Dynamic trusted aggregation method and system for vehicle networking federated learning

Abstract

This invention provides a dynamic trusted aggregation method and system for federated learning in vehicle-to-everything (V2X) networks, comprising: verifying the decentralized identity and verifiable credentials of a target node and generating a participation token; having a vehicle node in the target node receive a global model and generate model update parameters and lightweight compliance proofs for the global model based on local data, generating an update message for the global model; calculating and obtaining a network-side security score for the vehicle node; determining whether the network-side security score triggers a preset abnormal baseline; calculating a total trusted score, and determining whether the total trusted score triggers a preset mandatory isolation condition. If the result is yes, the aggregation weight of the vehicle node is set to zero and the vehicle node is forcibly isolated; if the result is no, the global model is updated based on the total trusted score. This invention reduces the risk of unauthorized or forged nodes participating in federated learning.

Description

Technical Field

[0001] This invention belongs to the fields of intelligent transportation, information security, and edge computing technology. Specifically, it relates to a dynamic trusted aggregation method and system for federated learning in vehicle-to-everything (V2X) networks. More specifically, it is a trusted update and dynamic aggregation method for V2X federated learning based on blockchain and a zero-trust architecture. Background Technology

[0002] With the rapid development of vehicle-to-everything (V2X) technology, edge devices such as on-board units (OBU) and roadside units (RSU) have generated massive amounts of data.

[0003] However, connected vehicle data is highly fragmented and privacy-sensitive, posing risks of data privacy breaches and communication bandwidth bottlenecks to traditional centralized machine learning methods. Federated learning, as a privacy-preserving distributed machine learning paradigm, enables multiple nodes to collaboratively train a global model without sharing local data through a mechanism of "data remaining stationary while model updates flow," making it an important solution to the data silo problem in connected vehicles.

[0004] Despite the advantages of federated learning in terms of privacy protection, existing security and aggregation mechanisms still have the following pain points and shortcomings in practical applications in the complex and dynamic environment of connected vehicles: First, existing federated learning defense algorithms, such as the Krum algorithm, the median algorithm, and the truncated mean algorithm, usually rely on the statistical distance or directional consistency between model updates when dealing with malicious node poisoning attacks.

[0005] However, perception data and driving trajectories in connected vehicle scenarios naturally exhibit a high degree of non-independent identically distributed characteristics (Non-IID) and long-tail features. When faced with this highly Non-IID long-tail data, existing algorithms may misjudge effective model updates generated from rare, real-world edge scenarios, such as extreme weather, sudden road conditions, or uncommon driving behaviors, as malicious poisoning or unknown types of poisoning attacks, and thus reject them. This simplistic rejection mechanism reduces the generalization ability and performance of the global model in both edge and critical scenarios.

[0006] Second, since federated learning cannot directly examine local training data, some blockchain-based model auditing solutions, if they directly verify model updates on-chain, may need to access plaintext model update parameters or only record hash digests, which is insufficient to support a trustworthy assessment. The former poses a privacy risk, while the latter makes it difficult to detect malicious updates and backdoor injection attacks in a timely manner. Therefore, a collaborative mechanism with clearly defined responsibilities is needed among privacy protection, trustworthy auditing, and anomaly detection.

[0007] Third, the computing resources of on-board units (OBUs) and edge sensing devices are limited, making it difficult to handle the computational overhead of complex homomorphic encryption or large-scale multi-party secure computation in traditional encryption methods. Meanwhile, the consensus mechanisms of traditional public blockchains, such as Proof-of-Work (PoW), have high computational complexity and excessive block confirmation latency, making it difficult to meet the low-latency response requirements of vehicle-to-everything (V2X) security early warning scenarios.

[0008] Patent document CN119363469A discloses a blockchain-driven reward and punishment mechanism and a federated learning method for detecting vehicle-to-everything (V2X) behavior. This scheme aims to optimize the resource utilization efficiency in the V2X federated learning environment. However, it does not meet the low-latency response requirements in V2X security early warning scenarios and still relies heavily on the statistical distance or directional consistency between model updates.

[0009] This problem urgently needs to be solved. Summary of the Invention

[0010] To address the shortcomings of existing technologies, the purpose of this invention is to provide a dynamic trusted aggregation method and system for federated learning in the Internet of Vehicles.

[0011] A dynamic trusted aggregation method for federated learning in vehicle networks, provided by the present invention, includes: Step 1: Verify the decentralized identity and verifiable credentials of the target node, and generate a participation token; Step 2: Based on the participation token, the vehicle node in the target node receives the global model and generates the model update parameters and lightweight compliance proof of the global model according to the local data. Then, based on the model update parameters and lightweight compliance proof, the update message of the global model is generated. Step 3: Based on the update message, calculate and obtain the network-side security score of the vehicle node; determine whether the network-side security score triggers the preset abnormal baseline; if the result is yes, add the corresponding node to the isolation candidate set; if the result is no, do not process it. Step 4: Based on the network-side security score and the trustworthiness of the vehicle node, calculate the total trust score. Based on whether the total trust score triggers the preset mandatory isolation condition, if the result is yes, set the aggregation weight of the vehicle node to zero and forcibly isolate the vehicle node; if the result is no, update the global model according to the total trust score.

