Compressed and quantized Internet of Vehicles privacy protection security hierarchical federal learning method

Abstract

The invention relates to a compressed and quantized Internet of Vehicles privacy protection security hierarchical federated learning method, which designs a vehicle end side encryption parameter compression and integer quantization method, a vehicle side completes model training on the basis of local data, and performs flattening, low-rank approximate compression and integer quantization on parameter tensors to obtain a compressed and quantized vehicle end side encryption parameter. And the transmission quantity is obviously reduced under the condition of ensuring the model precision. The quantized parameters are subjected to end-side homomorphic encryption at a vehicle end, and a ciphertext is kept invisible in transmission and aggregation stages. And secondly, a method for setting anonymous authentication batch verification and session key negotiation security encryption between the vehicle end and the fog layer is designed, the concurrent authentication pressure is reduced through a batch verification structure, and the communication security and the access efficiency are improved. And finally, designing a hierarchical ciphertext aggregation updating and Bloom filtering revocation control method, performing homomorphic domain aggregation on ciphertext parameters uploaded by different vehicles by a fog layer, executing one-time decryption by a center end and recovering a floating point model weight, and completing global model updating.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of Internet of Vehicles privacy protection federated learning, and particularly relates to a compressed quantization Internet of Vehicles privacy protection security hierarchical federated learning method. BACKGROUND

[0002] With the rapid development of intelligent transportation and Internet of Vehicles, vehicles gradually evolve into highly intelligent mobile nodes. Through vehicle-mounted sensors, vehicle-mounted computing units, and vehicle-to-vehicle (V2V), vehicle-to-road (V2I), vehicle-to-fog (V2F), vehicle-to-cloud (V2C) communication methods, vehicle terminals realize road state perception, driving assistance, environmental monitoring, and automatic driving decision functions. In this process, a large amount of high-dimensional, multi-modal data is generated and transmitted in real time. Federated learning provides a solution to the problems of data silos and privacy leakage. Vehicle terminals only save data locally and complete model updates locally, and then upload parameter gradients or weights to the server for aggregation. However, in the Internet of Vehicles scenario, vehicle terminals are highly dynamic, have complex access conditions, and have limited computing power, which poses many challenges to traditional federated learning.

[0003] Firstly, the data generated by vehicle-mounted sensors presents significant non-independent and identically distributed characteristics. Differences in road conditions and driving habits in different areas result in extremely uneven label distribution, which further slows down the convergence of the global model or even causes bias. Secondly, the parameter size of deep learning models is large, and direct uploading will result in huge communication overhead, which will easily cause delay and packet loss under high-speed movement and complex channel conditions, seriously affecting real-time performance. Thirdly, although federated learning avoids the leakage of raw data, clear parameter transmission may still be subject to theft or reverse inference. Existing encryption and privacy protection methods often have high computational or communication costs in the Internet of Vehicles scenario, making them difficult to apply. In addition, if vehicle identity authentication is fixed, it will pose a privacy risk. Existing solutions are inefficient in supporting anonymity and large-scale concurrency, and lack flexible and fast revocation mechanisms. Finally, traditional federated learning relies on a central server for model aggregation, which poses a single point of failure and attack risk. The Internet of Vehicles urgently needs a hierarchical aggregation architecture that integrates at the fog node first and then decrypts and fuses by the center to improve robustness and scalability. SUMMARY

[0004] The purpose of the present application is to overcome the shortcomings of the prior art and provide a compressed quantization Internet of Vehicles privacy protection security hierarchical federated learning method that simultaneously realizes volume compression, end-to-end security, and hierarchical aggregation in one parameter upload process, keeps identity controlled and revocable throughout the entire process, relieves bandwidth and latency pressure, prevents single vehicle update leakage, and improves the stability and convergence speed of global training.

