A distributed machine learning high-robustness security aggregation method for a vehicle networking scene

Abstract

The application discloses a distributed machine learning high-robustness security aggregation method for a vehicle networking scene, a partially connected communication topology is constructed by adopting a regular graph, and a parallel vector coding algorithm is used to quickly code high-dimensional gradient data by using a multi-core processor of a vehicle, so that the calculation efficiency is significantly improved. In view of an offline vehicle, a gradient recovery mechanism cooperates with online vehicles to reconstruct a protected zero value as a virtual gradient of the offline vehicle, so that the offline vehicle cannot destroy the aggregation process, and the integrity and privacy of the aggregation are ensured. A pseudo-random number generator is used to generate a mask vector to blind the gradient input by the vehicle, so that the roadside unit aggregator cannot obtain sensitive information of a single vehicle. The method provided by the application considers the efficiency and safety, and realizes the privacy protection of the gradient data.

Description

Technical Field

[0001] This invention relates to the fields of vehicle-to-everything (V2X) communication technology and distributed machine learning security technology, specifically to a robust and secure aggregation method for distributed machine learning in V2X scenarios. Background Technology

[0002] As a core component of intelligent transportation systems, the Internet of Vehicles (IoV) provides strong support for achieving autonomous driving, alleviating traffic congestion, and improving road safety through communication between vehicles and between vehicles and infrastructure. With the development of artificial intelligence technology, distributed machine learning is widely used in IoV scenarios because it can achieve distributed model training while protecting data privacy. In distributed machine learning, vehicles train models using local data and only upload model gradients or parameters to the server, rather than the raw data, thus protecting user privacy to a certain extent.

[0003] However, directly applying distributed machine learning to the Internet of Vehicles (IoV) faces significant challenges. First, IoVs are highly dynamic; the high-speed movement of vehicles causes frequent changes in network topology, leading to unstable communication links and a high risk of vehicle disconnection. Vehicle disconnection compromises the data integrity of traditional distributed machine learning aggregation protocols, resulting in model training failures or slow convergence. Second, the model gradients generated by vehicles are typically high-dimensional vectors, requiring substantial computational and communication resources for encryption and transmission, while vehicles have relatively limited computing power and storage space. Furthermore, malicious attackers could attempt to infer sensitive vehicle information from uploaded gradients by eavesdropping on communication links or colluding with aggregators.

[0004] Existing secure aggregation schemes, such as the SecAgg scheme based on pairwise masks, can protect individual gradient privacy, but their communication complexity is... These solutions are difficult to apply to large-scale vehicle-to-everything (V2X) scenarios. While homomorphic encryption-based schemes can achieve privacy protection, they incur huge computational overhead and cannot meet the low-latency requirements of V2X. Threshold-based secret sharing schemes tolerate user disconnection to some extent, but their key distribution and reconstruction processes are complex and also face communication bottlenecks. Furthermore, most existing schemes fail to fully utilize the computing resources of vehicles and roadside units for parallel processing, resulting in low efficiency in processing high-dimensional gradient data.

[0005] For example, Chinese patent application CN112770291B proposes a distributed intrusion detection method and system based on federated learning and trust assessment, which uses a masking mechanism to protect model parameters, but relies on centralized distribution and lacks a systematic fault-tolerant design for offline vehicles; Chinese patent application CN112583575B proposes a privacy protection method based on homomorphic encryption federated learning in vehicle networking, which introduces federated learning based on homomorphic encryption into vehicle networking. Although it improves the Paillier and RSA algorithms to achieve full homomorphic encryption, the computational overhead of homomorphic encryption is still huge, making it difficult to meet the real-time requirements of low-latency communication in vehicle networking.

[0006] Therefore, there is an urgent need for a distributed machine learning security aggregation method that is suitable for the dynamic environment of vehicle networking, has high communication efficiency, low computational overhead, and can effectively tolerate vehicle disconnection. Summary of the Invention

[0007] Purpose of the invention: The purpose of this invention is to address the shortcomings of existing technologies and provide a robust and secure distributed machine learning aggregation method for vehicle networking scenarios.

