A verifiable privacy recommendation system based on secure multi-party computation

By adopting secure multi-party computing and verifiable protocols in the recommendation system, using Paillier homomorphic encryption and Shamir secret sharing technology, blindly processing and secret sharing of user characteristics, solving the problem of malicious opponents destroying user privacy and realizing privacy Protect and validate availability of computational results to ensure the performance of the recommended system.

CN114915402BActive Publication Date: 2024-11-12SHANDONG DUOFANG SEMICON CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210673043.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-06-14
Publication Date
2024-11-12
Estimated Expiration
2042-06-14

AI Technical Summary

Technical Problem

Existing recommendation systems are difficult to effectively protect user privacy when facing malicious adversaries, and existing privacy protection research mainly focuses on semi-honest models and cannot cope with the destructive capabilities of malicious adversaries.

Method used

Using a verifiable privacy recommendation system based on secure multi-party computing, user features are blinded by using Paillier homomorphic encryption and norms of feature vectors in the secure vector inner product protocol, and Shamir secret sharing is adopted in the secure gradient descent protocol. , users commit and verify their characteristics and share to determine whether there is a malicious attacker.

Benefits of technology

While protecting user privacy, verify the correctness and availability of calculation results, ensure that the performance of the recommendation system is not affected by the privacy protection protocol, and effectively resist malicious attacks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114915402B_ABST
    Figure CN114915402B_ABST
Patent Text Reader

Abstract

The present invention discloses a verifiable privacy recommendation system based on secure multi-party computing, comprising a server and multiple users. The users interact with other users locally, calculate similarity and generate a local similarity network, and cooperate with users in the network to train a linear model. After completion, the users interact with the server to train a feature interaction model. A verifiable secure vector inner product protocol is used to protect user privacy when the users interact with local data, a verifiable secure gradient descent protocol is used to protect user privacy when the users interact with data in a local similarity network, and a secure aggregation strategy is used to protect user privacy when the users interact with the server.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of privacy protection technology, and more specifically to a verifiable privacy recommendation system based on secure multi-party computing. Background Art

[0002] At present, recommendation systems are based on machine learning and other technical support. They can build models by analyzing user behaviors or interest preferences, predict user ratings or preferences, find users' personalized needs, and proactively recommend items that users may be interested in. Recommendation systems can improve user experience, discover "long-tail products" in commodity transactions, and serve the economy and society.

[0003] However, in the recommendation system, the recommendation server or the users who participate in the recommendation generally perform the recommendation work under a semi-honest model or a malicious model. The former means that the participants honestly perform the work according to the protocol, but try to obtain the secret information of other honest participants as much as possible; the latter means that the participants can destroy the execution of the protocol through any behavior, including tampering with data and terminating the protocol. At present, most of the existing research on privacy protection of recommendation systems is to protect privacy under the semi-honest adversary model. However, malicious adversaries have a stronger ability to destroy user privacy. Once they exist, they will cause huge losses to users and even recommendation service providers.

[0004] Therefore, providing a recommendation system that protects user privacy during the recommendation process and can verify the correctness and usability of calculation results is an urgent problem that needs to be solved by those skilled in the art. Summary of the invention

[0005] In view of this, the present invention provides a verifiable privacy recommendation system based on secure multi-party computation; in the secure vector inner product protocol, the user uses Paillier homomorphic encryption and the norm of the feature vector to blind the user features to protect the user's privacy. In terms of verification, the user commits to his or her own features and makes them public. Other users can determine whether the user's commitment is consistent with the calculation result through calculation, thereby determining whether the user is a malicious attacker. In the secure gradient descent protocol, the user uses Shamir secret sharing to protect the user's privacy. In terms of verification, the user commits to and makes public his or her share. In the stage of aggregating shares and reconstructing secrets, other users can determine whether the user is a malicious attacker by determining whether the user's commitment is consistent with the calculation result.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A verifiable privacy recommendation system based on secure multi-party computing includes a server and multiple users. The users interact with other users locally, calculate similarities and generate local similarity networks, and cooperate with users in the network to train linear models. After completion, the users interact with the server to train feature interaction models.