[0005] The present application is a compressed quantization Internet of Vehicles privacy protection security hierarchical federated learning method, the steps of which are: S1: In a connected vehicle environment, each vehicle performs local training of a deep learning model based on a locally collected dataset. The vehicle-side dataset is divided into non-independent identically distributed (Non-IID) partitions according to categories to reflect the heterogeneity of in-vehicle data and accurately reflect the label skew of different vehicles, thus allocating data accordingly. Each client follows a Dirichlet distribution: .when When the smaller value is taken, the label distribution of each client is more skewed relative to the overall distribution, satisfying the requirement that... ; S2: To reduce the bandwidth consumption of the vehicle-side model upload without significantly sacrificing model accuracy and to avoid exposing training parameters during transmission, a vehicle-side encrypted parameter compression and integer quantization method is proposed. The vehicle-side compresses and quantizes the locally trained model parameters. First, the trained parameter tensors... Flattened into a matrix ,in Then, singular value decomposition is performed on the matrix to extract its main feature information. By truncating the singular values, the top [values] are selected. By analyzing the singular values ​​and their corresponding singular vectors, a low-rank approximation compressed matrix can be obtained. Furthermore, the compressed parameter matrix is ​​further compressed using integer quantization, further reducing the amount of data transmitted. Quantization bit width and scaling factor It was used to adjust the quantization accuracy. A hierarchical end-to-end homomorphic encryption was designed for the quantized model parameters. The vehicle-side utilizes the quantized model parameters... Perform homomorphic encryption to generate ciphertext. This ensures that the fog layer can aggregate without touching the plaintext, guaranteeing that sensitive information is not exposed during the model upload process; S3: To ensure a verifiable and continuously secure channel for model update transmission between vehicles and fog, a secure encryption method combining anonymous authentication, batch verification, and session key negotiation is proposed. Secure handshake and data confidentiality are addressed in epoch numbering. It operates within a specific timeframe. The vehicle side is marked with long-term identification. and system master key Formation of epochal pseudonyms In the elliptic curve group Generate temporary key pairs Fog node side generates corresponding During a single connection, the vehicle organizes authentication messages. ,in For random points in the session, For timestamps, message hashes are... The authentication mark is recorded as Fog nodes perform consistency checks on individual packets and, when multiple vehicles access the network simultaneously, [they will then...]. By merging the data into a single batch verification path, the verification load and queuing latency under concurrent conditions are significantly reduced. After successful authentication, the vehicle and fog node surround... Derivation of shared key in conjunction with session randomness The session key is then obtained through a key derivation function. The model state is represented as follows after serialization: The initialization vector is denoted as The ciphertext and integrity marker are denoted as follows: and The vehicle side will The data is sent to the fog layer, forming an end-to-end encrypted and tamper-detectable transmission channel. To mitigate the risks associated with replay and clock skew, the time window width, message counter, and random number freshness are explicitly specified in the session metadata. To cope with denial-of-service pressure, fog nodes set thresholds for the number of authentication requests and failure rate per unit time, triggering lightweight rate limiting and a blacklist observation window. This link ensures that identity is not exposed, handshake overhead is controllable, and transmitted content is invisible in the relay fog layer. S4: To successfully complete global model aggregation and achieve scalable node revocation capabilities without leaking the update content of individual vehicle models, a hierarchical encrypted aggregation update and Bloom filter revocation control method is proposed. (The fog layer is related to the data from the first...) encrypted parameters of each vehicle Element-level aggregation of homomorphic fields is performed to obtain aggregated ciphertext. This corresponds to element-level summation in the plaintext space, and the entire process does not touch any individual vehicle plaintext parameters. The fog layer performs fast revocation access control in identity management. To ensure the correct connection of encrypted operators, the fog layer processes batch numbers and epoch numbers in the processing queue. Client count Perform consistency checks with the weight shape metadata and provide a verifiable batch summary for the fog layer at the output. Fog layer receives... Then, a one-time decryption is performed to obtain an integer summation vector. Then based on the number of clients After averaging, we get and with scaling factor Dequantization to restore floating-point weights Refresh the global parameter set To address failed or non-compliant nodes, the system publishes a set of Bloom filters on an epoch-time basis. The fog layer performs the revocation determination by member query before authentication and aggregation. The revocation set supports a scrolling window, displaying the previous epoch. With current epoch Joint cache to reduce the probability of false positives caused by errors when crossing the epoch boundary. Revoked sessions are immediately terminated at the transport layer and form traceable event entries in the security log; S5: After the global model update is completed, the parameters In an independent test set

[0006] Complete prediction and evaluation: , , The above ensures that the fog layer does not hold plaintext, the central end minimizes the number of decryptions, and the vehicle end obtains a convergent consistent global model.