[0012] Preferably, between step three and step four, the method further includes: having the verification node or RSU aggregation node of the target node anomaly detect the global model of the vehicle node, obtaining a verification result, and hashing the verification result into the permissioned chain; the verification result includes a statistical risk score; the statistical risk score is obtained based on the CKA anomaly score; In step four, the credibility of the vehicle node includes the statistical credibility of the model update; the statistical credibility of the model update is obtained based on the statistical risk score.

[0013] Preferably, in step three, the expression for the network-side security score is:

[0014] in, Indicates network-side security score; Indicates the probability of an attack; This indicates a communication quality score.

[0015] Preferably, in step four, the mathematical expression for the total credibility score is:

[0016] in, Indicates the overall credibility score. This indicates that the rating will be limited to Within the interval, As a penalty item, This represents the historical reliability score of the comprehensive vehicle node. Indicates network security score. This represents the result of a zero-knowledge proof. This indicates the statistical confidence level of the model update. This refers to the credibility of historical behavior, derived from vehicle node historical proof results, network scores, statistical risks, and isolation records. It is also known as historical behavior credibility. The weighting coefficients are satisfied. .

[0017] Preferably, in step four, the global model is updated based on the total confidence score, and the mathematical expression is:

[0018] in, This represents the updated global model. Represents the global model; Indicates aggregate weight; This indicates that the model is updating its parameters; Represents vehicle node In the wheel isolation markings, among which This indicates that the node is not isolated and can participate in aggregation. This indicates that the node is isolated. This represents the summation of all unisolated vehicle nodes.

[0019] The mathematical expression for the aggregation weight is:

[0020] in, Indicates aggregate weight, and Representing vehicle nodes respectively With vehicle nodes The number of local samples, and Representing the corresponding vehicle nodes With vehicle nodes In the The overall credibility score of the wheel This represents the summation of all vehicle nodes that meet the condition of not being isolated.

[0021] A dynamic trusted aggregation system for federated learning in vehicle networks, provided by the present invention, includes: Module 1: Verify the decentralized identity and verifiable credentials of the target node, and generate a participation token; Module 2: Based on the participation token, the vehicle node in the target node receives the global model, and generates the model update parameters and lightweight compliance proof of the global model according to the local data. Then, based on the model update parameters and lightweight compliance proof, an update message of the global model is generated. Module 3: Based on the update message, calculate and obtain the network-side security score of the vehicle node; determine whether the network-side security score triggers a preset abnormal baseline; if the result is yes, add the corresponding node to the isolation candidate set; if the result is no, do not process it. Module 4: Based on the network-side security score and the trustworthiness of the vehicle node, calculate the total trust score. Based on whether the total trust score triggers a preset mandatory isolation condition, if the result is yes, set the aggregation weight of the vehicle node to zero and forcibly isolate the vehicle node; if the result is no, update the global model according to the total trust score.

[0022] Preferably, between the operation of trigger module three and the operation of trigger module four, the method further includes: having the verification node or RSU aggregation node of the target node abnormally detect the global model of the vehicle node, obtaining a verification result, and hashing the verification result into the permissioned chain; the verification result includes a statistical risk score; the statistical risk score is obtained based on the CKA anomaly score; In step four, the credibility of the vehicle node includes the statistical credibility of the model update; the statistical credibility of the model update is obtained based on the statistical risk score.

[0023] Preferably, in module three, the expression for the network-side security score is:

[0024] in, Indicates network-side security score; Indicates the probability of an attack; Represents communication quality score, symbol It represents the product.

[0025] Preferably, in module four, the mathematical expression for the total credibility score is:

[0026] in, Indicates the overall credibility score. This indicates that the rating will be limited to Within the interval, As a penalty item, This represents the historical reliability score of the comprehensive vehicle node. Indicates network security score. This represents the result of a zero-knowledge proof. This indicates the statistical confidence level of the model update. This represents the credibility of historical behavior, derived from vehicle node historical proof results, network scores, statistical risks, and isolation records. It is referred to as historical behavior credibility. The weighting coefficients are satisfied. .

[0027] Preferably, in module four, the global model is updated based on the total credibility score, and the mathematical expression is:

[0028] in, This represents the updated global model. Represents the global model; Indicates aggregate weight; This indicates that the model is updating its parameters; Represents vehicle node In the The wheel isolation markings, among which, This indicates that the node is not isolated and can participate in aggregation. This indicates that the node is isolated. This represents the summation of all unisolated vehicle nodes.

[0029] The mathematical expression for the aggregation weight is:

[0030] in, Indicates aggregate weight, and Representing vehicle nodes respectively With vehicle nodes The number of local samples, and Representing the corresponding vehicle nodes With vehicle nodes In the The overall credibility score of the wheel This represents the summation of all vehicle nodes that meet the condition of not being isolated.

[0031] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention reduces the risk of unauthorized or forged nodes participating in federated learning and improves the credibility of node admission by using DID / VC and permissioned chain admission mechanisms.

[0032] 2. This invention can improve the trusted aggregation capability of connected vehicle federated learning, reduce the risk of misjudgment between abnormal updates and real edge scenarios, and support the traceability and auditing of the training process.