[0008] A verifiable secure vector inner product protocol is used to protect user privacy when users interact with local data. A verifiable secure gradient descent protocol is used to protect user privacy when users interact with data in local similar networks. A secure aggregation strategy is used to protect user privacy when users interact with servers.

[0009] Preferably, when the user interacts with other users locally, a verifiable secure gradient descent protocol is used to protect user privacy, specifically:

[0010] Given a set of users, each user i has an m-dimensional user feature vector In the secure vector inner product protocol, two large prime numbers p and q are selected, and g is The generator of , public g;

[0011] For each user i, the feature vector is divided by the norm of the feature vector to obtain a new feature vector At the same time, X i′ Make a commitment,

[0012] User u and user v perform similarity calculation. User u generates a public key pk and a private key sk through the Paillier homomorphic encryption algorithm, publishes pk and keeps sk private. User u encrypts X with pk u' , get the ciphertext And send it to user v. After receiving C, user v calculates And send it to user u. After user u gets C', he uses sk to perform the decryption operation, and the result is the similarity sim(u,v);

[0013] In the verification phase, user u calculates the commitment of user v to obtain the power modulus of the eigenvalue of the Commi commitment. The similarity between the calculated result and g is modulo pg sim(u,v) By comparison, if they are the same, the similarity is available, otherwise it means that the similarity is not available and user v may be a malicious adversary.

[0014] Preferably, when the user interacts with data in a local similar network, a verifiable secure gradient descent protocol is used to protect privacy, including:

[0015] In the secure gradient descent protocol, assuming that the user to perform gradient descent is u, the participants in the protocol execution are all neighbors in the similar network of user u, and the neighbor v generates a Lagrangian polynomial of order t-1

[0016] p v (x) = a v,0 +a v,1 x+a v,2 x 2 +…+a v,(t-1) x t-1

[0017] Among them, a v,0 It is the secret to be shared;

[0018] User u randomly selects two large prime numbers p and q, q|(p-1), g is The generators of {p,q,g} make a commitment to all neighbors, and each neighbor v makes a commitment to its own polynomial coefficients and make it public;

[0019] In the secret sharing phase, each neighbor divides his secret into N sub-shares, where N is the size of the user u network, retains his own share, and distributes the other N-1 shares to the corresponding neighbors;

[0020] In the secret synthesis phase, neighbors receive the sub-share SS sent by other users. v Afterwards, based on Synthetic Master Share MS v And send the master share to user u based on Synthesize the master secret;

[0021] In the verification phase, after receiving the sub-share, the neighbor verifies Is it true? If so, it proves that the sub-share is available; otherwise, it proves that the user v who sent the sub-share may be malicious. User u receives the main share ms sent by the neighbor. v After verification Is it true? If so, it proves that the primary share is available; otherwise, it proves that the user v who sent the primary share may be a malicious user.

[0022] Preferably, the adopting of a security aggregation strategy to protect user privacy in the interaction between the user and the server specifically includes:

[0023] All users u receive their signing keys and the authentication keys of other users The user generates two pairs of keys through Diffie-Hellman key negotiation Concatenate the two public keys to get And use the signature key Sign it to get σ u , concatenate and σ u Send it to the server, and the server counts the number of users who send the message, denoted as U 1 (U 1 ≥t). The server broadcasts to all users in U 1

[0024] The user judges whether U 1 is ≥t, ensure that all public key pairs are different, and use to verify whether the signature σ u is correct. If according to the above operations, it is proved that the message is legal, then user u selects a random number b u as the random seed of the PRG, and uses secret sharing to share the key and b u with the user group v, and gets the share values and The user performs encryption authentication on to get e u,v and sends it to the server. The server counts the number of users at this time, denoted as U 2 , and then broadcasts e u,v to the users in U 2 .