[0007] Further, the specific steps of the vehicle side encryption parameter compression and integer quantization in S2 are:

[0008] S2.1 The vehicle end flattens the local model parameter tensor into a matrix , then singular value decomposition is performed on the matrix : . Wherein, and are orthogonal matrices, is a diagonal matrix containing singular values ; S2.2 Select the first singular values, calculate the low-rank approximation: . Then restore the matrix back to the original tensor shape to get the compressed parameter ; S2.3 Define the quantization bit width , and perform integer quantization according to the scaling factor . The scaling factor is: , wherein . Then, quantize each element according to the following formula: . The quantized parameter matrix is the final ciphertext input transmitted to the fog node; S2.4 Each vehicle end randomly selects a value , and encrypts the parameter using the server public key: . Wherein, is a random number, and the ciphertext is securely uploaded to the fog node for further processing.

[0009] Further, the specific steps of the anonymous authentication batch verification and session key negotiation secure encryption in S3 are: S3.1 Each client Computes its temporary identity , where is the long-term identity, is the master key, is the current epoch number. Then, the client computes the partial private key and randomly selects , thus obtaining the complete private key , and then computes the temporary public key ; S3.2 The vehicle end generates a random number and computes . At the same time, the hash value is computed: and the signature is generated: . The vehicle sends to the fog node; S3.3 After receiving the message, the fog node performs identity verification: . If the verification is passed, the fog node can perform batch anonymous authentication, and the verification of multiple client signatures is combined into one verification: ; S3.4 After the verification is passed, the client and the fog node perform key agreement: , . Finally, the shared key is obtained. This session key is used for subsequent symmetric encryption and decryption operations; S3.5 After the session key is generated, the client encrypts the model using the session key: and uploads the ciphertext. After receiving the ciphertext, the fog node decrypts it using the same session key: . This ensures the security and confidentiality of the transmitted data.

[0010] Further, the specific steps of the hierarchical ciphertext aggregation update and Bloom filter revocation control described in S4 are as follows: S4.1 The fog node receives ciphertexts from all clients and performs aggregation in the ciphertext domain: . This aggregation process ensures that the ciphertext does not expose any sensitive data, and the fog node can only operate within the ciphertext domain; S4.2 The center server receives the aggregated ciphertext and decrypts it: . Then, the server performs dequantization and global model update according to the following formula: . For unquantized parameters, the ordinary average method is used: ; S4.3 The server uses the Bloom filter Periodically publish the identification of revoked vehicles and perform a revocation check before processing updates at the fog node: If a vehicle is revoked, the fog node aborts the session and does not process model updates for that vehicle.

[0011] The compressed quantization privacy protection security hierarchical federated learning method of the Internet of Vehicles of the application, the system first completes the initialization of the elliptic curve parameters and the homomorphic encryption key, and each vehicle generates a temporary identity and a private key to realize anonymous registration. Then each vehicle performs model training based on the non-independent and identically distributed data set locally, and constructs singular value decomposition and truncated reconstruction compression on the obtained model parameters, and then completes quantization according to the scaling factor. Then the vehicle encrypts the quantized parameters using the homomorphic encryption algorithm, and establishes a secure communication channel through anonymous authentication and key negotiation. After the fog node receives the encrypted parameters, it performs homomorphic aggregation, and the central server decrypts the aggregation result to obtain the global model update. Finally, the server restores the aggregated values and refreshes the global model, and then returns the results to each vehicle, and implements fast revocation and management of node identity through the Bloom filter mechanism, forming a continuous and stable training closed loop.

[0012] The application has the following beneficial effects:

[0013] 1) The application realizes efficient compression and lightweight transmission of the model by performing singular value decomposition and truncated reconstruction on the model parameters obtained by training the vehicle terminal, and combining with fixed-point quantization processing. On the premise of ensuring the accuracy of the model, the data size and communication bandwidth occupation of parameter transmission are significantly reduced, and the real-time interaction ability and overall operation efficiency in the Internet of Vehicles environment are improved.

[0014] 2) The application introduces anonymous authentication, temporary key negotiation and homomorphic encryption aggregation technology in the security mechanism, forming an end-to-end encrypted transmission link. The node identity of the vehicle is dynamically generated in a pseudo-anonymous way, effectively preventing the leakage of the real identity; the parameters are always in an encrypted state during transmission and aggregation, which can resist various network attacks such as eavesdropping, tampering and replay, significantly improving the privacy protection ability and data security level of the federated learning process.

