Multi-participant longitudinal federal learning method based on data alignment and privacy protection

Abstract

The invention relates to a multi-participant longitudinal federated learning method based on data alignment and privacy protection, and belongs to the field of data security. According to the invention, by fusing security calculation of multi-party data, the requirements of data privacy protection and data collaboration are met, and meanwhile, the user is supported to solve the problems of data leakage and privacy invasion; according to the method, a safe multi-party data alignment mechanism is constructed in combination with zero-knowledge proof and homomorphic encryption technologies, data privacy is protected through an efficient encryption means, safe identification of common user data is achieved through multi-party private set intersection calculation, and the safety and efficiency of data collaboration are improved; according to the violation processing and reward and punishment mechanism, through supervision of a third-party review mechanism, it is ensured that the participants comply with protocol regulations in the data alignment and model training process, data abuse is prevented, and the protocol reliability and the compliance of the participants are enhanced.

Description

Technical Field

[0001] The present invention belongs to the field of data security technology and relates to a multi-party vertical federated learning method based on data alignment and privacy protection. Background Art

[0002] With the continuous strengthening of data privacy protection and data security regulations (such as GDPR and CCPA), data holders are facing higher privacy protection requirements. In this context, traditional data sharing and centralized processing methods face challenges and cannot be directly applied to data collaboration in privacy-sensitive fields such as finance, medical care, and telecommunications. Federated Learning (FL), as a distributed machine learning framework, allows data holders to jointly train models without sharing original data, thereby achieving privacy protection for data collaboration. In federated learning, Vertical Federated Learning (VFL) is particularly suitable for scenarios where different industry institutions have different dimensions of feature data for the same user group. Through VFL, participants from different industries can build collaborative models based on different data features of the same user to improve business accuracy, such as user credit scoring or personalized recommendations. VFL is generally divided into two steps: the first step: sample alignment; the second step: aligned samples for model encryption training. However, a key issue facing VFL is how to align data samples while ensuring privacy to identify users commonly owned by participants (i.e., intersection users) and achieve data collaboration.

[0003] For example, Liu et al. introduced the asymmetric private set intersection protocol (APSI) and the "real and virtual" method in "Asymmetrical Vertical Federated Learning" [in arXiv preprint arXiv:2004.07427, 2020.] to ensure the protection of the ID privacy of weak participants during sample ID alignment and model training, while allowing strong participants to perform effective model training without obtaining the specific sample IDs of weak participants. However, this scheme is only for protecting the privacy of weak participants and does not consider the impact of malicious behavior of weak participants on training. At the same time, this scheme only realizes data alignment between two participants and fails to expand to multiple participants. Zhou et al. proposed a systematic privacy-preserving federated learning framework (PFLF) in "Privacy-Preserving Federated Learning Framework with General Aggregation and Multiparty Entity Matching" [in Wireless Communications and Mobile Computing, 2021, 2021(1):6692061.], which proposed a multi-party entity matching protocol to solve the entity alignment problem through secure multi-party computation. Although the framework achieves multi-scenario data aggregation and multi-party matching while protecting data privacy, it still has some limitations in practical applications. The framework assumes that all participants are honest but curious, and does not take into account the attack behavior of malicious participants.

[0004] Existing technologies usually use Private Set Intersection (PSI) technology to achieve data sample alignment, and use encryption to ensure that each party can only obtain the set information of the intersection users without leaking the information of non-intersection users. PSI technology performs well in two-party scenarios, but in multi-party scenarios, the traditional two-party PSI protocol has many shortcomings. In vertical federated learning, if the participants forge a false ID set for private intersection alignment or use an incomplete user set for intersection calculation, it may cause security threats. The following are the shortcomings of existing solutions:

[0005] (1) Insufficient scalability: Most of the PSI protocols used by existing VFLs only support two-party intersection alignment calculations and cannot be directly extended to multi-party scenarios. In multi-party data collaboration, if the pairwise intersection method is used, the computational complexity will increase exponentially with the increase in the number of participants. At the same time, the integration of pairwise intersection results also faces many challenges, and data inconsistency and error accumulation are prone to occur, which seriously affects the efficiency and accuracy of multi-party data collaboration and is difficult to meet the needs of large-scale multi-party data alignment in practical applications.

[0006] (2) Privacy leakage risk: Forging a false ID set may lead to attribute inference attacks, where the attacker may infer sensitive information about the training data by analyzing the output or gradient of the model. This attack is particularly noteworthy in vertical federated learning, because the user's original data does not leave the local area, but private information may still be leaked through the output of the model. Using incomplete intersection result data for training may result in incomplete data for model training, affecting the accuracy and generalization ability of the model. In addition, if the participants deliberately provide false data, it may pollute the model and affect the model's prediction results, thereby reducing the reliability and effectiveness of the model. Forged or incomplete data may cause the model to fail to learn effectively, thereby affecting the performance of the model in practical applications. The model may not be able to accurately capture patterns and trends in the data, resulting in reduced prediction accuracy.