[0008] Technical solution: The present invention provides a distributed machine learning-based robust and secure aggregation method for vehicle-to-everything (V2X) scenarios, comprising the following steps:

[0009] Step S1: System Initialization Phase

[0010] The Key Generation Center (KGC) generates common system parameters and registers Roadside Units (RSUs) and vehicles; the RSUs register the parameters based on the number of vehicles within their communication range. Generate a - A regular graph is used to determine the set of neighboring nodes for each vehicle; the current vehicle generates a shared key with its neighboring vehicles through a key negotiation mechanism, and calculates its own secret aggregation key based on the shared key;

[0011] The Key Generation Center (KGC) generates common system parameters and registers Roadside Units (RSUs) and vehicles; RSUs register based on the number of vehicles. generate - Regular graphs This involves determining the set of neighbor nodes for each vehicle, used to construct a sparse interaction topology between vehicles during the online aggregation phase. Each vehicle only exchanges blinding shares with its neighbor nodes, thus optimizing the communication complexity for each vehicle. Down to It also supports online vehicle collaboration to generate protection zero-value ciphertexts for offline vehicles to recover from faults;

[0012] The current vehicle communicates with [other entities] through a key negotiation mechanism. - Neighbor nodes in the regular graph establish paired shared keys, which are used to generate paired masks for blinding their own model updates during the online aggregation phase; simultaneously, the vehicle calculates its own secret aggregation key based on the paired shared keys with each neighbor node, and splits it into its secret sharing key using the Shamir secret sharing scheme. Each share is distributed to all registered vehicles, not just neighboring nodes, to ensure that no less than the threshold is reached. Online vehicles can jointly recover the secret information of offline vehicles, thereby achieving fault tolerance; among them, pairwise shared key negotiation with neighboring nodes serves the goal of privacy protection, while the network-wide distribution of secret aggregated key shares serves the goal of system robustness. The two topologies serve different stages respectively.

[0013] The vehicle uses a secret sharing scheme to split the secret aggregation key into a set of shared keys. Each share, and through an authenticated encrypted channel, the proceeds... Each share was distributed to all, including himself. Registered vehicles; vehicles here Share generation algorithm for running secret sharing scheme Finally, its secret aggregation key is generated. of Integer field share ;

[0014] Step S2, Online Aggregation Stage

[0015] In each round of distributed machine learning training, the vehicle uses a pseudo-random number generator (PRG) to generate a blinding mask to encode and blind the local gradient, and uses a threshold variant of the Joye-Libert secure aggregation scheme (TJL) to encrypt the blinded gradient and generate ciphertext.

[0016] Meanwhile, the vehicle generates a random seed using a secret sharing scheme. -out-of- The share, encrypted, is sent to the RSU along with the ciphertext; the RSU collects the ciphertext and share of all online vehicles, identifies the set of offline vehicles, and instructs online vehicles to collaborate in generating a protection zero value for the offline vehicles.

[0017] RSU collects from at least The protection zero value share and random seed share of all online vehicles are used to reconstruct the protection zero value of offline vehicles and restore the blinding mask of all online vehicles; RSU aggregates the ciphertext and protection zero value of all online vehicles, and after removing the blinding mask, obtains the plaintext sum of the gradients of all online vehicles. .

[0018] Furthermore, the detailed method of step 1 is as follows:

[0019] Step 1.1: Key Generation Center (KGC) selects security parameters. The parameter generation algorithm for running the key negotiation scheme KA To generate common parameters Common parameters Key generation for the vehicle is then performed, followed by the initialization algorithm of the TJL security aggregation scheme. Generate common parameters Configure security parameters for secret sharing of integer fields. Set a secret sharing reconstruction threshold. (Equivalent to the minimum number of online vehicles required to recover the secret information of a disconnected vehicle), total number of vehicles Gradient dimension Blinding range Generate system common parameters and will Broadcast to all roadside unit RSUs and vehicles;

[0020] Step 1.2: The RSU collects the registered users within its communication range. Vehicle identification information can be used to run graph generation algorithms that reduce communication complexity and simplify topology structures. Generate a - A regular graph G (distinct from the fully connected topology of existing technologies) identifies each other vehicle. The set of neighboring nodes NG ( ), and ensure the reconstruction threshold RSU will - The regular graph G is broadcast to all online vehicles;

[0021] Step 1.3, Vehicle Based on common parameters The key generation algorithm KA.Gen (which runs the key negotiation algorithm) This generates two key pairs, which are the communication key pairs. and secret shared key pair ;

[0022] vehicle public key and The public keys are sent to the RSU via a secure channel; after the RSU collects all vehicle public keys, it will generate a list containing all vehicle public keys. The announcement was broadcast to all vehicles;