[0025] The user judges whether U 2 is ≥t. If U 2 <t, exit the program. Otherwise, the user generates a random number s u,v shared by user u and user v through Diffie - Hellman key negotiation, and expands it into a random vector p u,v = PRG(s u,v )·Δ u,v , where, if u > v, Δ u,v = 1; otherwise, Δ u,v = -1. At the same time, expand b u into the vector p u = PRG(b u ). Calculate and send y u to the server. The server counts the number of received y u , denoted as U 3 , and broadcasts U 3 to the users among them.

[0026] The user judges whether U 3 is ≥t. If U 3 ​<t, exit the program. Use Signature U 3 , to obtain the signature value σ' u Send it to the server. The server counts the received σ' u The quantity of, denoted as U 4 , and then send {v, σ' v} to the users in U 4 .

[0027] The user judges whether U 4 is ≥ t, use to verify whether σ' u is correct. If according to the above operations, it is proved that the message is legal, then the user decrypts the authentication e u,v , to obtain judge whether u' is equal to u and v' is equal to v at this time. At this time, the offline users are U 3 / U 2 , and the online users are U 3 . Send the of the offline users and the of the online users to the server. The server counts the number of users sending messages, denoted as U 5 . According to secret sharing, reconstruct for each offline user according to recalculate p v,u . Reconstruct b u for each online user according to secret sharing, and then obtain p again using PRG u .

[0028] The server calculates which is the aggregation value.

[0029] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses a verifiable privacy recommendation system based on secure multi-party computation; in the secure vector inner product protocol, the user uses Paillier homomorphic encryption and the norm of the feature vector to blind the user features to protect user privacy. In terms of verification, the user makes a commitment to and discloses his own features, and other users can judge whether the user's commitment is consistent with the operation result through calculation, so as to judge whether the user is a malicious attacker. In the secure gradient descent protocol, the user uses Shamir secret sharing to protect user privacy. In terms of verification, the user makes a commitment to and discloses the share he shared. In the stage of aggregating shares and reconstructing secrets, other users can judge whether the user's commitment is consistent with the operation result, so as to judge whether the user is a malicious attacker. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0031] Figure 1 The accompanying drawing is a schematic diagram of the system model structure provided by the present invention.

[0032] Figure 2 The accompanying drawing is a schematic diagram of the privacy protection model structure provided by the present invention.

[0033] Figure 3 The accompanying drawing is a schematic diagram of the influence of the similarity ratio and the similarity threshold value Thre provided by the present invention on the recommendation performance.

[0034] Figure 4 The accompanying drawing is a schematic diagram of a curve showing a change in the total training time when the values ​​of p and Thre are changed according to the present invention.

[0035] Figure 5 The accompanying drawing is a schematic diagram showing the effect of the number of items provided by the present invention on the protocol running time.

[0036] Figure 6 The accompanying drawing is a schematic diagram showing the effect of the number of users provided by the present invention on the protocol running time. DETAILED DESCRIPTION

[0037] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0038] The embodiment of the present invention discloses a verifiable privacy recommendation system based on secure multi-party computing, including a server and multiple users, wherein the user interacts with other users locally to perform data interaction, calculates similarity and generates a local similarity network, and cooperates with users in the network to train a linear model. After completion, the user interacts with the server to train a feature interaction model;

[0039] A verifiable secure vector inner product protocol is used to protect user privacy when users interact with local data. A verifiable secure gradient descent protocol is used to protect user privacy when users interact with data in local similar networks. A secure aggregation strategy is used to protect user privacy when users interact with servers.

[0040] To further optimize the above technical solution, when users interact with other users locally, a verifiable secure gradient descent protocol is used to protect user privacy. Specifically:

[0041] Given a set of users, each user i has an m-dimensional user feature vector In the secure vector inner product protocol, two large prime numbers p and q are selected, and g is The generator of , public g;

[0042] For each user i, the feature vector is divided by the norm of the feature vector to obtain a new feature vector At the same time, X i′ Make a commitment,

[0043] User u and user v perform similarity calculation. User u generates a public key pk and a private key sk through the Paillier homomorphic encryption algorithm, publishes pk and keeps sk private. User u encrypts X with pk u' , get the ciphertext And send it to user v. After receiving C, user v calculates And send it to user u. After user u gets C', he uses sk to perform the decryption operation, and the result is the similarity sim(u,v);

[0044] In the verification phase, user u calculates the commitment of user v to obtain the power modulus of the eigenvalue of the Commi commitment. The similarity between the calculated result and g is modulo pg sim(u,v) By comparison, if they are the same, the similarity is available, otherwise it means that the similarity is not available and user v may be a malicious adversary.