[0033] 3. This invention uses a zero-knowledge proof system to verify whether the model update parameters meet the preset lightweight constraints without obtaining the original data, thus achieving privacy-preserving compliance verification of model updates. Furthermore, this invention uses HSIC / CKA similarity and context grouping to avoid simple mis-removal of rare real-world scenarios to a certain extent, reducing the probability of misjudging real Non-IID data.

[0034] 4. This invention enhances the system's protection against backdoor poisoning and abnormal communication behavior by dynamically and reliably weighting multi-source information, thereby improving the robustness of the model aggregation process. Furthermore, the key interaction processes, proof results, detection summaries, and model versions provided by this invention are all stored on the blockchain for evidence, providing a basis for subsequent accountability and model rollback, thus achieving a traceable and auditable security closed loop. Attached Figure Description

[0035] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the system architecture topology provided by the present invention; Figure 2 This is a schematic diagram of the method flow provided by the present invention; Figure 3 A schematic diagram illustrating the logic of off-chain detection and on-chain audit and evidence preservation provided by this invention; Figure 4 This is a schematic diagram of dynamic credibility scoring and aggregate weight calculation provided by the present invention. Detailed Implementation

[0036] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0037] To address the problems existing in current connected vehicle federated learning systems, such as insufficient vehicle node identity trustworthiness, difficulty in verifying model updates, difficulty in distinguishing malicious poisoning from real Non-IID data, difficulty in balancing on-chain auditing and privacy protection, and difficulty in timely feedback of edge security risks, the purpose of this invention is to provide a dynamic trusted aggregation method, system, and device for connected vehicle federated learning.

[0038] The core objectives of this invention specifically include: first, to achieve lightweight compliance verification of vehicle model update parameters without uploading original vehicle data; second, to reduce the probability of misjudging extreme road scenarios as malicious poisoning in highly Non-IID scenarios; third, to enable trusted nodes to obtain higher aggregation weights through a dynamic scoring mechanism, while degrading or isolating abnormal nodes; and fourth, to utilize permissioned chains to achieve traceable auditing of proof results, detection summaries, score changes, and model versions.

[0039] This invention discloses a dynamic trusted aggregation method, system, and device for federated learning in the Internet of Vehicles (IoV). Addressing the issues of insufficient identity trust, susceptibility to misjudgment of genuine Non-IID updates, and the difficulty in balancing on-chain auditing and privacy protection in IoV federated learning, this invention achieves zero-trust access based on DID / VC, utilizes a zero-knowledge proof system to generate lightweight compliance proofs for model update parameters, performs network-side scoring through RSU, and verifier nodes perform anomaly detection off-chain using HSIC / CKA, context grouping, and historical consistency. The detection summaries and scoring results are stored on-chain as evidence, and finally, aggregation weights are allocated or isolation is performed based on the dynamic trusted score.

[0040] Example 1: A dynamic trusted aggregation method for federated learning in vehicle networks according to the present invention includes: Step S1: Zero-trust admission based on DID / VC: Before the start of each round of federated learning task, vehicle nodes... RSU edge nodes and validator nodes verify decentralized identity tokens through permissioned chain smart contracts or off-chain identity verification modules. and verifiable credentials After successful verification, generate the federated learning task and the current training round. Bound participation token Among them, footnotes This indicates the vehicle node number.

[0041] If the verification fails, the system will not generate a participation token for the current round and will refuse the vehicle node from submitting a model update for the current round; the verification failure record will also be written to the permissioned chain or local audit log for subsequent traceability and risk assessment.

[0042] Specifically, participation tokens In step S2, it serves as a binding credential for model updates and zero-knowledge proofs, proving that the update comes from an admitted node and corresponds to the current task and the current training round. Step S2, Lightweight training and privacy compliance proof generation on the device side: Vehicle node Receive global model Parameters are updated based on the model generated from local data. It generates lightweight compliance proofs based on a zero-knowledge proof system. ; The lightweight compliance proof is used to prove, without disclosing the original vehicle data, that the model update parameters meet the binding constraints related to identity, round, model version and update commitment, as well as the compliance constraints related to update magnitude, dimension and format. Specifically, in step S2, the model update parameters are quantized, and the vehicle node generates a lightweight compliance proof based on a zero-knowledge proof system. This is used to prove that the quantization update vector satisfies constraints such as task round binding, model version binding, consistent update commitment, consistent dimensions, consistent quantization format, restricted overall update norm, restricted single coordinate amplitude, and restricted update variance.

[0043] Step S3, RSU Edge Gateway Cleaning and Network-Side Dynamic Scoring: After receiving the model update message submitted by the vehicle node, the RSU performs message verification and intrusion detection, and calculates the network-side security score of the vehicle node based on its communication behavior. If an abnormal baseline is triggered, the corresponding node will be added to the isolation candidate set. Step S4, Off-chain Model Update Anomaly Detection and On-chain Audit and Evidence Storage: Validator nodes or RSU aggregation nodes perform anomaly detection on the vehicle node's model update off-chain to obtain a statistical risk score. The verification results, off-chain detection summary, trust score, isolation status, and model version hash are written to the permissioned chain. Step S5, Dynamic Trusted Aggregation and Forced Isolation: Calculate the total trust score by integrating the trustworthiness of vehicle nodes, including the trustworthiness of their proofs, network-side security scores, model update statistical trustworthiness, and historical behavior trustworthiness. If a node triggers the mandatory isolation condition, then its aggregate weights are applied. Set to 0; otherwise, perform weighted aggregation based on the total credibility score and update the global model.