[0023] Step 1.4, Vehicle After receiving the list, for any other vehicle Using one's own private key and the other party's public key Run the key negotiation algorithm to generate the channel key. , used to encrypt the share of the secret key;

[0024] ;

[0025] At the same time, they use their own secret shared private key. Share public key with the other party's secret Run the key negotiation algorithm to generate a shared key. ;

[0026] ;

[0027] Step 1.5, Vehicle According to all other vehicles The generated shared key Calculate your own secret aggregation key :

[0028] ;

[0029] in, A collection of all registered vehicles;

[0030] Each pair of vehicles Define symbolic functions as follows:

[0031] ;

[0032] Among them, the order relationship between vehicles is based on the vehicle's unique identifier (such as a vehicle ID or a hash of a public key). Sort; when At that time, the vehicle Shared key Its secret aggregation key is recorded with a positive sign; when At that time, the vehicle Shared key Its secret aggregation key is recorded with a negative sign;

[0033] Step 1.6, Vehicle Share generation algorithm for running secret sharing scheme Generate its secret aggregation key of Integer field share ,in ;

[0034] For each other vehicle ,vehicle Use the corresponding communication channel key Encryption algorithm that runs authentication encryption algorithm Generate ciphertext ;vehicle All ciphertext collection Send to RSU;

[0035] Step 1.7: The Roadside Unit (RSU) collects the encrypted data sent by all vehicles. And based on the receiver index in the ciphertext , the corresponding ciphertext set Distributed to the corresponding vehicles ;vehicle Upon receipt, use the communication channel key. Decryption algorithm for running authentication encryption algorithm Obtain and store data from vehicles Share of its secret aggregation key .

[0036] Furthermore, the detailed process of step 2 is as follows:

[0037] Step 2.1, in a round of distributed machine learning training Initially, each online vehicle Randomly sample a random seed Run the pseudo-random number generator Expand it to a blinding mask vector with the same dimension as the local gradient vector. ;vehicle Its local gradient vector is processed using the parallel vector encoding algorithm PVE.Encode. Encode to obtain the encoded long integer. ;vehicle Calculate the blinded input It then runs the TJL scheme's encryption algorithm to generate gradient ciphertext. ;

[0038] ;

[0039] Step 2.2, Vehicle Share generation algorithm for running secret sharing scheme Generate its random seed of -out-of- Individual shares ;

[0040] For each other vehicle ,vehicle Use communication channel key Run authentication encryption algorithm Generate ciphertext ;vehicle All share ciphertext and gradient ciphertext Send to RSU;

[0041] Step 2.3: The RSU collects the share ciphertext and gradient ciphertext sent by all vehicles and records the current... The collection of online vehicles is ;

[0042] like If the aggregation fails, the current round of aggregation terminates; otherwise, RSU traverses the set. Each vehicle Share cipher Send to the corresponding online vehicle , ;

[0043] Step 2.4, each online vehicle After receiving the share ciphertext, use the corresponding communication channel key. Decryption algorithm for running authentication encryption algorithm Restore other online vehicles random seed share ;

[0044] Meanwhile, online vehicles According to the set of disconnected vehicle identifiers provided by the RSU (i.e. (This refers to a list consisting of the IDs of every registered vehicle that is not online in this round), utilizing its locally stored information about offline vehicles. Secret aggregate key share Run the protection zero-value generation algorithm of the TJL scheme. Generate a protection zero-value share for all offline vehicles. ;vehicle The random seed share of all recovered online vehicles and its own generated protection zero-value share Send to RSU;

[0045] Step 2.5, RSU from online vehicle collection At least collected The share sent by each vehicle, denoted as the set of these vehicles. , RSU for each online vehicle Reconstructing the secret sharing scheme algorithm Recover the random seed for the vehicle. And run the pseudo-random number generator. Recalculate its blinding mask RSU runs the protection zero-value merging algorithm of the TJL scheme. Restore the protection value to zero for all disconnected vehicles. When the number of honest vehicles exceeds half of the total number of vehicles, privacy and security can be guaranteed, and it can tolerate a maximum of [number missing] vehicles. A vehicle went offline; under the malicious threat model, when the reconstruction threshold... At that time, it can resist malicious attacks and tolerate up to [number of attacks]. One vehicle went offline;

[0046] Step 2.6: RSU runs the aggregation algorithm of the TJL scheme. Obtain the aggregation result after blinding ;

[0047] in, For system aggregation keys, satisfy ;

[0048] RSU calculates the final aggregation result. The expression is:

[0049] ;

[0050] This result is the plaintext sum of the local gradients for all online vehicles.