[0015] 3) The application constructs a hierarchical aggregation architecture and an extensible Bloom filter revocation mechanism. The fog node is responsible for local aggregation and fast identity verification. The central server only bears global decryption and model update, and only needs to decrypt once to get the global parameter update, which reduces the calculation burden and effectively disperses the aggregation pressure, improves the robustness, scalability and Bloom filter-based revocation mechanism of the system. It realizes the fast elimination of abnormal or failed nodes, and can complete identity revocation detection in constant time. In a large-scale Internet of Vehicles environment, it ensures the continuous stability and reliable execution of the global model training process. BRIEF DESCRIPTION OF DRAWINGS

[0016] Fig. 1 is a specific flowchart of the compression quantization-based vehicle networking privacy protection secure hierarchical federated learning method of the present application, Fig. 2 is a specific implementation of the compression quantization-based vehicle networking privacy protection secure hierarchical federated learning method of the present application, Fig. 3 is a calculation graph of the present application and the comparative method under different node numbers in the present embodiment, Fig. 4 is a communication overhead comparison graph of the present application and the comparative method under different node numbers in the present embodiment, Fig. 5 is an accuracy graph of the present application and the comparative method under different rounds in the present embodiment, Fig. 6 is a loss comparison graph of the present application and the comparative method under different rounds in the present embodiment. DETAILED DESCRIPTION

[0017] The specific embodiments of the present application are described below in conjunction with the accompanying drawings, so that those skilled in the art can better understand the present application. As shown in the drawings: Figs. 1-6

[0018] Non-IID data division of vehicle-mounted data:

[0019] The vehicle client dataset is divided in a non-independent and identically distributed (non-IID) manner to reflect the heterogeneity of vehicle-mounted data and truly reflect the label skew of different vehicles. The data is allocated to clients follow Dirichlet distribution:

[0020] (1)

[0021] When takes a smaller value, the label distribution of each client is more skewed relative to the overall, satisfying .

[0022] Vehicle-end parameter compression and integer quantization:

[0023] The parameter tensor is flattened into a matrix and singular value decomposition is performed, the main singular components are retained, a low-rank approximation is formed, and then restored to the tensor shape. Under the premise of not significantly damaging the accuracy, the data volume to be uploaded is reduced. The parameter tensor obtained by training is flattened into , . The singular value decomposition is as follows:

[0024] (2)

[0025] Select the truncated rank:

[0026] , (3)

[0027] The ​Recover to tensor , so that most of the energy is preserved in the first principal component. The floating-point weight is converted to integer representation by bit width , so that the bandwidth and computing power burden of the subsequent ciphertext aggregation stage is further converged. The scaling factor is:

[0028] (4)

[0029] where . Then element-wise quantization .

[0030] End-to-end encryption and system initialization:

[0031] The integer parameters are homomorphically encrypted before uploading to the end, so that the fog layer can complete the aggregation without touching the plaintext. The client takes a random number , constructs the ciphertext . The system side parameters and keys are set on the elliptic curve , the group and the base point , the master key , and the system public key is constructed. The homomorphic encryption end , , define . Let , the public parameter set contains .

[0032] Anonymous authentication and batch verification:

[0033] The vehicle generates anonymous protection and revocable management within the round . The vehicle constructs Round-based anonymity:

[0034] (5)

[0035] The vehicle uses anonymous to form part of the private key , randomly selects to generate the full private key and public key , . The fog node generates in the same round, and synchronously refreshes with element progression to strengthen forward security.

[0036] In single-session authentication, the vehicle first gives a random point and security parameter:

[0037] , (6)

[0038] Vehicle-side upload message The fog node. The fog node side completes the verification according to the following formula:

[0039] (7)

[0040] In the concurrent scenario, the fog side combines the verification of multiple sessions into a linear combination:

[0041] (8)

[0042] Put The point multiplication relationship is compressed in a verification path, which significantly relieves the verification load when multiple vehicles access.

[0043] Session key negotiation and model encryption:

[0044] The vehicle and the fog node derive a shared key on the temporary quantity of elliptic curve exchange, and then derive a symmetric session key to protect subsequent transmission: the vehicle calculates , and the fog node calculates . It can be obtained .