[0045] To further optimize the above technical solution, when users interact with data in local similar networks, a verifiable secure gradient descent protocol is used to protect privacy, including:

[0046] In the secure gradient descent protocol, assuming that the user to perform gradient descent is u, the participants in the protocol execution are all neighbors in the similar network of user u, and the neighbor v generates a Lagrangian polynomial of order t-1

[0047] p v (x) = a v,0 +a v,1 x+a v,2 x 2 +…+a v,(t-1) x t-1

[0048] Among them, a v,0It is the secret to be shared;

[0049] User u randomly selects two large prime numbers p and q, q|(p-1), g is The generators of {p,q,g} make a commitment to all neighbors, and each neighbor v makes a commitment to its own polynomial coefficients and make it public;

[0050] In the secret sharing phase, each neighbor divides his secret into N sub-shares, where N is the size of the user u network, retains his own share, and distributes the other N-1 shares to the corresponding neighbors;

[0051] In the secret synthesis phase, neighbors receive the sub-share SS sent by other users. v Afterwards, based on Synthetic Master Share MS v And send the master share to user u based on Synthesize the master secret;

[0052] In the verification phase, after receiving the sub-share, the neighbor verifies Is it true? If so, it proves that the sub-share is available; otherwise, it proves that the user v who sent the sub-share may be malicious. User u receives the main share ms sent by the neighbor. v After verification Is it true? If so, it proves that the primary share is available; otherwise, it proves that the user v who sent the primary share may be a malicious user.

[0053] To further optimize the above technical solution, a security aggregation strategy is adopted to protect user privacy in the interaction between users and servers, including:

[0054] All users u receive their signing keys and the authentication keys of other users The user generates two pairs of keys through Diffie-Hellman key negotiation Concatenate the two public keys to get And use the signing key Sign it to get σ u ,connect and σ u Sent to the server, the server counts the number of users who sent messages, recorded as U 1 (U 1 ≥t). The server sends 1 Broadcast to all users in

[0055] User judgment U 1 Is ≥ t, ensure that all public key pairs are different and use Verify the signature σ u to see if it is correct. If the message is proven to be legal according to the above operations, then user u selects a random number b u as the random seed of the PRG, and uses secret sharing to share the key with the user group v and b u , obtaining the share values and The user encrypts and authenticates to get e u,v and sends it to the server. The server counts the number of users at this time, denoted as U 2 , then broadcasts e u,v to the users in U 2 .

[0056] The user determines whether U 2 is ≥ t. If U 2 < t, exit the program. Otherwise, the user generates a random number s shared by user u and user v through Diffie - Hellman key negotiation u,v , and expands it into a random vector p u,v = PRG(s u,v )·Δ u,v , where, if u > v, Δ u,v = 1; otherwise, Δ u,v = -1. At the same time, b u is also expanded into the vector p u = PRG(b u ). Calculate and send y u to the server. The server counts the number of received y u , denoted as U 3 , and broadcasts U 3 to the users in it.

[0057] The user determines whether U 3 is ≥ t. If U 3 < t, exit the program. Use to sign U 3 to obtain the signature value σ' u and send it to the server. The server counts the number of received σ' u , denoted as U 4 , then sends {v, σ' v} to the users in U 4 .

[0058] The user determines whether U 4 is ≥ t, and uses to verify σ' uIs it correct? If the above operation proves that the message is legitimate, the user decrypts and authenticates the message. u,v ,get Determine whether u' is equal to u and v' is equal to v. At this time, the offline user is U 3 / U 2 , the online user is U 3 . and online users Sent to the server, the server counts the number of users who sent messages, recorded as U 5 . Based on secret sharing, reconstruct the according to Recalculate p v,u . Reconstruct b for each online user based on secret sharing u , and then use PRG again to get p u .