[0051] The parallel vector encoding algorithm in step 2.1 above This is used to convert high-dimensional floating-point gradient vectors into long integers suitable for the TJL scheme; based on a preset compression ratio... Determine the batch size; utilize parallel computing libraries to process the input vectors. The process is divided into multiple batches; multiple vehicle CPU cores are started to perform encoding operations on each batch in parallel, encoding multiple gradient components within the batch into a large integer; finally, the encoding results of all batches are merged to obtain the final encoded vector.

[0052] Furthermore, step 1.5 involves calculating the secret aggregation key. When the following formula is satisfied: ;

[0053] The above formula indicates that the sum of the secret aggregate keys of all vehicles is zero, thus making the system aggregate key... Or satisfy the requirements in the TJL scheme. The computational requirements.

[0054] Compared with the prior art, the present invention has the following advantages:

[0055] (1) This invention constructs a graph generation algorithm. - Regular graphs separate communication objects with shared masks among all other vehicles. The number of neighbors is reduced to a constant level. This reduces the communication overhead per vehicle to logarithmic levels.

[0056] (2) The present invention designs a parallel vector encoding algorithm, which divides the high-dimensional gradient vector into multiple batches and uses the vehicle's multi-core CPU for parallel encoding, which significantly shortens the data preprocessing time and improves the computational efficiency.

[0057] (3) The present invention can achieve anonymity and non-linkability of the message subject, eliminate the correlation between communication interaction and identity information, prevent attackers from maliciously tracking based on vehicle dynamic trajectory, and achieve traceability of vehicle identity in the event of a dispute.

[0058] This invention employs a dual mechanism to protect vehicle privacy: on the one hand, by utilizing the additive homomorphic property of the secret aggregation key, it ensures that the aggregator can only obtain the sum of all gradients; on the other hand, by using a blinding mask generated by a random seed, it prevents the aggregator from inferring the gradient information of a single vehicle from intermediate results. Attached Figure Description

[0059] Figure 1 This is a flowchart illustrating the implementation of the present invention;

[0060] Figure 2 This is a schematic diagram of the system framework of the present invention;

[0061] Figure 3 This is a flowchart of the secure aggregation protocol in an embodiment of the present invention;

[0062] Figure 4 This is a diagram of the parallel vector algorithm in this invention;

[0063] Figure 5 This is a comparison chart of communication overhead under different numbers of vehicles in an embodiment of the present invention. Detailed Implementation

[0064] The technical solution of the present invention will be described in detail below, but the scope of protection of the present invention is not limited to the embodiments described.

[0065] like Figure 1 and Figure 2 As shown, the distributed machine learning-based robust and secure aggregation method for vehicle-to-everything (V2X) scenarios of the present invention includes the following steps:

[0066] Step S1: System Initialization Phase

[0067] The Key Generation Center (KGC) generates common system parameters and registers Roadside Units (RSUs) and vehicles; the RSUs register the parameters based on the number of vehicles within their communication range. Generate a - A regular graph is used to determine the set of neighboring nodes for each vehicle; the current vehicle generates a shared key with its neighboring vehicles through a key negotiation mechanism, and calculates its own secret aggregation key based on the shared key;

[0068] The vehicle uses a secret sharing scheme to split the secret aggregation key into a set of shared keys. Each share, and through an authenticated encrypted channel, the proceeds... Each share was distributed to all, including himself. Registered vehicles;

[0069] Step S2, Online Aggregation Stage

[0070] In each round of distributed machine learning training, the vehicle uses a pseudo-random number generator (PRG) to generate a blinding mask to encode and blind the local gradient, and uses a threshold variant of the Joye-Libert secure aggregation scheme (TJL) to encrypt the blinded gradient and generate ciphertext.

[0071] Meanwhile, the vehicle generates a random seed using a secret sharing scheme. -out-of- The share, encrypted, is sent to the RSU along with the ciphertext; the RSU collects the ciphertext and share of all online vehicles, identifies the set of offline vehicles, and instructs online vehicles to collaborate in generating a protection zero value for the offline vehicles.

[0072] RSU collects from at least The protection zero value share and random seed share of all online vehicles are used to reconstruct the protection zero value of offline vehicles and restore the blinding mask of all online vehicles; RSU aggregates the ciphertext and protection zero value of all online vehicles, and after removing the blinding mask, obtains the plaintext sum of the gradients of all online vehicles. .