[0045] Decrypt the model state with the derived session key:

[0046] (8)

[0047] Triplet is securely received and processed in the fog layer.

[0048] Bloom filter revocation:

[0049] To control the invalid alias to continue participating in training, the center publishes the Bloom filter set in each round. The fog layer performs time judgment before accepting updates to support the dynamic entry and exit of a large number of vehicle ends with very low overhead:

[0050] (9)

[0051] Fog layer ciphertext aggregation, center-side one-time decryption and dequantization:

[0052] The fog layer aggregates the ciphertexts from each vehicle element by element:

[0053] (10)

[0054] The center side only performs one-time decryption and value taking on the aggregation result: . Then complete the mean and dequantization, back to the floating-point weight space: .

[0055] Update the obtained ​Complete prediction and evaluation on the test set: , , . The above ensures that the fog layer does not hold plaintext, the center end minimizes the number of decryption times, and the vehicle end obtains a convergent consistent global model.

Claims

1. A compression quantization privacy protection security hierarchical federated learning method for Internet of Vehicles, characterized in that, The steps are: S1: In a vehicle networking environment, each vehicle end performs local training of a deep learning model based on a locally collected data set, and the data set of the vehicle end is divided according to categories in a non-independent and identically distributed (Non-IID) manner to reflect the heterogeneity of vehicle-mounted data and truly reflect the label skew of different vehicles, and the data is allocated to Each client follows a Dirichlet distribution: When takes a smaller value, the label distribution of each client is more skewed relative to the overall, satisfying ; S2: To reduce the bandwidth consumption of vehicle-side model uploads and avoid exposing training parameters during transmission without significantly sacrificing model accuracy, a vehicle-side encrypted parameter compression and integer quantization method is proposed. The vehicle-side compresses and quantizes the locally trained model parameters. First, the trained parameter tensors are... Flattened into a matrix ,in Then, singular value decomposition is performed on the matrix to extract its main feature information. By truncating the singular values, the top [values] are selected. By analyzing the singular values ​​and their corresponding singular vectors, a low-rank approximation compressed matrix can be obtained. Furthermore, the compressed parameter matrix is ​​further compressed using integer quantization, which further reduces the amount of data transmitted and the quantization bit width. and scaling factor Used to adjust quantization accuracy, a hierarchical end-to-end homomorphic encryption was designed for the quantized model parameters. The vehicle side utilizes the quantized model parameters... Perform homomorphic encryption to generate ciphertext. This ensures that the fog layer can aggregate without touching the plaintext, guaranteeing that sensitive information is not exposed during the model upload process; S3: To ensure that the model update transmission process between vehicles and fog has a verifiable and continuously secure channel, a secure encryption method combining anonymous authentication batch verification and session key negotiation is proposed. Secure handshake and data confidentiality are ensured at the epoch numbering stage. Operating within a specific timeframe, the vehicle side features long-term markings. and system master key Formation of epochal pseudonyms In the elliptic curve group Generate temporary key pairs Fog node side generates corresponding During a single connection, the vehicle organizes authentication messages. ,in For random points in the session, For timestamps, message hashes are... The authentication mark is recorded as Fog nodes perform consistency checks on individual packets and, when multiple vehicles access the network simultaneously, [they will then...]. Merging them into a single batch verification path significantly reduces the verification load and queuing latency under concurrent conditions. After authentication, the vehicle and fog node surround... Derivation of shared key in conjunction with session randomness The session key is then obtained through a key derivation function. The model state is represented as follows after serialization: The initialization vector is denoted as The ciphertext and integrity marker are denoted as follows: and The vehicle side will The data is sent to the fog layer, forming an end-to-end encrypted and tamper-detectable transmission channel. To suppress the risks caused by replay and clock skew, the time window width, message counter, and random number freshness are explicitly marked in the session metadata. To cope with denial-of-service pressure, the fog node sets thresholds for the number of authentication requests and failure rate per unit time, triggering lightweight rate limiting and blacklist observation windows. This link ensures that the identity is not exposed, the handshake overhead is controllable, and the transmitted content is not visible in the relay fog layer. S4: To successfully complete global model aggregation and achieve scalable node revocation capabilities without leaking the update content of individual vehicle models, a hierarchical encrypted aggregation update and Bloom filter revocation control method is proposed. The fog layer controls the updates from the first... encrypted parameters of each vehicle Element-level aggregation of homomorphic fields is performed to obtain aggregated ciphertext. This corresponds to element-level summation in the plaintext space. The entire process does not touch any individual vehicle plaintext parameters. The fog layer performs fast revocation access control in identity management. To ensure the correct connection of ciphertext operators, the fog layer processes batch numbers and epoch numbers in the processing queue. Client count The system performs consistency checks with the weight shape metadata and provides a verifiable batch summary to the fog layer at the output. The fog layer receives... Then, a one-time decryption is performed to obtain an integer summation vector. Then based on the number of clients After averaging, we get and with scaling factor Dequantization to restore floating-point weights Refresh the global parameter set To address failed or non-compliant nodes, the system publishes a set of Bloom filters per epoch. The fog layer performs the revocation determination by member query before authentication and aggregation. The revocation set supports a scrolling window, displaying the previous epoch. With the current epoch Joint caching is used to reduce the probability of false positives caused by errors when crossing epoch boundaries. Revoked sessions are immediately terminated at the transport layer and a traceable event entry is formed in the security log. S5: After the global model update is completed, the parameters On the independent test set Complete prediction and evaluation: , , The above ensures that the fog layer does not hold plaintext, the center end minimizes the number of decryption times, and the vehicle end obtains a convergent consistent global model.