[0044] Specifically, the participation token in step S1 satisfies:

[0045] in, The hash function is used to prevent cross-task submissions, cross-round replay submissions, and submissions from unauthorized nodes. Indicates the federal learning task and the current training round. Bound participation token; Indicates the first Decentralized identity identifier for each node; Indicates the current training round; Indicates the Federal Learning Task Number; Represents a random number or a one-time random value, used to prevent participation tokens from being replayed; the symbol " "" indicates a string or data field concatenation operation. Specifically, the lightweight compliance proof in step S2 is used to prove that the model update parameters satisfy at least one of the following constraints: update commitment consistency, model dimension consistency, quantization format consistency, update overall norm constraint, single coordinate amplitude constraint, and update variance constraint; the zero-knowledge proof system includes Groth16, Plonk, zk-STARK, or Bulletproofs.

[0046] Specifically, in step S2, the vehicle nodes use symmetric quantization to convert the model update parameters into integer quantized update vectors:

[0047] in, To quantify the scaling factor, Update the parameter coordinates for the model; before performing aggregation, the aggregation node performs inverse quantization or equivalent fixed-point aggregation on qualified updates based on the quantization scaling factor; This represents the integer quantization update vector. Represents vehicle node In the Model update parameters submitted in round The Middle The values ​​of each coordinate. This represents the floor function.

[0048] Specifically, the network-side security score in step S3 is as follows:

[0049]

[0050] in, To score the quality of communication, and To adjust parameters; the abnormal baseline includes attack probability exceeding a threshold, message signature verification failure, inconsistent model update hash, invalid participation token, duplicate submission, or submission in an expired round. Represents the natural constant An exponential function with base 0; Represents vehicle node In the Communication latency of submitting model update messages in rounds; Represents vehicle node In the Packet loss rate during round-robin communication; Indicates network-side security score; Indicates the probability of an attack; Specifically, the abnormal baseline is used to determine whether there are obvious communication or message anomalies at vehicle nodes; if the abnormal baseline is triggered, the network-side security score is directly adjusted. Set the value to zero and add the node to the isolation candidate set; if no abnormal baseline is triggered, calculate the network-side security score based on the attack probability, communication latency, and packet loss rate. . Specifically, the off-chain model update anomaly detection in step S4 includes: using the global model and the client model to analyze the public reference sample set. Extract the intermediate layer representation and calculate the CKA similarity:

[0051] in, and These are the intermediate layer representations extracted from the common reference sample set for the client model and the global model, respectively. The Hilbert-Schmidt independence criterion is used; and the vehicles are grouped according to their traffic context. Within the same context cluster, the CKA anomaly score, orientation anomaly score, and norm anomaly score are calculated to obtain the statistical risk score. Represents vehicle node In the CKA similarity to the corresponding wheel; Specifically, CKA, or Centered Kernel Alignment, is a term used in federated learning to measure the similarity between different models. The system first divides vehicle nodes into different contextual clusters based on contextual information such as weather, road type, and region. Then, within each cluster, it calculates the CKA deviation, update direction deviation, and update norm deviation for each node using the median absolute deviation method, and synthesizes these to obtain a statistical risk score. The direction anomaly score measures the deviation between the vehicle node model's update direction and the average update direction within the same contextual cluster; the norm anomaly score measures the deviation between the vehicle node model's update magnitude and the normal update magnitude range within the same contextual cluster.

[0052] Specifically, the statistical confidence level of the model update in step S5 is:

[0053] in, This indicates the statistical confidence level of the model update.

[0054] Total credibility score, i.e. The expression is:

[0055] in, This indicates that the rating will be limited to Within the interval, As a penalty item, The weighting coefficients are satisfied. .

[0056] in, This represents the historical reliability score of the comprehensive vehicle node. Indicates network security score. This represents the result of a zero-knowledge proof. This indicates the statistical confidence level of the model update. This indicates the credibility of historical behavior based on vehicle node historical proof results, network scores, statistical risks, and isolation records.

[0057] Specifically, if a node meets any of the following conditions: zero-knowledge proof verification fails, network intrusion detection attack probability exceeds a threshold, statistical risk score exceeds a strong isolation threshold, or total trust score falls below a minimum threshold, then an isolation flag is set. And make the aggregate weight For unisolated nodes, the aggregation weight is:

[0058] in, This indicates the number of local samples at the vehicle node. Indicates aggregate weight, and Representing vehicle nodes respectively and The number of local samples, Represents vehicle node In the The overall credibility score of the wheel Represents vehicle node In the Wheel isolation markings. This represents the summation over all unisolated vehicle nodes; Global model updated to The expression is:

[0059] in, This indicates the number of local samples at the vehicle node. This represents the updated global model. This indicates nodes that are not isolated and can participate in aggregation.