[0007] Therefore, an improved multi-party private set intersection protocol is needed that can achieve efficient intersection calculation with the participation of multiple parties while ensuring data privacy protection and compliance supervision to meet the requirements of vertical federated learning for data alignment and privacy protection. Summary of the invention

[0008] In view of this, the purpose of the present invention is to provide a multi-party vertical federated learning method based on data alignment and privacy protection, based on the combination of privacy set intersection, zero-knowledge proof, homomorphic encryption and third-party review technology, so as to solve the privacy protection and compliance problems during multi-party data collaboration in the vertical federated learning environment, and provide an efficient and secure data sharing solution. The present invention introduces a third-party audit agency (Third-Party Auditor, TPA) to supervise and verify the key steps of intersection calculation and federated learning training, ensure that all parties follow the protocol during data alignment and training, and avoid falsifying data or improper use of data. TPA uses zero-knowledge proof and homomorphic encryption to verify the authenticity and integrity of data, effectively preventing possible privacy leaks and unfair competition.

[0009] In order to achieve the above object, the present invention provides the following technical solutions:

[0010] A multi-party vertical federated learning method based on data alignment and privacy protection includes the following steps:

[0011] S1: Protocol initialization phase: Initialize the protocol, and the trusted institution TA registers and signs the ID;

[0012] S2: User ID verification stage: Each participant uses zero-knowledge proof to prove the legitimacy of its user ID;

[0013] S3: Multi-Party Private Set Intersection (MPSI) stage: The third-party review agency TPA uses the verified user ID hash set to perform intersection calculation to ensure that the intersection set only contains the real user IDs of each party;

[0014] S4: Encryption verification and homomorphic encryption check phase: TPA ensures that all parties only train on the complete intersection user set S through encryption verification and dynamic homomorphic encryption check;

[0015] S5: Violation handling and reward and punishment mechanism stage: When violations are detected by the participants, TPA handles them according to the predetermined reward and punishment mechanism.

[0016] Further, the protocol initialization in step S1 includes:

[0017] The protocol has P1,...,P t Each party holds a set of private data, S1, S2, ..., S t , each set contains n i elements, the goal of the protocol is to obtain the intersection through secure calculation: Ensure that each party can only train on the intersection users; among the participants, P1 to P t-1 is called client, P t On the server side, TPA is responsible for the whole process supervision as the review agency, and TA is the trusted agency to register and sign the ID;

[0018] The trusted institution TA registers and signs the ID, specifically including:

[0019] Each participant P i Set its user ID Send it to TA, TA verifies each ID received, and for the ID that passes the verification, TA uses an asymmetric encryption algorithm and uses TA's private key to sign the ID-related information for the participant P i Issue a certificate Cert(ID) containing the signature and TA public key information, and send it back to the corresponding participant P i save.

[0020] Further, step S2 specifically includes the following steps:

[0021] S21: Generate user ID hash: First, the certificate Cert (ID i ) and the corresponding user ID set ID i Bind; then perform user ID hash calculation together with the binding result and the user ID set. Each participant P i Set its user ID Perform hash conversion to obtain a hash set:

[0022]

[0023] Where h is a one-way hash function, which ensures that TPA cannot restore the original value of the user ID;

[0024] S22: Generate zero-knowledge proof: Party P i Generate a zero-knowledge proof π for each hash value i,j , so that:

[0025] π j =ZKP(h(ID j ),AC)

[0026] Among them, AC represents the attributes of the ID that meet the authenticity requirements;

[0027] Adopt the zero-knowledge proof process based on the Schnorr protocol: Assume that the hash value h(ID i,j )There is a secret x j The signing key from the user ID satisfies the following relationship:

[0028]

[0029] Where g is the generator and p is a large prime number;

[0030] The prover wishes to prove that he knows x j , so that Established, but not directly disclosed x j ; Combined with the certificate Cert(ID i ) to ensure that the entire certification process can reflect the authenticity of the ID and the validity of the certificate;

[0031] S23: Participant P i The hash set H i , the corresponding zero-knowledge proof {π i,j} and the bound certificate and ID set are submitted to TPA;

[0032] S24: Verify zero-knowledge proof: TPA receives each participant P i Submitted hash set H i , the corresponding zero-knowledge proof set {π i,j} and the bound certificate and ID set, and perform the following verification steps:

[0033] First, obtain the public key of the trusted organization from the certificate and verify the certificate Cert (ID i )’s signature is valid, check whether the certificate is within the validity period, and whether the ID information in the certificate is consistent with the submitted ID set; if the certificate verification fails, the ID set submitted by the participant is directly judged to be invalid and marked as a violation; if the certificate verification passes, the zero-knowledge proof is verified. For each h(ID i,j )∈H i , TPA uses the verification algorithm Verify to check the corresponding proof π i,j Is it valid: Verify(h(ID i,j ),π i,j )=True, if the verification is successful, h(ID i,j ) is recorded as a valid user ID; otherwise, it is marked as invalid;

[0034] S25: Record the hash set of legal user IDs: After verification is completed, TPA records the hash set of legal user IDs of each participant.