[0073] The detailed method for step 1 in this embodiment is as follows:

[0074] Step 1.1: Key Generation Center (KGC) selects security parameters. Run the key negotiation algorithm to generate common parameters and become common parameters Configure security parameters for secret sharing of integer fields. Set a secret sharing reconstruction threshold. Total number of vehicles Gradient dimension and blinding range Generate system common parameters and will Broadcast to all roadside unit RSUs and vehicles;

[0075] Step 1.2: The RSU collects the registered users within its communication range. Vehicle identification information is generated using a graph generation algorithm. - Regular graph G, determining each other vehicle The set of neighboring nodes NG ( ), and ensure the reconstruction threshold RSU will - The regular graph G is broadcast to all online vehicles;

[0076] Step 1.3, Vehicle Based on common parameters The key negotiation algorithm generates two key pairs, which are the communication key pairs. and secret shared key pair ;

[0077] vehicle public key The public keys are sent to the RSU via a secure channel; after the RSU collects all vehicle public keys, it will generate a list containing all vehicle public keys. The announcement was broadcast to all vehicles;

[0078] Step 1.4, Vehicle After receiving the list, for any other vehicle Using one's own private key and the other party's public key Run the key negotiation algorithm to generate the channel key. , used to encrypt the share of the secret key;

[0079] ;

[0080] At the same time, they use their own secret shared private key. Share public key with the other party's secret Run the key negotiation algorithm to generate a shared key. ;

[0081] ;

[0082] Step 1.5, Vehicle According to all other vehicles The generated shared key Calculate your own secret aggregation key :

[0083] ;

[0084] in, A collection of all registered vehicles;

[0085] Each pair of vehicles Define symbolic functions as follows:

[0086] ;

[0087] Among them, the order relationship between vehicles is sorted based on the vehicle's unique identifier; when At that time, the vehicle Shared key Its secret aggregation key is recorded with a positive sign; when At that time, the vehicle Shared key Its secret aggregation key is recorded with a negative sign;

[0088] Step 1.6, Vehicle The share generation algorithm of the secret sharing scheme generates its secret aggregation key. of Integer field share ,in ;

[0089] For each other vehicle ,vehicle Use the corresponding communication channel key The encryption algorithm that runs the authentication encryption algorithm generates ciphertext. ;vehicle All ciphertext collection Send to RSU;

[0090] Step 1.7: The Roadside Unit (RSU) collects the encrypted data sent by all vehicles. And based on the receiver index in the ciphertext , the corresponding ciphertext set Distributed to the corresponding vehicles ;vehicle Upon receipt, use the communication channel key. The decryption algorithm that runs the authentication encryption algorithm obtains and stores the information from the vehicle. Share of its secret aggregation key .

[0091] like Figure 3 and Figure 4 As shown, the detailed process of step 2 in this embodiment is as follows:

[0092] Step 2.1, in a round of distributed machine learning training Initially, each online vehicle Randomly sample a random seed Run the pseudo-random number generator Expand it to a blinding mask vector with the same dimension as the local gradient vector. ;vehicle Use parallel vector encoding method to local gradient vectors Encode to obtain the encoded long integer. ;vehicle Calculate the blinded input It then runs the TJL scheme's encryption algorithm to generate gradient ciphertext. ;

[0093] ;

[0094] Step 2.2, Vehicle The share generation algorithm of the secret sharing scheme generates its random seed. of -out-of- Individual shares ;

[0095] For each other vehicle ,vehicle Use communication channel key Run the authentication encryption algorithm to generate ciphertext ;vehicle All share ciphertext and gradient ciphertext Send to RSU;

[0096] Step 2.3: The RSU collects the share ciphertext and gradient ciphertext sent by all vehicles and records the current... The collection of online vehicles is ;

[0097] like If the aggregation fails, the current round of aggregation terminates; otherwise, RSU traverses the set. Each vehicle Share cipher Send to the corresponding online vehicle , ;

[0098] Step 2.4, each online vehicle After receiving the share ciphertext, use the corresponding communication channel key. The decryption algorithm that runs the authentication encryption algorithm recovers other online vehicles. random seed share ;