[0060] Example: A dynamic trusted aggregation method for federated learning in vehicle networks according to the present invention includes the following steps: Step S1: Zero-trust access based on DID / VC: Before the start of each round of federated learning tasks, vehicle nodes, RSU edge nodes, and validator nodes verify decentralized identity identifiers through permissioned blockchain smart contracts or off-chain identity verification modules. and verifiable credentials After successful verification, the current training round is generated. Participation token ; Step S2, Lightweight training and privacy compliance proof generation on the device side: Vehicle node Receive global model Parameters are updated based on the model generated from local data. The model update parameters are quantified; furthermore, vehicle nodes generate lightweight compliance proofs based on a zero-knowledge proof system. This is used to prove that the quantization update vector satisfies constraints such as task round binding, model version binding, consistent update commitment, consistent dimension, consistent quantization format, restricted overall update norm, restricted single coordinate amplitude, and restricted update variance. Step S3, RSU Edge Gateway Cleaning and Network-Side Dynamic Scoring: After receiving update messages, the RSU identifies abnormal communication behavior through the intrusion detection system and combines it with attack probability. Calculate the network-side security score of vehicle nodes based on communication latency and packet loss rate. If an abnormal baseline is triggered, the score will be set to zero and the corresponding node will be added to the isolation candidate set. Step S4, Off-chain model update anomaly detection and on-chain audit evidence storage: Verification nodes or RSU aggregation nodes use the global model and client model to analyze the public reference sample set. Extract the intermediate layer representation and calculate the centered kernel-aligned CKA similarity:

[0061] Then, based on the vehicle traffic context, the system groups the vehicles and calculates the CKA anomaly score, direction anomaly score, and norm anomaly score. Historical update consistency is then incorporated for correction to obtain the statistical risk score. The permissioned chain records the verification results, detection digest, verification node signature, trust score, isolation status, and model version hash. Step S5, Dynamic Trusted Aggregation and Forced Isolation: Integrating Historical Trust Scores of Vehicle Nodes Zero-knowledge proof results Network security score Model update statistical reliability And the credibility of historical behavior, calculate the first Overall Credibility Score of Wheel If a node triggers the mandatory isolation condition, then its aggregate weights are applied. Set it to 0, otherwise set its total credibility score. The weights are normalized together with the number of local samples to form aggregate weights, and then a global model update is performed.

[0062] In this invention, model immune detection refers to the off-chain identification process of suspected abnormal model updates based on the similarity of public reference sample representations, model update direction, model update norm, and historical behavior consistency; the abnormal baseline refers to the edge interception rules composed of attack probability threshold, message integrity, signature validity, round token validity, and duplicate submission; the statistical risk score refers to the risk measure obtained by combining representation similarity anomaly, direction anomaly, norm anomaly, and historical consistency correction; and the historical behavior credibility refers to the reputation index calculated based on the past proof results of vehicle nodes, network score, statistical risk, and isolation records.

[0063] System architecture and operating environment, combined Figure 1 As shown, the present invention provides a dynamic trusted aggregation system for federated learning in vehicle networks, which mainly includes the following entities: 1. Vehicle Node (OBU): Responsible for local data acquisition, local model training, and generating model update parameters. It uses the INT8 symmetric quantization algorithm to convert floating-point model update parameters into 8-bit integer quantized update vectors, thereby reducing the overhead of proof circuitry and communication. Subsequently, a zero-knowledge proof system such as Groth16, Plonk, zk-STARK, or Bulletproofs is selected to construct the proof process and generate lightweight compliance proofs.

[0064] 2. Edge Node (RSU): Responsible for receiving vehicle update messages, performing communication-side security checks through the Intrusion Detection System (IDS), calculating network-side trust scores, intercepting abnormal messages, and forwarding qualified updates.

[0065] 3. Verification Node: Responsible for verifying zero-knowledge proofs, performing off-chain HSIC / CKA statistical checks, and generating check summaries and signatures.

[0066] 4. Permissioned Blockchain Network: This can utilize Hyperledger Fabric's Raft ordering service or BFT / SmartBFT ordering service, or FISCO BCOS, Tendermint, or other permissioned blockchain frameworks that support smart contracts and consortium consensus. Other consortium blockchains can employ PBFT-like Byzantine fault-tolerant consensus mechanisms. The permissioned blockchain is responsible for identity access control, proof verification result recording, trust score recording, isolation records, detection summaries, and model version auditing.

[0067] 5. Global Aggregation Node: Responsible for performing federated aggregation based on dynamic trust weights to generate the next round of global model.

[0068] In this embodiment, the following parameters are configured during the system initialization phase: Federated learning task number. Initial global model Update the upper bound of the global norm. Upper bound of single-axis update amplitude Update the upper bound of variance. IDS attack probability anomaly threshold Strong isolation threshold for statistical anomalies Minimum threshold for total credibility score Simultaneously configure dynamic credibility scoring weight parameters. And satisfy .

[0069] Data Structure and Interaction Message Definition. In one implementation, the model update message submitted by a vehicle node to the RSU or verification node includes at least the task number, training round, vehicle identity identifier or its hash, verifiable credential hash, global model hash, model update commitment, zero-knowledge proof, quantization format, quantization scaling factor, and traffic context metadata. The traffic context metadata may include weather, road segment type, region, speed range, or sensor type, for subsequent context grouping detection.