[0035] Further, step S3 specifically includes the following steps:

[0036] S31: MPSI input and encrypted inverse Bloom filter are generated and sent to the server P t ;

[0037] S32: MPSI calculation: The server first calculates the hash value for each element in the user ID set and searches for matching items in the client's encrypted inverse Bloom filter; the server calculates the sum of the ciphertexts and re-randomizes them; the server distributes the encryption results to the selected clients, and the clients decrypt and feed back to the server; the server summarizes the decryption results to form an intersection set, which is submitted to TPA for confirmation and then used for federated training.

[0038] Further, the input of the MPSI and the generation of the encrypted inverse Bloom filter in step S31 include the following steps:

[0039] S311: Input: Each client P1-P t-1 Holds a shared decryption key sk i , the key belongs to an additive homomorphic threshold encryption scheme with zero decryption algorithm ShDec0; the public key pk is visible to all parties;

[0040] S312: Initialization: Server P t Randomly select a set of hash functions And send it to the client;

[0041] S313: Local encrypted inverse Bloom filter generation: Each client performs the following steps to generate an encrypted inverse Bloom filter, thereby achieving privacy-preserving user ID encryption: Calculate its user ID set S i Bloom filter BF i , Get the inverse Bloom filter IBF of the Bloom filter i , IBF i Encrypt each element in the , get the encrypted inverse Bloom filter and send it to the server P t .

[0042] Further, in step S32, the server P t Perform the following steps to calculate the intersection user set:

[0043] The set of user IDs S held by the server t Each element y in j , calculate k hash values ​​h d (y j ), and for each hash value h d (y j ) Inverse Bloom Filter IBF on each client i Find the value in:

[0044]

[0045] Compute the sum of the ciphertexts and re-randomize:

[0046]

[0047] The server will result Any distributed to t-1 clients clients and request them to j Decryption is performed;

[0048] Compliance check: Each client P i Calculate its decryption share sh i,j And feed it back to the server, as follows:

[0049] (1) The client first j After randomization, Sent back to the server, the server combines it into And send it to each client;

[0050] (2) The server summarizes the decrypted shares sh fed back by the client i,j , and finally form the decryption share set sh 1,j ,...,sh t-1,j ;

[0051] TPA verifies the decryption process: TPA monitors the client's decryption feedback process, ensures that all feedback is sent through a secure channel, and prevents the server or client from interfering with the calculation results by manipulating feedback data; TPA randomly checks the client's feedback decryption share to ensure compliance with the decryption process;

[0052] (3) The server uses the combined algorithm to decrypt. For each j∈{1,...,n t}, the server uses a combined algorithm to calculate: Dec(e j )←Comb(sh 1,j ,...,sh t-1,j ); if the decryption result is zero, Dec(e j )=0, then y j In the set of user IDs of all participants, the server will j Add to the intersection set S = {y j}∪S; otherwise, discard y j ;

[0053] The final intersection user set output: server P t Finally, the intersection set S is output and the result is submitted to TPA for final confirmation; after TPA verifies the intersection result, it sends the compliant intersection set to all parties.

[0054] Further, step S4 specifically includes the following steps:

[0055] S41: Encrypted Verification Mechanism: Before each round of training begins, TPA requires all parties to submit an encrypted hash set S of user IDs used for training to verify the compliance of the training data; each party generates an encrypted hash set, and TPA verifies whether these sets are consistent with the known complete intersection set; if the verification passes, it indicates that the parties are training on the correct intersection users; if it fails, TPA will record the violation;

[0056] S42: Dynamic homomorphic encryption check: TPA uses dynamic homomorphic encryption check technology to prevent participants from using non-intersection user data in some rounds; first, randomly generate checkpoints: TPA generates checkpoints at random rounds during the training process, requiring each participant to resubmit the encrypted training data set Then homomorphic encryption verification: TPA uses homomorphic encryption technology at the checkpoint to verify Is it consistent with E(S) to ensure that each round of training data comes from the complete intersection?

[0057] S43: Training phase: After identifying the common entities, the parties use the data of these common entities to collaboratively train a machine learning model.