[0099] Meanwhile, online vehicles According to the set of disconnected vehicle identifiers provided by RSU (A list consisting of the IDs of every registered vehicle that is not online in this round), utilizing its locally stored information about offline vehicles. Secret aggregate key share The protection zero-value generation algorithm is run to generate protection zero-value shares for all disconnected vehicles. ;vehicle The random seed share of all recovered online vehicles and its own generated protection zero-value share Send to RSU;

[0100] Step 2.5, RSU from online vehicle collection At least collected The share sent by each vehicle, denoted as the set of these vehicles. , RSU for each online vehicle The secret-sharing scheme reconstruction algorithm was used to recover the random seed of the vehicle. And run the pseudo-random number generator. Recalculate its blinding mask The RSU uses the TJL scheme's protection zero-value merging algorithm to recover the protection zero values ​​for all disconnected vehicles. ;

[0101] Step 2.6: RSU runs the aggregation method of the TJL scheme to obtain the blinded aggregation results. ;

[0102] in, For system aggregation keys, satisfy ;

[0103] RSU calculates the final aggregation result. The expression is:

[0104] ;

[0105] This result is the plaintext sum of the local gradients for all online vehicles.

[0106] To further verify the technical solution of the present invention, the technical solution of the present invention is applied to the following scheme. This embodiment includes multiple vehicle nodes and a roadside unit (RSU), and the specific steps are as follows.

[0107] Step i (System Initialization): KGC selects security parameters Generate common parameters And broadcast. The RSU collects data within its communication range. A vehicle, generate one - A regular graph, and broadcast the graph structure to all vehicles. Each vehicle The key generation algorithm is run to generate a communication key pair and a secret shared key pair, and the public key is sent to the RSU. The RSU collects all public keys and then broadcasts the list of public keys. (Vehicle) After receiving the list, it is compared with each other vehicle. Negotiate and generate channel key and shared keys .vehicle According to the formula Calculate its secret aggregation key. Then, the vehicle... Generate its secret aggregation key Individual shares and using the channel key The encrypted message is sent to the RSU. The RSU then distributes the ciphertext to the corresponding vehicle based on the recipient's index. .vehicle After decryption, obtain and store the shares from all other vehicles. .

[0108] Step ii (Online Aggregation): Assuming that in distributed machine learning... Wheels, online vehicle collection Each online vehicle Sampling random seed Generate a blinding mask And use the parallel vector encoding algorithm PVE.Encode to set its local gradient Encoded as .vehicle Calculate the blinded input The gradient ciphertext was obtained by encrypting it using the TJL scheme. Meanwhile, the vehicle Generate random seed of -out-of- Shares, encrypted together Send to RSU. After collecting all data, RSU determines the set of disconnected vehicles. If the number of online vehicles The RSU distributes the share of online vehicles in encrypted form to each online vehicle. Each online vehicle... Decrypt and recover the random seed shares of other online vehicles, and utilize the locally stored shares of offline vehicles. Generate a protection zero-value share for all offline vehicles. Then these shares are sent back to the RSU. The RSU receives at least After collecting shares from the online vehicles, a random seed for all online vehicles is reconstructed. And recalculate its blinding mask. At the same time, the zero-value protection share was merged to obtain Finally, RSU aggregates all gradient ciphertexts. and protect zero value Subtracting the sum of the blinding masks of all online vehicles, we obtain the plaintext sum of the gradients of all online vehicles in this round. .

[0109] Experiments show that, by building a prototype in a simulation environment, the communication overhead of this invention is significantly lower than that of the traditional SecAgg scheme when the number of vehicles is large. Furthermore, compared with other existing fault-tolerant and safe aggregation schemes, the results show that the technical solution of this invention has significant advantages in both computational efficiency and fault tolerance. Regarding communication efficiency, this invention achieves this by constructing… - Regular graphs separate communication objects with shared masks among all other vehicles. The number of neighbors is reduced to a constant level. This reduces the communication overhead per vehicle to [number], thus lowering the communication overhead per vehicle to [number]. Regarding computational efficiency, the parallel vector encoding algorithm of this invention divides high-dimensional gradient vectors into multiple batches and utilizes the multi-core CPU of the vehicle for parallel encoding, significantly shortening the data preprocessing time. In terms of fault tolerance, this invention combines secret sharing and the TJL threshold cryptography scheme, enabling the remaining online vehicles to reconstruct their contributions through collaboration when some vehicles go offline, without the participation of the offline vehicles, thus ensuring the integrity of the aggregation process. Regarding privacy protection, this invention employs a dual mechanism to protect vehicle privacy: on the one hand, through the additive homomorphic property of the secret aggregation key, it ensures that the aggregator can only obtain the sum of all gradients; on the other hand, through a blinding mask generated by a random seed, it prevents the aggregator from inferring the gradient information of individual vehicles from intermediate results.