[0059] Server Computing This is the aggregate value.

[0060] We applied the two protocols to a distributed recommendation system based on factorization machine (FM) and selected a 538k real Black Friday dataset for experiments. The following table shows the impact of the feature interaction model size K value on the AUC of the original FM recommendation system and our recommendation system. As can be seen from the table, our privacy protection protocol does not affect the recommendation performance, and even due to the participation of similar neighbors, the recommendation effect is slightly better than the original FM recommendation system.

[0061] Model FM VePriRec K=5 0.62431 0.64075 K=10 0.62847 0.64155 K=15 0.62991 0.64362

[0062] In addition, we tested the impact of the parameters of the two protocols on the performance, the efficiency of the two protocols, and the impact of the number of users and the number of items on the efficiency of the protocols.

[0063] exist Figure 3 In this paper, we study the impact of similarity ratio p and similarity threshold Thre on recommendation performance. Among them, p represents the influence of neighbors in a user's similarity network on his choice. The larger p is, the greater the influence is. Thre represents the threshold of the influence of neighbors on the user. It controls the size of the user's similarity network. The higher the threshold, the fewer neighbors there are. In the similarity ratio experiment, we set K = 10, Thre = 0.9, and the AUC curve of VePriRec is shown in the figure below. Figure 3 As shown in Figure 2, when 0.5≤p≤0.8, AUC is positively correlated with p; when 0.8≤p≤1, AUC is negatively correlated with p. This indicates that the influence of similar user preferences is limited. When p=1, the SGD protocol is equivalent to the DGD protocol.

[0064] In the similarity threshold experiment, we set K = 10, p = 0.8, and the AUC curve of VePriRec is as follows: Figure 3 As shown in the figure. When 0.5≤Thre≤0.9, AUC is positively correlated with Thre. Even if the size of the similarity network becomes smaller, the more similar users have a greater impact on the recommendation. Thre=1 means that in this similarity network, all neighbors have the same features as the user, but this situation is very rare and can be ignored.

[0065] exist Figure 4 In the example, we control the number of items to 2000. When changing the values ​​of p and Thre, the change in the total training time is shown in the curve. As Thre increases, the size of the user's similarity network decreases, that is, the number of neighbors participating in secret sharing decreases, and the calculation time decreases. When Thre = 1, the user's similarity network only has himself, and no secret sharing is required. Therefore, the protocol execution time is the shortest when Thre = 1. The value of p affects the update effect of the user's neighbors on his linear model, so p does not affect the training time.

[0066] exist Figure 5 In , we study the effect of the number of items on the protocol running time. When the number of items increases, the running time of the similarity network construction protocol increases, while the gradient descent protocol hardly changes. Figure 6 In , we study the effect of the number of users on the protocol running time. When the number of users increases, the time of both protocols increases, but at different rates.

[0067] The beneficial effects of this application are:

[0068] Using a distributed computing model, the data, models, and computing processes that contain user privacy are handed over to users on their own devices; the parts that do not contain user privacy are handed over to the recommendation server. While protecting user privacy, it reduces the storage pressure on user devices;

[0069] Two privacy protection protocols are designed to protect the privacy of users and verify the availability of calculation results when there may be malicious participants in the system. At the same time, it is ensured that the performance of the recommendation system will not be affected by the privacy protection protocol.