[0058] Further, the training phase described in step S43 includes the following steps:

[0059] S431: The trusted coordinator creates a key pair and sends the public key to the client and the server;

[0060] S432: The client and the server encrypt and exchange the intermediate results, which are used to help calculate the gradient and loss value;

[0061] S433: The client and the server calculate the encryption gradient and add the additional mask respectively, the server calculates the encryption loss, and the client and the server send the encryption result to the trusted coordinator;

[0062] S434: The trusted coordinator decrypts the gradient and loss information and sends the results back to the client and server; the client and server remove the mask on the gradient information and update the model parameters based on the gradient information.

[0063] Further, step S5 specifically includes the following steps:

[0064] Violation detection: TPA uses zero-knowledge proof and cryptographic verification technology at each key step to regularly check the compliance of all parties;

[0065] Penalty measures: TPA will take the following measures against participants who are confirmed to have violated the regulations:

[0066] (1) Model weight reduction: reduce its contribution weight in the federated model;

[0067] (2) Financial penalties: Financial penalties are imposed on the violating party;

[0068] (3) Cooperation restrictions: In the event of serious violations, the eligibility to participate in federated learning will be suspended;

[0069] Incentives: For compliant parties, TPA issues compliance certificates and gives appropriate rewards to encourage all parties to follow the agreement.

[0070] The beneficial effects of the present invention are as follows: the present invention realizes the authenticity verification and secure sharing of data under the premise of ensuring the privacy of the participants; at the same time, the signature of the trusted certification agency and the real-time supervision of the third-party review agency ensure the fairness and compliance of the data alignment and federated training process, and effectively prevent the behavior of falsifying data, thereby improving the efficiency and reliability of cross-industry multi-party collaboration, enhancing user trust and meeting legal compliance requirements.

[0071] (1) This paper improves the traditional two-party PSI protocol and proposes a privacy intersection alignment method that can be directly extended to multi-party scenarios, effectively solving the problem of insufficient scalability of existing technologies in multi-party data collaboration. This method reduces the computational complexity that grows exponentially with the increase in the number of participants by introducing multi-party private set intersection calculations, and ensures the consistency and accuracy of the data, meeting the needs of large-scale multi-party data alignment in practical applications.

[0072] (2) The present invention introduces strict privacy protection measures in the data alignment stage and the federated training stage to ensure that all participants use real IDs for intersection alignment and only use complete intersection user data during the training process. Through user ID verification and encryption verification mechanisms, the present invention effectively prevents attribute inference attacks and privacy leakage risks, while ensuring the data integrity and authenticity of model training, improving the accuracy, generalization ability and reliability of the model, thereby enhancing the security of data use and the performance of the model in vertical federated learning.

[0073] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0075] Figure 1 This is a block diagram of the vertical federated learning system described in the present invention;

[0076] Figure 2 This is a flow chart of the intersection of multiple privacy sets described in the present invention. DETAILED DESCRIPTION

[0077] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0078] It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and thus the drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0079] In the following description, numerous details are discussed to provide a more thorough explanation of the embodiments of the present invention. However, it is obvious to those skilled in the art that the embodiments of the present invention can be implemented without these specific details. In other embodiments, well-known structures and devices are shown in the form of block diagrams rather than in detail to avoid making the embodiments of the present invention difficult to understand.

[0080] See also Figure 1-2 ,This paper proposes a data alignment method for vertical federated learning ,based on the intersection of multiple privacy sets, which is suitable for data alignment and privacy protection in ,vertical federated learning environments.,By introducing a trusted third-party review agency and combining ,techniques such as zero-knowledge proof and homomorphic encryption, this method achieves the ,legality verification of user IDs, compliance supervision of intersection calculations, and security assurance of joint training ,while ensuring the data privacy of all participating parties, ,thereby ensuring the fairness, security and privacy protection of the data ,collaboration process.

[0081] like Figure 1 This is a block diagram of the vertical federated learning system described in the present invention, which includes the ID sample alignment in the first stage and the model training in the second stage.

[0082] like Figure 2 This is the multi-party privacy set intersection flow chart of the present invention, in which a secure multi-party ID set intersection operation is performed between the client and the server.

[0083] In order to achieve the above object, the present invention provides the following technical solutions:

[0084] S1: Protocol initialization phase:

[0085] Assume that there are P1,…,P t Each party holds a set of private data, S1, S2, ..., S t , each set contains n i elements, the goal of the protocol is to obtain the intersection through secure calculation: Ensure that each party can only train on the intersection of users. t-1 is called client, P t On the server side, TPA acts as a review agency responsible for the entire supervision process, and TA is a trusted agency that registers and signs the ID.