[0110] In summary, the present invention utilizes - Regular graphs are used to construct partially connected communication topologies, reducing the original quadratic complexity. Optimized to logarithmic communication complexity Alternatives: By using parallel vector coding technology, the high-dimensional gradient coding process is accelerated using the vehicle's multi-core processor, improving computational efficiency; By combining threshold secret sharing and TJL cryptography, an efficient gradient reconstruction mechanism for disconnected vehicles is designed. Under a malicious threat model, the system can tolerate a vehicle disconnection rate of up to 33% while ensuring privacy.

[0111] As described above, although the invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the invention itself. Various changes in form and detail may be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. A robust and secure distributed machine learning aggregation method for vehicle-to-everything (V2X) scenarios, characterized in that: Includes the following steps: Step S1: System Initialization Phase The Key Generation Center (KGC) generates common system parameters and registers Roadside Units (RSUs) and vehicles; RSUs register based on the number of vehicles. generate - Regular graphs The system determines the set of neighbor nodes for each vehicle, which is used to build a sparse interaction topology between vehicles during the online aggregation phase. Each vehicle only exchanges blinding shares with its neighbor nodes, and supports online vehicles to collaboratively generate protection zero-value ciphertexts for offline vehicles to recover from faults. The current vehicle communicates with [other entities] through a key negotiation mechanism. - Neighboring nodes in the regular graph establish paired shared keys, which are used to generate paired masks to blind their own model updates during the online aggregation phase; simultaneously, the vehicle calculates its own secret aggregation key based on the paired shared keys with each neighboring node, and splits it into its secret sharing components using the Shamir secret sharing scheme. Each share is distributed to all registered vehicles; Among them, pairwise shared key negotiation with neighboring nodes serves the goal of privacy protection, while the network-wide distribution of secret aggregated key shares serves the goal of system robustness. The two topologies serve different stages respectively. Step S2, Online Aggregation Stage In each round of distributed machine learning training, the vehicle uses a pseudo-random number generator (PRG) to generate a blinding mask to encode and blind the local gradient, and uses a threshold variant of the Joye-Libert secure aggregation scheme (TJL) to encrypt the blinded gradient and generate ciphertext. Meanwhile, the vehicle generates a random seed using a secret sharing scheme. -out-of- The share, encrypted, is sent to the RSU along with the ciphertext; the RSU collects the ciphertext and share of all online vehicles, identifies the set of offline vehicles, and instructs online vehicles to collaborate in generating a protection zero value for the offline vehicles. RSU collects from at least The protection zero value share and random seed share of all online vehicles are used to reconstruct the protection zero value of offline vehicles and restore the blinding mask of all online vehicles; RSU aggregates the ciphertext and protection zero value of all online vehicles, and after removing the blinding mask, obtains the plaintext sum of the gradients of all online vehicles. .

2. The distributed machine learning-based robust and secure aggregation method for vehicle-to-everything (V2X) scenarios as described in claim 1, characterized in that: The detailed method for step 1 is as follows: Step 1.1: Key Generation Center (KGC) selects security parameters. Run the key negotiation algorithm to generate common parameters Generate common parameters Configure security parameters for secret sharing of integer fields. Set a secret sharing reconstruction threshold. Total number of vehicles Gradient dimension and blinding range Generate system common parameters and will Broadcast to all roadside unit RSUs and vehicles; Step 1.2: The RSU collects the registered users within its communication range. Vehicle identification information is generated using a graph generation algorithm. - Regular graph G, determining each other vehicle The set of neighboring nodes NG ( ), and ensure the reconstruction threshold ; RSU will - The regular graph G is broadcast to all online vehicles; Step 1.3, Vehicle Based on common parameters The key negotiation method generates two key pairs, which are the communication key pairs. and secret shared key pair ; vehicle public key and The public keys are sent to the RSU via a secure channel; after the RSU collects all vehicle public keys, it will generate a list containing all vehicle public keys. The announcement was broadcast to all vehicles; Step 1.4, Vehicle After receiving the list, for any other vehicle Using one's own private key and the other party's public key Run the key negotiation algorithm to generate the channel key. , used to encrypt the share of the secret key; ； At the same time, they use their own secret shared private key. Share public key with the other party's secret Generate a shared key by running the key negotiation method. ; ； Step 1.5, Vehicle According to all other vehicles The generated shared key Calculate your own secret aggregation key : ； in, A collection of all registered vehicles; Each pair of vehicles Define symbolic functions as follows: ； Among them, the order relationship between vehicles is sorted based on the vehicle's unique identifier; when At that time, the vehicle Shared key Its secret aggregation key is recorded with a positive sign; when At that time, the vehicle Shared key Its secret aggregation key is recorded with a negative sign; Step 1.6, Vehicle The share generation algorithm of the secret sharing scheme generates its secret aggregation key. of Integer field share ,in ; For each other vehicle ,vehicle Use the corresponding communication channel key The encryption method that runs the authentication encryption algorithm generates ciphertext. ;vehicle All ciphertext collection Send to RSU; Step 1.7: The Roadside Unit (RSU) collects the encrypted data sent by all vehicles. And based on the receiver index in the ciphertext , the corresponding ciphertext set Distributed to the corresponding vehicles ;vehicle Upon receipt, use the communication channel key. The decryption method of the authentication encryption method is used to obtain and store the data from the vehicle. Share of its secret aggregation key .

