A multi-party secure aggregation method, system and storage medium for classification tasks

Abstract

The present invention discloses a multi-party secure aggregation method, system and storage medium for classification tasks, comprising the following steps: reasoning query data to generate a #imgabs0#-dimensional vector result, and randomly generating a #imgabs1#-dimensional vector; adding the #imgabs2#-dimensional vector result to the #imgabs3#-dimensional vector to obtain a model inference result of the #imgabs4#th participant; encrypting the model inference result based on a private key, adding the #imgabs5#-dimensional vector to the Gaussian noise parameter of the #imgabs6#th participant to obtain a noise parameter; sending the encrypted model inference result and noise parameter to a demander; obtaining a result ciphertext fed back by the demander, decrypting the result ciphertext using a private key, and sending the decrypted result ciphertext to the demander, wherein the demander calculates a multi-party aggregation result that satisfies differential privacy. The invention provides a technology for ensuring that the inference results of each participant are obtained without leaking the individual annotation information of any participant, thereby achieving the goal of not leaking the private data of each party except the calculation result.

Description

Technical Field

[0001] The present invention relates to the field of network and information security technology, and in particular to a multi-party security aggregation method, system and storage medium for classification tasks. Background Art

[0002] Machine learning benefits from large training datasets, but it may not always be possible for any single entity to collect all the data, especially when working with privacy-sensitive data. In many cases, such as healthcare and finance, different parties may wish to collaborate and learn from each other's data but are prohibited from doing so due to privacy regulations. Some laws and regulations prohibit explicit data sharing between parties by centrally collecting datasets. For collaborative approaches using federated learning, model gradient sharing still presents a significant attack risk. Due to the implicit memory nature of machine learning, attackers can obtain model gradient parameters to conduct membership inference attacks and model theft attacks.

[0003] Based on this, a method currently proposed for joint learning is to record inference results through shared data. However, the model inference results are all 0s and 1s (the model inference results are vectors of the form [0,1,0,0,0...]). During encrypted transmission, they are vulnerable to known-plaintext attacks. A third party can infer the inference results of each participant through frequency statistics. Furthermore, in scenarios with multiple users, they are susceptible to differential attacks. Summary of the Invention

[0004] Based on the technical problems existing in the background technology, the present invention proposes a multi-party secure aggregation method, system and storage medium for classification tasks, which ensures that the reasoning results of each participant are obtained without leaking the individual annotation information of any participant, and realizes a technology that does not leak the private data of each party except the calculation results.

[0005] The present invention proposes a multi-party secure aggregation method for classification tasks, comprising the following steps:

[0006] Receive query data, public key, private key and Gaussian noise scale ;

[0007] Perform inference generation on the query data Dimensional vector results , randomly generated dimensional vector , Indicates the Participants

[0008] Will Dimensional vector results and dimensional vector Add up to get Model inference results of each participant ;

[0009] Model inference results based on private key Encryption based on Gaussian noise scale Calculate the Gaussian noise parameters and dimensional vector With the The noise parameter is obtained by adding the Gaussian noise parameters of the participants ;

[0010] The encrypted model inference results And the noise parameters Send to a semi-trusted third party;

[0011] Obtain the ciphertext of the result of the demander's feedback , use the private key to encrypt the result Decrypt and decrypt the result ciphertext Send to a semi-trusted third party, and the demander will cipher the results feedback from each participant After combining and combining the total noise parameters Get a multi-party aggregation result that satisfies differential privacy ;

[0012] Multi-party aggregation results The calculation process is as follows:

[0013]

[0014]

[0015]

[0016]

[0017]

[0018]

[0019]

[0020]

[0021] in, represents the total inference result after decryption, represents the total noise parameter, represents an integer, Indicates that each participant has a ciphertext of the result The combined result after decryption, Indicates the Participants encrypt the result The result after decryption, Indicates the The noise parameters of the participants, Indicates the The private keys of each participant, represents the number of participants, Represents a Gaussian noise scale Calculate the Gaussian noise parameters, Table semi-random number, express of Power, and Represents a custom function, , .

[0022] Further, after receiving the query data, public key, private key and Gaussian noise scale In the example, the query data received is as follows: Each participant receives the data to be discussed and sent by the same demander using the communication protocol. The dimension vector is the result data you want to receive, and the data to be discussed and the label vector are the query data.

[0023] Further, after receiving the query data, public key, private key and Gaussian noise scale In the example, the public key, private key, and Gaussian noise scale The generation process is as follows:

[0024] Select an integer , calculate the private key and the public key ( , , ),in and are two strong prime numbers;

[0025] Selecting a privacy budget and error ,based on , select the Gaussian noise scale , represents the parameterized differential privacy budget;

[0026] Use the key sharing protocol to share the private key Decompose into , and and Assigned to participants .

[0027] Furthermore, the private key The calculation process is as follows:

[0028]

[0029]

[0030] in, Represents the result ciphertext, ( ) represents a randomly selected number pair, Table semi-random number, least common multiple, Represents the greatest common divisor.

[0031] Furthermore, based on the private key pair model reasoning results Encryption is in progress. The encryption formula is as follows:

[0032]

[0033] in, Indicates the result of model inference based on the private key The encryption result of Indicates the An integer randomly chosen by the participants.