[0070] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0071] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A verifiable privacy recommendation system based on secure multi-party computing, characterized in that: It includes a server and multiple users. The users interact with other users locally, calculate similarity and generate a local similarity network, and train a linear model in cooperation with users in the network. After completion, the users interact with the server to train a feature interaction model. A verifiable secure vector inner product protocol is used to protect user privacy when users interact with local data. A verifiable secure gradient descent protocol is used to protect user privacy when users interact with data in local similar networks. Specifically: Given a set of users, each user i has an m-dimensional user feature vector In the secure vector inner product protocol, two large prime numbers p and q are selected, and g is The generator of , public g; For each user i, the feature vector is divided by the norm of the feature vector to obtain a new feature vector At the same time, X i′ Make a commitment, User u and user v perform similarity calculation. User u generates a public key pk and a private key sk through the Paillier homomorphic encryption algorithm, publishes pk and keeps sk private. User u encrypts X with pk u′ , get the ciphertext And send it to user v. After receiving C, user v calculates And send it to user u. After user u gets C', he uses sk to perform the decryption operation, and the result is the similarity sim(u,v); In the verification phase, user u calculates the commitment of user v to obtain the power modulus of the eigenvalue of Commit commitment. The similarity between the calculated result and g modulo pg sim(u,v) By comparison, if they are the same, the similarity is available, otherwise it means that the similarity is not available and user v may be a malicious adversary; In the secure gradient descent protocol, assuming that the user to perform gradient descent is u, the participants in the protocol execution are all neighbors in the similar network of user u, and the neighbor v generates a Lagrangian polynomial of order t-1 p v (x)=a v,0 +a v,1 x+a v,2 x 2 +…+a v,(t-1) x t-1 Among them, a v,0 It is the secret to be shared; User u randomly selects two large prime numbers p and q, and g is The generator of {p,q,g} is announced to all neighbors, and each neighbor v makes a commitment to his own polynomial coefficients and make it public; In the secret sharing phase, each neighbor divides his secret into N sub-shares, where N is the size of the user u network, retains his own share, and distributes the other N-1 shares to the corresponding neighbors; In the secret synthesis phase, neighbors receive the sub-share SS sent by other users. v Afterwards, based on Synthetic Master Share MS v And send the master share to user u based on Synthesize the master secret; In the verification phase, after receiving the sub-share, the neighbor verifies Is it true? If so, it proves that the sub-share is available; otherwise, it proves that the user v who sent the sub-share may be a malicious user. User u receives the main share ms sent by the neighbor. v After verification Is it true? If so, it proves that the primary share is available; otherwise, it proves that the user v who sent the primary share may be a malicious user A secure aggregation strategy is used to protect user privacy in the interaction between users and servers, including: All users u receive their signing keys and the authentication keys of other users The user generates two pairs of keys through Diffie-Hellman key negotiation Concatenate the two public keys to get And use the signing key Sign it to get σ u ,connect and σ u Send to the server, the server counts the number of users who sent messages, recorded as U1, U1 ≥ t, the server broadcasts to all users in U1 The user determines whether U1 ≥ t, ensuring that all public key pairs are different and using Verify signature u Is it correct? If the above operation proves that the message is legal, then user u selects a random number b u As a random seed for PRG, using secret sharing, share the key with user group v and b u , get the share value and User Encryption authentication to obtain e u,v And send it to the server, the server counts the number of users at this time, record it as U2, and then e u,v Broadcast to users in U2; The user determines whether U2 is ≥ t. If U2 < t, the program exits. Otherwise, the user generates a random number s shared by user u and user v through Diffie-Hellman key negotiation u,v , and expands it into a random vector p using PRG u,v = PRG(s u,v )·Δ u,v , where if u > v, Δ u,v = 1; otherwise, Δ u,v = -1; at the same time, expand b u into vector p as well u = PRG(b u ), calculate and send y u to the server. The server counts the number of received y u , denoted as U3, and broadcasts U3 to the users among them; The user determines whether U3 is ≥ t. If U3 < t, the program exits and signs U3 to obtain the signature value σ'. u It is sent to the server, and the server counts the received σ'. u The quantity is denoted as U4, and then {v, σ' v} is sent to the users in U4; The user determines whether U4 ≥ t, using Verify σ' u Is it correct? If the above operation proves that the message is legitimate, the user decrypts and authenticates the message. u,v ,get Determine whether u' is equal to u, and whether v' is equal to v. At this time, the offline user is U3 / U2, and the online user is U3. and online users The server counts the number of users who sent messages, which is recorded as U5. Based on secret sharing, it reconstructs the message for each offline user. according to Recalculate p v,u , reconstruct b for each online user based on secret sharing u , and then use PRG again to get p u ; Server Computing This is the aggregate value.