2. The compressively quantized V2X privacy-preserving secure federated learning method of claim 1, wherein, The specific steps of the vehicle-side encryption parameter compression and integral quantization in the step S2 are: S2.1 The vehicle end flattens the local model parameter tensor into a matrix where Then the matrix is singular value decomposed: where and are orthogonal matrices, is a diagonal matrix containing the singular values ; S2.2 Selection before of singular values, compute low-rank approximation: , then restore the matrix back to original tensor shape, get compressed parameters ; S2.3 defining quantization bit width integer quantization according to a scaling factor of 1. where Then, each element is quantized as follows: The quantized parameter matrix is the final ciphertext input transmitted to the fog node; S2.4 Each vehicle end randomly selects a value and encrypts the parameters with the server public key: where, is a random number, and the ciphertext is securely uploaded to the fog node for further processing.

3. The compressively quantized V2X privacy-preserving secure federated learning method of claim 1, wherein, The specific steps of the anonymous authentication batch verification and session key negotiation secure encryption in the step S3 are: S3.1 Each client computes its temporary identity where is the long-term identity, is the master key, is the current epoch number, then the client computes the partial private key and randomly selects so that the complete private key is obtained, then the temporary public key is computed; S3.2 Vehicle end generates a random number and calculates while calculating the hash value: and generates the signature: The vehicle sends to the fog node; S3.3 After the fog node receives the message, identity authentication is performed: If the verification is passed, the fog node can perform batch anonymous authentication, and multiple client signature verifications are combined into one verification: ; S3.4 After the verification is passed, the client performs key agreement with the fog node: , The shared key is finally derived This session key is used for subsequent symmetric encryption and decryption operations; S3.5 After the session key is generated, the client encrypts the model using the session key: and uploads the ciphertext. After receiving the ciphertext, the fog node decrypts it using the same session key: , ensuring the security and confidentiality of the transmitted data.

4. The compressively quantized V2X privacy-preserving secure federated learning method of claim 1, wherein, The specific steps of the hierarchical ciphertext aggregation update and Bloom filtering revocation control in the step S4 are: The specific steps of the hierarchical ciphertext aggregation update and Bloom filtering revocation control in the step S4 are: S4.1 The fog node receives ciphertexts from all clients and performs aggregation over the ciphertext domain: This aggregation process guarantees that no sensitive data is exposed by the ciphertexts, and the fog node can only operate over the ciphertext domain; S4.2 The center server receives the aggregated ciphertext and decrypts it: Then, the server performs dequantization and global model update according to the following formula: For unquantized parameters, the ordinary average method is used: ; S4.3 Server performs revocation check via Bloom filter Periodically publish revocations of vehicle identities and perform revocation checks before processing updates at the fog node: If the vehicle is revoked, the fog node aborts the session and does not process model updates for that vehicle.

Images