[0086] ID registration: Each participant P i Set its user ID It is sent to TA, and TA conducts detailed verification of each ID received. The verification content includes but is not limited to checking whether the format of the ID complies with the regulations, whether there are duplicate IDs, whether there is a conflict with other registered IDs, etc. The verification method can adopt technical means such as database query and comparison with a known legal ID library. For the ID that passes the verification, TA issues a certificate Cert(ID) containing a signature. The signing process uses an asymmetric encryption algorithm, and TA uses its private key to sign ID-related information (such as the ID itself, issuance time, validity period, etc.). The certificate should also contain TA's public key information for subsequent verification of the signature. TA sends the issued certificate Cert(ID) back to the corresponding participant P i . Participant P i Keep the certificate in a safe place for use in the subsequent user ID verification stage.

[0087] S2: User ID verification stage:

[0088] Before the multi-party private set intersection calculation stage, TPA first verifies the authenticity of the user IDs submitted by each party to prevent any party from obtaining more data by forging a user ID. At this stage, each party uses zero-knowledge proof to prove the legitimacy of its user ID.

[0089] First, generate a hash: Before calculating the user ID hash, the participant first generates the certificate Cert (ID i ) and the corresponding user ID set ID i Binding can be done by using technologies such as encrypted hash chain or digital envelope to ensure the integrity and relevance of the certificate and ID set. The binding result is then processed together with the user ID set. Next is the user ID hash calculation. Each participant P i Set its user ID Perform hash conversion to obtain a hash set: Where h is a one-way hash function, which ensures that TPA cannot restore the original value of the user ID.

[0090] Then generate a zero-knowledge proof: To prove H i Each h(ID i,j ) is a valid user ID hash value and the corresponding ID has a valid certificate issued by a trusted organization. i Generate a zero-knowledge proof π for each hash value i,j , so that: π j =ZKP(h(ID j),AC). Among them, AC indicates that the ID meets the authenticity requirements. The ID is issued by a trusted certification authority and contains a signature to prevent ID forgery. To achieve zero-knowledge proof generation, we use a zero-knowledge proof process based on the Schnorr protocol: Assume that the hash value h(ID i,j )There is a secret x j The signing key from the user ID satisfies the following relationship: Where g is the generator and p is a large prime number, which ensures the difficulty of the discrete logarithm problem. The prover wants to prove that he knows x j , so that Established, but not directly disclosed x j . It is also necessary to combine the certificate Cert (ID i ), including the hash value and signature of the certificate, to ensure that the entire certification process can reflect the authenticity of the ID and the validity of the certificate. i The hash set H i , the corresponding zero-knowledge proof {π i,j} and the bound certificate and ID set are submitted to TPA.

[0091] Finally, TPA verifies the zero-knowledge proof: TPA receives each participant P i Submitted hash set H i , the corresponding zero-knowledge proof set {π i,j} and the bound certificate and ID set, and perform the following verification steps: First, use the public key of the trusted authority (obtained from the certificate) to verify the certificate Cert (ID i )’s signature is valid, check whether the certificate is within the validity period, and whether the ID information in the certificate is consistent with the submitted ID set. If the certificate verification fails, the ID set submitted by the participant is directly judged to be invalid and marked as a violation; if the certificate verification passes, the original zero-knowledge proof verification algorithm is continued. Then verify the zero-knowledge proof, for each h(ID i,j )∈H i , TPA uses the verification algorithm Verify to check the corresponding proof π i,j Is it valid: Verify(h(ID i,j ),π i,j )=True, if the verification is successful, h(ID i,j ) is recorded as a valid user ID; otherwise, it is marked as invalid. After recording the valid user ID hash set, TPA records the valid user ID hash set of each participant as: verified,i ={h(ID i,j )|Verify(h(ID i,j ),πi,j )=True}, this set will be used for subsequent intersection calculations.

[0092] S3: Multi-Party Private Set Intersection (MPSI) calculation phase:

[0093] After the user ID verification is completed, TPA uses the verified user ID hash set to perform intersection calculation to ensure that the intersection set only contains the real user IDs of each party, such as Figure 2 shown.

[0094] First, the input of MPSI is generated with the encrypted inverse Bloom filter: 1. Input: Each client P i (where 1≤i≤t-1) holds a shared decryption key sk i , the key belongs to an additive homomorphic threshold encryption scheme with a zero decryption algorithm (ShDec0). The public key pk is visible to all parties. 2. Initialization: Server P t Randomly select a set of hash functions And send it to the client P i , where 1≤i≤t-1. 3. Local encrypted inverse Bloom filter generation: Each client P i Perform the following steps to generate an encrypted inverse Bloom filter to achieve privacy-preserving user ID encryption: Calculate the user ID set S i Bloom filter BF i , Get the inverse Bloom filter IBF of the Bloom filter i , IBF i Encrypt each element in the , get the encrypted inverse Bloom filter and send it to the server P t .