[0070] On-chain audit logs should include at least the task number, training round, node identity hash, zero-knowledge proof verification result, network security score, statistical risk score, total trust score, isolation status, detection result hash, verification node signature, and model version hash. These fields are only used to illustrate one feasible data organization method; in actual deployment, vehicle identification and detection details can be hashed or digested according to privacy protection requirements.

[0071] Example: Trusted update and dynamic aggregation process for normal vehicle nodes. Combined with... Figure 2 With a legitimate and normal vehicle node For example, the complete process of participating in a round of federated learning is as follows: Step 1: Vehicle Identity Verification Node Submit its decentralized identity token to the permissioned blockchain smart contract or off-chain identity verification module and verifiable credentials System verification The issuing authority's signature, validity period, revocation status, and role permissions. After successful verification, it becomes... Generate the current round Training token:

[0072] Step 2: Local training, quantization, and lightweight proof to generate vehicle nodes Receive global model Model update parameters are obtained by training on local data. and to conduct Symmetric quantization:

[0073] in, This is the quantization scaling factor. The quantized update vector is used to generate the update commitment. . Generating zero-knowledge proofs based on the Groth16 algorithm Prove that the following lightweight statement is true:

[0074] in, Indicates the upper bound of the single-axis update magnitude; This represents a zero-knowledge proof generation algorithm; Represents a hash function; Indicates variance; Indicates the dimension of a vector; This represents a function to determine the updated data format. Represents an 8-bit integer quantization format; Represents vehicle node Quantization scaling factor, Represents the current global model The hash value; Represents vehicle node The update commitment for the quantization update vector in this round; Represents the squared 2-norm of a vector; This indicates that the upper bound of the updated variance has been updated.

[0075] The above proof does not require the complete unfolding of the deep neural network training process in a zero-knowledge circuit, nor does it separately determine whether the model update contains backdoor poisoning; its role is to verify that the submitted update meets the preset compliance constraints such as identity token, task round, model version, commitment and magnitude.

[0076] Step 3: RSU Network Side Scoring. After receiving packets, the RSU (Remote Subsystem for Security) detects the communication behavior and outputs the attack probability. A network-side security score is calculated by combining latency and packet loss rate. Not triggered The message is forwarded to the verification node.

[0077] Step 4: HSIC / CKA statistical testing. Combined with... Figure 3 The verification nodes utilize a common reference sample set. Extracting intermediate layer representations from the global model and the local model and Calculate the CKA similarity. Among them, Represents a public reference sample set The first in One reference sample.

[0078]

[0079] because No directional or norm anomalies were observed within the same context cluster, and a high statistical confidence level was obtained through statistical detection. .

[0080] Step 5: Dynamic Trust Aggregation. Combining Figure 4, the system calculates the total trust score by integrating various scores:

[0081] because Vehicle node Successfully participating in the aggregation, its weight is Aggregation nodes can perform inverse quantization on quantization updates based on the quantization scaling factor and then aggregate them, or perform equivalent aggregation within a fixed-point domain.

[0082] Example: Identification and retention process for real-world Non-IID updates in extreme scenarios. This applies to vehicle nodes on snowy roads. Its data distribution deviates from the normal scenario, and the calculated If compared directly with all vehicles, this update may be flagged as anomalous. This system uses its metadata... It is classified into the snowy context cluster. And calculate the relative outlier score within the cluster. .in, Represents vehicle node In the The similarity of the wheels to CKA This indicates that the weather type in the traffic context in which the node is located is snowy. This indicates that the road segment type in the traffic context in which the node is located is an urban road;

[0083] because There was nothing unusual within the group during the snowy weather, that is... , This represents the CKA anomaly score threshold, used to determine whether the deviation in representation similarity between vehicle nodes within the same context cluster exceeds a preset range; and combined with its good historical consistency, it ultimately corrects the statistical risk. The isolation level is low. The system does not enforce isolation of this node, i.e., the isolation flag is low. Instead, it assigns a lower but non-zero aggregation weight. To preserve edge scene features. This represents the median; This represents the absolute deviation function of the median; This indicates the smallest positive number added to avoid a denominator of zero; Indicates a snowy context family Each vehicle node in the first The CKA similarity set of wheels; Example: Detection and forced isolation process for malicious poisoning nodes. Targeting attack vehicles attempting to implant backdoors. It constructs an update that satisfies norm constraints, thus potentially verifiable through zero-knowledge proofs, i.e. .

[0084] However, the verification node is Figure 3 The off-chain detection shown revealed that its CKA similarity decreased to [value missing]. Furthermore, it deviates significantly from the average update direction of the same context cluster, i.e. This leads to its comprehensive statistical risk score. The smart contract triggers a mandatory isolation mechanism based on the detection digest of the verifier node's signature, setting... and . Represents vehicle node In the The cosine similarity between the model update direction of the wheel and the average update direction of the same context cluster. This represents the threshold for strong isolation of statistical risk. When the corrected statistical risk score exceeds this threshold, a mandatory isolation mechanism is triggered.