3. The distributed machine learning-based robust and secure aggregation method for vehicle-to-everything (V2X) scenarios according to claim 1, characterized in that: The detailed process of step 2 is as follows: Step 2.1, in a round of distributed machine learning training Initially, each online vehicle Randomly sample a random seed Run the pseudo-random number generator Expand it to a blinding mask vector with the same dimension as the local gradient vector. ;vehicle Use parallel vector encoding method to local gradient vectors Encode to obtain the encoded long integer. ; vehicle Calculate the blinded input It then runs the TJL scheme's encryption algorithm to generate gradient ciphertext. ; ； Step 2.2, Vehicle The share generation algorithm of the secret sharing scheme generates its random seed. of -out-of- Individual shares ; For each other vehicle ,vehicle Use communication channel key Run the authentication encryption algorithm to generate ciphertext ;vehicle All share ciphertext and gradient ciphertext Send to RSU; Step 2.3: The RSU collects the share ciphertext and gradient ciphertext sent by all vehicles and records the current... The collection of online vehicles is ; like If so, then this round of aggregation will terminate; Otherwise, RSU traverses the set. Each vehicle Share cipher Send to the corresponding online vehicle , ; Step 2.4, each online vehicle After receiving the share ciphertext, use the corresponding communication channel key. The decryption method using the authentication encryption algorithm recovers other online vehicles. random seed share ; Meanwhile, online vehicles According to the set of disconnected vehicle identifiers provided by RSU (A list consisting of the IDs of every registered vehicle that is not online in this round), utilizing its locally stored information about offline vehicles. Secret aggregate key share The protection zero-value generation method generates protection zero-value shares for all offline vehicles. ;vehicle The random seed share of all recovered online vehicles and its own generated protection zero value share Send to RSU; Step 2.5, RSU from online vehicle collection At least collected The share sent by each vehicle, denoted as the set of these vehicles. , ; RSU for each online vehicle The secret-sharing scheme reconstruction algorithm was used to recover the random seed of the vehicle. And run the pseudo-random number generator. Recalculate its blinding mask ; The RSU uses the TJL scheme's protection zero-value merging algorithm to recover the protection zero values ​​for all disconnected vehicles. ; Step 2.6: RSU runs the aggregation method of the TJL scheme to obtain the blinded aggregation results. ; in, For system aggregation keys, satisfy ; RSU calculates the final aggregation result. The expression is: 。 4. The distributed machine learning-based robust and secure aggregation method for vehicle-to-everything (V2X) scenarios according to claim 2, characterized in that, Step 1.5 Calculate the secret aggregation key When the following formula is satisfied: ； The above formula indicates that the sum of the secret aggregate keys of all vehicles is zero, making the system aggregate key... .

5. The distributed machine learning-based robust and secure aggregation method for vehicle-to-everything (V2X) scenarios according to claim 3, characterized in that, In step 2.5, when the number of honest vehicles exceeds half of the total number of vehicles, privacy and security can be guaranteed, and the maximum number of vehicles that can be tolerated can be increased. A vehicle went offline; under the malicious threat model, when the reconstruction threshold... At that time, it can resist malicious attacks and tolerate up to [number of attacks]. One vehicle went offline.

Images