[0095] Then perform MPSI calculation: Server-side P t Perform the following steps to calculate the intersection user set: For the user ID set S held by the server t Each element y in j , calculate k hash values ​​h d (y j ), and for each hash value h d (y j ) Inverse Bloom Filter IBF on each client i Find the value in: And j∈{1,...,n t}. Compute the sum of the ciphertexts and re-randomize to ensure privacy: The re-randomization operation is designed to break the determinism of homomorphic encryption and avoid information leakage. Any distributed to t-1 clients clients and request them to j To decrypt, usually

[0096] Compliance check: Each client P i Calculate its decryption share sh i,j And feed it back to the server. The details are as follows:

[0097] (1) The client first j After randomization, Sent back to the server, the server combines it into And send it to each client.

[0098] (2) The server summarizes the decrypted shares sh fed back by the client i,j , and finally form the decryption share set sh 1,j ,...,sh t-1,j .

[0099] TPA verifies the decryption process: TPA monitors the client's decryption feedback process, ensures that all feedback is sent through a secure channel, and prevents the server or client from interfering with the calculation results by manipulating the feedback data. TPA can randomly check the client's feedback decryption share to ensure compliance with the decryption process.

[0100] (3) The server uses the combined algorithm to decrypt. For each j∈{1,...,n t}, the server uses a combined algorithm to calculate: Dec(e j )←Comb(sh 1,j ,...,sh t-1,j ). If the decryption result is zero (ie Dec(e j )=0), then it means y j In the set of user IDs of all participants, the server will j Add to the intersection set S = {y j}∪S. Otherwise, discard y j .

[0101] The final intersection user set output: server P t Finally, the intersection set S is output and the result is submitted to TPA for final confirmation. After TPA verifies the intersection result, it sends the compliant intersection set to all parties for subsequent federated training.

[0102] S4: Encryption verification and homomorphic encryption check during the federated training phase:

[0103] During the federated training phase, TPA ensures that all parties only train on the complete intersection user set S through encryption verification and dynamic homomorphic encryption checks, preventing training on a small set of intersection or using non-intersection users.

[0104] First, the encryption verification mechanism: Before each round of training begins, TPA requires all parties to submit an encrypted hash set S of the user IDs used for training to verify the compliance of the training data. Encrypted hash generation: The participants submit their training user set for the rth round. Each user ID is hashed and encrypted to generate a set: Where E(·) represents the encryption operation.

[0105] Then TPA's intersection consistency verification: TPA holds the encrypted hash set E(S) of the complete intersection user set S. At the beginning of each round of training, TPA verifies: If the verification passes, it means that the participant is training on the complete intersection users; otherwise, TPA records the violation.

[0106] Next is the dynamic homomorphic encryption check: TPA uses dynamic homomorphic encryption check technology to prevent participants from using non-intersecting user data in some rounds. Randomly generate checkpoints: TPA generates checkpoints at random rounds during the training process, requiring each participant to resubmit the encrypted training data set Homomorphic encryption verification: TPA uses homomorphic encryption technology at the checkpoint to verify Is it consistent with E(S) to ensure that each round of training data comes from the complete intersection.

[0107] Finally, there is the training phase. After determining the common entities, the parties can use the data of these common entities to collaboratively train a machine learning model. The training process can be divided into the following four steps: 1. The trusted coordinator creates a key pair and sends the public key to the client and server; 2. The client and server encrypt and exchange the intermediate results, which are used to help calculate the gradient and loss values; 3. The client and server calculate the encrypted gradients and add additional masks respectively. The server also calculates the encrypted loss. The client and server send the encrypted results to the trusted coordinator; 4. The trusted coordinator decrypts the gradient and loss information and sends the results back to the client and server. The client and server remove the mask on the gradient information and update the model parameters based on this gradient information.

[0108] S5: Violation handling and reward and punishment mechanism:

[0109] In this solution, TPA is responsible for supervising the behavior of each participant to ensure compliance with the protocol and the authenticity of the data. When a participant is detected to have violated the rules (such as forging a user ID or improperly using intersection data), TPA will handle it according to the predetermined reward and punishment mechanism:

[0110] First, violation detection. TPA uses zero-knowledge proof and encryption verification technology in each key step to regularly check the compliance of all parties.

[0111] Then, for the participants who are confirmed to have violated the regulations, TPA will take the following measures: Model weight reduction: reduce the weight of their contribution in the federated model. Economic fines: impose economic penalties on the violating parties. Cooperation restrictions: in the case of serious violations, suspend their participation in federated learning.