[0085] Simultaneously, the permissioned blockchain records the task number, training round, node identity hash, proof verification result, network security score, statistical risk score, total trust score, isolation status, detection result hash, and verification node signature corresponding to the node, for subsequent auditing and model rollback. This embodiment illustrates that lightweight compliance proofs and off-chain statistical detection can complement each other: the former verifies whether the submitted update meets the basic format and magnitude constraints, while the latter further identifies suspected poisoning or abnormal updates.

[0086] Example 4: Edge interception process for network attack nodes. This process targets nodes that launch DoS attacks, replay attacks, or expired round submissions. In step S3, RSU detects abnormally high frequency requests through IDS and outputs the attack probability. RSU directly assigns network-side security scores. Set the value to zero and intercept the update to reduce the computational burden on subsequent off-chain detection and aggregation nodes.

[0087] Smart contract execution logic: In one implementation, the permissioned blockchain smart contract executes the following steps: 1. Receive the vehicle node identity identifier, verifiable credential or identity verification module signature, and complete the identity validity verification; 2. Verify the participation token based on the task number, training round, and random number to determine if there are cross-task submissions, cross-round submissions, or replay submissions; 3. Receive and record the verification results of zero-knowledge proofs; 4. Verify the signatures of the verification nodes on the detection summary and scoring report; 5. Update the total trust score of the vehicle node based on the verification results, network security score, statistical risk score, and historical behavior trustworthiness. 6. If the verification fails, the statistical risk exceeds the threshold, or the total credibility score falls below the minimum threshold, an isolation flag is set. And make the aggregate weight Otherwise set This marks the node as eligible to participate in aggregation; 7. Write audit logs, isolation status, and model version hashes to the permission chain.

[0088] Other optional implementation methods. The technical solution of the present invention has good scalability: Specifically, the zero-knowledge proof system in step S2 can be replaced with zk-STARK, Bulletproofs, or other proof systems capable of proving preset lightweight compliance constraints, depending on the deployment conditions. Permissioned blockchain networks are not limited to Hyperledger Fabric; they can also use underlying blockchain architectures that support smart contracts, such as FISCO BCOS and Tendermint. In the dynamic credibility scoring formula, the weighting parameter is the total credibility score. It can be dynamically adjusted according to the actual application scenario. For example, it can be increased in high-security scenarios. and It places greater emphasis on proof and statistical testing, and can be adjusted higher in high real-time scenarios. It places greater emphasis on RSU edge filtering.

[0089] According to the present invention, a dynamic trusted aggregation system for federated learning in vehicle networks is provided to implement the method, comprising: vehicle nodes for local data collection, local model training, generating model update parameters and lightweight compliance proofs; edge nodes for receiving vehicle update messages, completing communication-side security detection, and calculating network-side security scores; verification nodes for verifying zero-knowledge proofs, performing off-chain statistical detection, and generating detection summaries and signatures; a permissioned blockchain network for identity access, recording proof verification results, trusted scores, isolation records, detection summaries, and model version auditing; and a global aggregation node for performing federated aggregation according to dynamic trusted weights to generate a global model.

[0090] An electronic device according to the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method described.

[0091] The present invention also provides a dynamic trusted aggregation system for federated learning in vehicle networks. The dynamic trusted aggregation system for federated learning in vehicle networks can be implemented by executing the process steps of the dynamic trusted aggregation method for federated learning in vehicle networks. That is, those skilled in the art can understand the dynamic trusted aggregation method for federated learning in vehicle networks as a preferred embodiment of the dynamic trusted aggregation system for federated learning in vehicle networks.

[0092] A dynamic trusted aggregation system for federated learning in vehicle networks, provided by the present invention, includes: Module 1: Verify the decentralized identity and verifiable credentials of the target node, and generate a participation token; Module 2: Based on the participation token, the vehicle node in the target node receives the global model, and generates the model update parameters and lightweight compliance proof of the global model according to the local data. Then, based on the model update parameters and lightweight compliance proof, an update message of the global model is generated. Module 3: Based on the update message, calculate and obtain the network-side security score of the vehicle node; determine whether the network-side security score triggers a preset abnormal baseline; if the result is yes, add the corresponding node to the isolation candidate set; if the result is no, do not process it. Module 4: Based on the network-side security score and the trustworthiness of the vehicle node, calculate the total trust score. Based on whether the total trust score triggers a preset mandatory isolation condition, if the result is yes, set the aggregation weight of the vehicle node to zero and forcibly isolate the vehicle node; if the result is no, update the global model according to the total trust score.

[0093] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0094] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A dynamic trusted aggregation method for federated learning in vehicle networks, characterized in that, include: Step 1: Verify the decentralized identity and verifiable credentials of the target node, and generate a participation token; Step 2: Based on the participation token, the vehicle node in the target node receives the global model and generates the model update parameters and lightweight compliance proof of the global model according to the local data. Then, based on the model update parameters and lightweight compliance proof, the update message of the global model is generated. Step 3: Based on the update message, calculate and obtain the network-side security score of the vehicle node; Determine whether the network-side security score triggers a preset abnormal baseline; If the result is yes, the corresponding node is added to the isolation candidate set; if the result is no, nothing is done. Step 4: Based on the network-side security score and the trustworthiness of the vehicle node, calculate the total trust score. Based on whether the total trust score triggers the preset mandatory isolation condition, if the result is yes, set the aggregation weight of the vehicle node to zero and forcibly isolate the vehicle node; if the result is no, update the global model according to the total trust score.

