).

[0145] Finally, define the Q-network model, where the input layer is the dimension of the state space, the hidden layer consists of two fully connected layers, each containing 24 neurons, using ReLU as the activation function, and the output layer is the dimension of the action space, using a linear activation function. Then, initialize the DQN agent, set the discount factor \(\gamma\), the initial exploration rate, the minimum exploration rate, the exploration rate decay, and prepare the memory capacity to store experiences.

[0146] During the training process, initialize the environment and the agent, and simulate multiple episodes through iteration. In each episode, the agent selects an action according to the current state and executes it, updates the state, and calculates the reward. Store the experience of each step in the memory, and randomly sample a small batch of experiences from the memory for replay to update the Q-network. Through continuous iteration and training, the agent gradually learns the optimal incentive strategy and the strategy of vehicle nodes, so as to maximize the cumulative reward and meet the constraints. Through the optimized strategy, the optimal solution in the Stackelberg game model is achieved.

Claims

1. A distributed federated learning vehicle network knowledge sharing method based on blockchain, characterized in that: The following steps are involved: S1: Adopting Neural Factorization Machine (NFM) as the prediction network, combined with a dynamic adaptive node selection algorithm, to select the most suitable vehicle node to participate in learning; S2: Based on the vehicle node selected in S1, the vehicle node performs NFM model training locally and uploads the training results as transaction data to the roadside unit; S3: The roadside unit generates candidate blocks based on the transaction data uploaded in S2, and ensures the transparency and immutability of the data through a consensus mechanism based on proof of knowledge; S4: Based on the candidate blocks generated in S3, the edge cloud service uses knowledge distillation technology to update the global model and optimize the model performance; S5: Based on the optimized global model in S4, a non-cooperative game incentive model for vehicle nodes and roadside units is constructed, and the deep Q network (DQN) is used to reasonably allocate incentives to encourage nodes to actively participate in learning and contribute.

2. According to claim 1, a distributed federated learning vehicle networking knowledge sharing method based on blockchain is characterized in that: In S1, a neural factorization machine (NFM) is used as a prediction network, combined with a dynamic adaptive node selection algorithm to select the most suitable vehicle node for learning, specifically including: The vehicle node collects its own feature data, including vehicle features, road features, environmental features, user features, and time features, and classifies these feature vectors to distinguish between continuous variables and embedded vectors; Standardize and normalize the collected data to ensure data consistency and usability; A dynamic adaptive node selection algorithm is designed to minimize the partial global loss function and select the most suitable vehicle nodes to participate in learning, providing optimized node selection for subsequent steps.

3. According to a blockchain-based distributed federated learning vehicle networking knowledge sharing method according to claim 1, it is characterized in that: The vehicle node in S2 transmits the learning results as transaction data to the roadside unit, and uploads the consensus blocks to the edge cloud through a consensus mechanism based on knowledge proof, including: Based on the vehicle node selected in S1, the vehicle node performs NFM model training locally and uploads the training results to the roadside unit; The roadside unit collects and verifies the local model parameters uploaded by the vehicle nodes, generates candidate blocks, and provides a data basis for the subsequent consensus mechanism; Through a consensus mechanism based on proof of knowledge, the transactions with the highest accuracy are selected for block generation to ensure the transparency and immutability of the data, providing reliable data support for subsequent model updates. Upload the blocks with successful consensus to the edge cloud service to provide reliable data support for subsequent global model updates.

4. According to a blockchain-based distributed federated learning vehicle networking knowledge sharing method according to claim 1, it is characterized in that: The edge cloud service in S4 uses knowledge distillation technology to update the global model, specifically including: Based on the candidate blocks generated in S3, the edge cloud service performs knowledge distillation; Through knowledge distillation technology, the optimized teacher model knowledge is transferred to the roadside unit to ensure the efficiency and consistency of the model; The edge cloud service further aggregates and optimizes the model parameters to generate the final global model, providing an optimized model basis for subsequent incentive allocation.

5. According to a blockchain-based distributed federated learning vehicle networking knowledge sharing method according to claim 1, it is characterized in that: The non-cooperative game incentive model of vehicle nodes and roadside units is constructed in S5, specifically including: Based on the global model optimized in S4, the incentive function of the vehicle node is designed, and its incentive is calculated based on the communication cost, computing cost and data quality index of the vehicle node; Design the incentive function of the roadside unit and calculate its incentive based on the contribution and participation frequency of the vehicle node; The Deep Q Network (DQN) is used to solve the Stackelberg game model, optimize the incentive allocation strategy, and encourage vehicle nodes and roadside units to actively participate in learning and contribution, ensuring the continuous and efficient operation of the system.

6. A distributed federated learning vehicle networking knowledge sharing method based on blockchain according to claim 2, characterized in that: The dynamic adaptive node selection algorithm specifically includes: Select the most suitable vehicle node to participate in learning according to the calculation cost, communication cost, current position, current speed, direction angle and data quality index of the vehicle node; By combining the gradient descent method with constrained optimization, we solve the problem of minimizing some global loss functions, ensuring that the selected nodes meet the communication range and time resource constraints, and providing optimized node selection for subsequent steps.

7. A distributed federated learning vehicle networking knowledge sharing method based on blockchain according to claim 3, characterized in that: The consensus mechanism based on knowledge proof specifically includes: The roadside unit collects and verifies the learning transactions uploaded by the vehicle nodes to ensure their validity and accuracy; The roadside unit packages the verified transactions into candidate blocks and selects the transactions with the highest accuracy for block generation based on the learning results; By calculating the PoK score of each candidate block, the optimal candidate block is selected for broadcasting to ensure the transparency and immutability of the data, providing reliable data support for subsequent model updates.

8. According to claim 4, a distributed federated learning vehicle network knowledge sharing method based on blockchain is characterized in that: The knowledge distillation technology specifically includes: Through the cross entropy loss function and KL divergence, the knowledge of complex models is transferred to simpler models to achieve model compression and knowledge sharing; Through knowledge distillation technology, the model parameters are optimized, the system's learning efficiency and the model's generalization ability are improved, providing an optimized model foundation for subsequent incentive allocation.

9. According to claim 5, a distributed federated learning vehicle network knowledge sharing method based on blockchain is characterized in that: The deep Q network (DQN) solves the Stackelberg game model specifically including: Define the state space and action space, including the specific actions of adjusting incentive strategies for roadside units and vehicle node adjustment strategies and incentives; The Q network model is used to estimate the expected total reward after taking action α in state s and optimize the incentive allocation strategy; Through continuous iteration and training, the intelligent agent gradually learns the optimal incentive strategy and vehicle node strategy, maximizes the cumulative reward and satisfies the constraints, ensuring the continuous and efficient operation of the system.

Images