[0112] Finally, as for incentives, TPA will issue compliance certificates and provide appropriate rewards to compliant parties to encourage all parties to follow the agreement.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A multi-party vertical federated learning method based on data alignment and privacy protection, characterized by: The following steps are involved: S1: Protocol initialization phase: Initialize the protocol, and the trusted institution TA registers and signs the ID; S2: User ID verification stage: Each participant uses zero-knowledge proof to prove the legitimacy of its user ID; S3: Multi-party private set intersection calculation stage: The third-party review agency TPA uses the verified user ID hash set to perform intersection calculation to ensure that the intersection set only contains the real user IDs of each party; S4: Encryption verification and homomorphic encryption check phase: TPA ensures that all parties only train on the complete intersection user set S through encryption verification and dynamic homomorphic encryption check; S5: Violation handling and reward and punishment mechanism stage: When violations are detected by the participants, TPA handles them according to the predetermined reward and punishment mechanism.

2. The multi-party vertical federated learning method based on data alignment and privacy protection according to claim 1, characterized in that: The protocol initialization in step S1 includes: The protocol has P1,...,P t Each party holds a set of private data, S1, S2, ..., S t , each set contains n i elements, the goal of the protocol is to obtain the intersection through secure calculation: Ensure that each party can only train on the intersection users; among the participants, P1 to P t-1 is called client, P t On the server side, TPA is responsible for the whole process supervision as the review agency, and TA is the trusted agency to register and sign the ID; The trusted institution TA registers and signs the ID, specifically including: Each participant P i Set its user ID Send it to TA, TA verifies each ID received, and for the ID that passes the verification, TA uses an asymmetric encryption algorithm and uses TA's private key to sign the ID-related information for the participant P i Issue a certificate Cert(ID) containing the signature and TA public key information, and send it back to the corresponding participant P i save.

3. The multi-party vertical federated learning method based on data alignment and privacy protection according to claim 1, characterized in that: Step S2 specifically includes the following steps: S21: Generate user ID hash: First, the certificate Cert (ID i ) and the corresponding user ID set ID i Bind; then perform user ID hash calculation together with the binding result and the user ID set. Each participant P i Set its user ID Perform hash conversion to obtain a hash set: Where h is a one-way hash function, which ensures that TPA cannot restore the original value of the user ID; S22: Generate zero-knowledge proof: Party P i Generate a zero-knowledge proof π for each hash value i,j , so that: π j =ZKP(h(ID j ),AC) Among them, AC represents the attributes of the ID that meet the authenticity requirements; Adopt the zero-knowledge proof process based on the Schnorr protocol: Assume that the hash value h(ID i,j )There is a secret x j The signing key from the user ID satisfies the following relationship: h j =g xj modp Where g is the generator and p is a large prime number; The prover wishes to prove that he knows x j , so that h j =g xj modp is established, but does not directly disclose x j ; Combined with the certificate Cert(ID i ) to ensure that the entire certification process can reflect the authenticity of the ID and the validity of the certificate; S23: Participant P i The hash set H i , the corresponding zero-knowledge proof {π i,j } and the bound certificate and ID set are submitted to TPA; S24: Verify zero-knowledge proof: TPA receives each participant P i Submitted hash set H i , the corresponding zero-knowledge proof set {π i,j } and the bound certificate and ID set, and perform the following verification steps: First, obtain the public key of the trusted organization from the certificate and verify the certificate Cert (ID i )’s signature is valid, check whether the certificate is within the validity period, and whether the ID information in the certificate is consistent with the submitted ID set; if the certificate verification fails, the ID set submitted by the participant is directly judged to be invalid and marked as a violation; if the certificate verification passes, the zero-knowledge proof is verified. For each h(ID i,j )∈H i , TPA uses the verification algorithm Verify to check the corresponding proof π i,j Is it valid: Verify(h(ID i,j ),π i,j )=True, if the verification is successful, h(ID i,j ) is recorded as a valid user ID; otherwise, it is marked as invalid; S25: Record the hash set of legal user IDs: After verification is completed, TPA records the hash set of legal user IDs of each participant.

4. The multi-party vertical federated learning method based on data alignment and privacy protection according to claim 1, characterized in that: Step S3 specifically includes the following steps: S31: MPSI input and encrypted inverse Bloom filter are generated and sent to the server P t ; S32: MPSI calculation: The server first calculates the hash value for each element in the user ID set and searches for matching items in the client's encrypted inverse Bloom filter; the server calculates the sum of the ciphertexts and re-randomizes them; the server distributes the encryption results to the selected clients, and the clients decrypt and feed back to the server; the server summarizes the decryption results to form an intersection set, which is submitted to TPA for confirmation and then used for federated training.