2. The dynamic trusted aggregation method for federated learning in vehicle networking according to claim 1, characterized in that, Between steps three and four, the method further includes: having the verification node or RSU aggregation node of the target node perform anomaly detection on the global model of the vehicle node, obtaining a verification result, and hashing the verification result into the permissioned chain; the verification result includes a statistical risk score; the statistical risk score is obtained based on the CKA anomaly score; In step four, the credibility of the vehicle node includes the statistical credibility of the model update; the statistical credibility of the model update is obtained based on the statistical risk score.

3. The dynamic trusted aggregation method for federated learning in vehicle networking according to claim 1, characterized in that, In step three, the expression for the network-side security score is: in, Indicates network-side security score; Indicates the probability of an attack; This indicates a communication quality score.

4. The dynamic trusted aggregation method for federated learning in vehicle networking according to claim 1, characterized in that, In step four, the mathematical expression for the total credibility score is: in, Indicates the overall credibility score. This indicates that the rating will be limited to Within the interval, As a penalty item, This represents the historical reliability score of the comprehensive vehicle node. Indicates network security score. This represents the result of a zero-knowledge proof. This indicates the statistical confidence level of the model update. Indicating the credibility of historical behavior, The weighting coefficients are and satisfy the following conditions: .

5. The dynamic trusted aggregation method for federated learning in vehicle networking according to claim 1, characterized in that, In step four, the global model is updated based on the total confidence score, and the mathematical expression is: in, This represents the updated global model. Represents the global model; Indicates aggregate weight; This indicates that the model is updating its parameters; Represents vehicle node In the wheel isolation markings, This indicates the vehicle node number participating in the current round of candidate aggregation; This represents the set of vehicle nodes that meet the condition of not being isolated. To express summation; The mathematical expression for the aggregation weight is: in, Indicates aggregate weight, and Representing vehicle nodes respectively With vehicle nodes The number of local samples, and Representing the corresponding vehicle nodes With vehicle nodes In the The overall credibility score of the wheel This represents the summation of all vehicle nodes that meet the condition of not being isolated.

6. A dynamic trusted aggregation system for federated learning in vehicle networks, characterized in that, include: Module 1: Verify the decentralized identity and verifiable credentials of the target node, and generate a participation token; Module 2: Based on the participation token, the vehicle node in the target node receives the global model, and generates the model update parameters and lightweight compliance proof of the global model according to the local data. Then, based on the model update parameters and lightweight compliance proof, an update message of the global model is generated. Module 3: Based on the update message, calculate and obtain the network-side security score of the vehicle node; Determine whether the network-side security score triggers a preset abnormal baseline; If the result is yes, the corresponding node is added to the isolation candidate set; if the result is no, nothing is done. Module 4: Based on the network-side security score and the trustworthiness of the vehicle node, calculate the total trust score. Based on whether the total trust score triggers a preset mandatory isolation condition, if the result is yes, set the aggregation weight of the vehicle node to zero and forcibly isolate the vehicle node; if the result is no, update the global model according to the total trust score.

7. The dynamic trusted aggregation system for federated learning in vehicle networking according to claim 6, characterized in that, Between the operation of trigger module three and the operation of trigger module four, the following steps are also included: causing the verification node or RSU aggregation node of the target node to abnormally detect the global model of the vehicle node, obtaining the verification result, and writing the verification result hash into the permissioned chain; the verification result includes a statistical risk score; the statistical risk score is obtained based on the CKA anomaly score; In step four, the credibility of the vehicle node includes the statistical credibility of the model update; the statistical credibility of the model update is obtained based on the statistical risk score.

8. The dynamic trusted aggregation system for federated learning in vehicle networking according to claim 6, characterized in that, In module three, the expression for the network-side security score is: in, Indicates network-side security score; Indicates the probability of an attack; This indicates a communication quality score.

9. The dynamic trusted aggregation system for federated learning in vehicle networking according to claim 6, characterized in that, In module four, the mathematical expression for the total credibility score is: in, Indicates the overall credibility score. This indicates that the rating will be limited to Within the interval, As a penalty item, This represents the historical reliability score of the comprehensive vehicle node. Indicates network security score. This represents the result of a zero-knowledge proof. This indicates the statistical confidence level of the model update. Indicating the credibility of historical behavior, The weighting coefficients are and satisfy the following conditions: .

10. The dynamic trusted aggregation system for federated learning in vehicle networking according to claim 6, characterized in that, In module four, the global model is updated based on the total credibility score, and the mathematical expression is: in, This represents the updated global model. Represents the global model; Indicates aggregate weight; This indicates that the model is updating its parameters; Represents vehicle node In the wheel isolation markings, This indicates the vehicle node number participating in the current round of candidate aggregation; This represents the set of vehicle nodes that meet the condition of not being isolated. To express summation; The mathematical expression for the aggregation weight is: in, Indicates aggregate weight, and Representing vehicle nodes respectively With vehicle nodes The number of local samples, and Representing the corresponding vehicle nodes With vehicle nodes In the The overall credibility score of the wheel This represents the summation of all vehicle nodes that meet the condition of not being isolated.

Images