5. The multi-party vertical federated learning method based on data alignment and privacy protection according to claim 4, characterized in that: The MPSI input and encrypted inverse Bloom filter generation in step S31 include the following steps: S311: Input: Each client P1-P t-1 Holds a shared decryption key sk i , the key belongs to an additive homomorphic threshold encryption scheme with zero decryption algorithm ShDec0; the public key pk is visible to all parties; S312: Initialization: Server P t Randomly select a set of hash functions And send it to the client; S313: Local encrypted inverse Bloom filter generation: Each client performs the following steps to generate an encrypted inverse Bloom filter, thereby achieving privacy-preserving user ID encryption: Calculate its user ID set S i Bloom filter BF i , Get the inverse Bloom filter IBF of the Bloom filter i , IBF i Encrypt each element in the , get the encrypted inverse Bloom filter and send it to the server P t .

6. The multi-party vertical federated learning method based on data alignment and privacy protection according to claim 4, characterized in that: In step S32, the server P t Perform the following steps to calculate the intersection user set: The set of user IDs S held by the server t Each element y in j , calculate k hash values ​​h d (y j ), and for each hash value h d (y j ) Inverse Bloom Filter IBF on each client i Find the value in: Compute the sum of the ciphertexts and re-randomize: The server will result Any distributed to t-1 clients clients and request them to j Decryption is performed; Compliance check: Each client P i Calculate its decryption share sh i,j And feed it back to the server, as follows: (1) The client first j After randomization, Sent back to the server, the server combines it into And send it to each client; (2) The server summarizes the decrypted shares sh fed back by the client i,j , and finally form the decryption share set sh 1,j ,...,sh t-1,j ; TPA verifies the decryption process: TPA monitors the client's decryption feedback process, ensures that all feedback is sent through a secure channel, and prevents the server or client from interfering with the calculation results by manipulating feedback data; TPA randomly checks the client's feedback decryption share to ensure compliance with the decryption process; (3) The server uses the combined algorithm to decrypt. For each j∈{1,...,n t }, the server uses a combined algorithm to calculate: Dec(e j )←Comb(sh 1,j ,...,sh t-1,j ); if the decryption result is zero, Dec(e j )=0, then y j In the set of user IDs of all participants, the server will j Add to the intersection set S = {y j }∪S; otherwise, discard y j ; The final intersection user set output: server P t Finally, the intersection set S is output and the result is submitted to TPA for final confirmation; after TPA verifies the intersection result, it sends the compliant intersection set to all parties.

7. The multi-party vertical federated learning method based on data alignment and privacy protection according to claim 1, characterized in that: Step S4 specifically includes the following steps: S41: Encrypted Verification Mechanism: Before each round of training begins, TPA requires all parties to submit an encrypted hash set S of user IDs used for training to verify the compliance of the training data; each party generates an encrypted hash set, and TPA verifies whether these sets are consistent with the known complete intersection set; if the verification passes, it indicates that the parties are training on the correct intersection users; if it fails, TPA will record the violation; S42: Dynamic homomorphic encryption check: TPA uses dynamic homomorphic encryption check technology to prevent participants from using non-intersection user data in some rounds; first, randomly generate checkpoints: TPA generates checkpoints at random rounds during the training process, requiring each participant to resubmit the encrypted training data set Then homomorphic encryption verification: TPA uses homomorphic encryption technology at the checkpoint to verify Is it consistent with E(S) to ensure that each round of training data comes from the complete intersection? S43: Training phase: After identifying the common entities, the parties use the data of these common entities to collaboratively train a machine learning model.

8. The multi-party vertical federated learning method based on data alignment and privacy protection according to claim 7, characterized in that: The training phase described in step S43 includes the following steps: S431: The trusted coordinator creates a key pair and sends the public key to the client and the server; S432: The client and the server encrypt and exchange the intermediate results, which are used to help calculate the gradient and loss value; S433: The client and the server calculate the encryption gradient and add the additional mask respectively, the server calculates the encryption loss, and the client and the server send the encryption result to the trusted coordinator; S434: The trusted coordinator decrypts the gradient and loss information and sends the results back to the client and server; the client and server remove the mask on the gradient information and update the model parameters based on the gradient information.

9. The multi-party vertical federated learning method based on data alignment and privacy protection according to claim 1, characterized in that: Step S5 specifically includes the following steps: Violation detection: TPA uses zero-knowledge proof and cryptographic verification technology at each key step to regularly check the compliance of all parties; Penalty measures: TPA will take the following measures against participants who are confirmed to have violated the regulations: (1) Model weight reduction: reduce its contribution weight in the federated model; (2) Financial penalties: Financial penalties are imposed on the violating party; (3) Cooperation restrictions: In the event of serious violations, the eligibility to participate in federated learning will be suspended; Incentives: For compliant parties, TPA issues compliance certificates and gives appropriate rewards to encourage all parties to follow the agreement.

Images