[0034] Further, in dimensional vector With the The noise parameter is obtained by adding the Gaussian noise parameters of the participants In the noise parameter The calculation formula is as follows:

[0035]

[0036] in, Indicates the Gaussian noise parameters of each participant.

[0037] Furthermore, after obtaining the result ciphertext of the demander's feedback , use the private key to encrypt the result Decryption in progress;

[0038] The demander encrypts the result The calculation formula is as follows:

[0039]

[0040] Private key pair result ciphertext The decryption formula is as follows:

[0041]

[0042] in, Indicates the Participants encrypt the result The result after decryption, Indicates the The private key of each participant.

[0043] A multi-party secure aggregation system for classification tasks, including a receiving module, an inference generation module, an addition module, an encryption calculation module, a sending module and a feedback acquisition module;

[0044] The receiving module is used to receive query data, public key, private key and Gaussian noise scale ;

[0045] The inference generation module is used to generate inferences on the query data Dimensional vector results , randomly generated dimensional vector , Indicates the Participants

[0046] The adding module is used to Dimensional vector results and dimensional vector Add up to get Model inference results of each participant ;

[0047] The encryption calculation module is used to infer the model results based on the private key Encryption based on Gaussian noise scale Calculate the Gaussian noise parameters and dimensional vector With the The noise parameter is obtained by adding the Gaussian noise parameters of the participants ;

[0048] The sending module is used to send the encrypted model inference results And the noise parameters Send to a semi-trusted third party;

[0049] The feedback acquisition module is used to obtain the result ciphertext of the demander's feedback , use the private key to encrypt the result Decrypt and decrypt the result ciphertext Send to a semi-trusted third party, and the demander will cipher the results feedback from each participant After combining and combining dimensional vector and the sum of the Gaussian noise parameters Get a multi-party aggregation result that satisfies differential privacy ;

[0050] Multi-party aggregation results The calculation process is as follows:

[0051]

[0052]

[0053]

[0054]

[0055]

[0056]

[0057]

[0058]

[0059] in, represents the total inference result after decryption, represents the total noise parameter, represents an integer, Indicates that each participant has a ciphertext of the result The combined result after decryption, Indicates the Participants encrypt the result The result after decryption, Indicates the The noise parameters of the participants, Indicates the The private keys of each participant, represents the number of participants, Represents a Gaussian noise scale Calculate the Gaussian noise parameters, Table semi-random number, express of Power, and Represents a custom function, , .

[0060] A computer-readable storage medium stores a plurality of programs, wherein the plurality of programs are used to be called by a processor and execute the multi-party security aggregation method as described above.

[0061] Those skilled in the art will understand that all or part of the steps of implementing the above-mentioned method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above-mentioned method embodiment; and the aforementioned storage medium includes: ROM, RAM, disk or optical disk, etc. Various media that can store program codes.

[0062] The advantages of the multi-party secure aggregation method, system and storage medium for classification tasks provided by the present invention are: the multi-party secure aggregation method, system and storage medium for classification tasks provided in the structure of the present invention ensure that the reasoning results of each participant are obtained without leaking the individual annotation information of any participant, and realize the technology of not leaking the privacy data of each party except the calculation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a structural schematic diagram of the present invention;

[0064] Figure 2 A framework diagram for implementing multi-party security aggregation methods;

[0065] Figure 3 for Figure 2 The symbolic flow chart of the framework to implement the multi-party secure aggregation method;

[0066] Figure 4 is a Gaussian distribution graph. DETAILED DESCRIPTION

[0067] The technical solutions of the present invention are described in detail below through specific embodiments. Numerous specific details are set forth in the following description to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art may make similar modifications without departing from the scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0068] The multi-party security aggregation mechanism described in this invention mainly includes: request submission, key generation, data encryption, data decryption and result combination. This technology involves two types of entities: Participants ...... With a demander, all entities are semi-trusted (honest and curious). A privacy-preserving technique is provided for securely aggregating classification result vectors (multi-party model inference results). Each party labels records using a local classifier, and the results are aggregated using a multi-party secure computation method. This process does not require the involvement of an additional trusted third party and can obtain computation results without revealing any participant's individual annotation information. This technology ensures that no private data other than the computation result is leaked. It also effectively avoids floating-point encryption and decryption operations, reducing precision errors and improving computational efficiency. The details are as follows.

[0069] like Figures 1 to 4 As shown, the present invention proposes a multi-party secure aggregation method for classification tasks, including the following steps S1 to S6:

[0070] S1: The participant receives the query data, public key, private key and Gaussian noise scale .

[0071] S2: Inferring and generating the query data Dimensional vector results , randomly generated dimensional vector , Indicates the Participants.

[0072] S3: Dimensional vector results and dimensional vector Add up to get Model inference results of each participant .

[0073] S4: Model inference results based on private key Encryption based on Gaussian noise scale Calculate the Gaussian noise parameters and dimensional vector With the The noise parameter is obtained by adding the Gaussian noise parameters of the participants .

[0074] S5: Encrypted model inference results And the noise parameters Sent to a semi-trusted third party.

[0075] S6: Obtain the ciphertext of the result of the demander's feedback , use the private key to encrypt the result Decrypt and decrypt the result ciphertext Send to a semi-trusted third party, and the demander will cipher the results feedback from each participant After combining and combining the total noise parameters Get a multi-party aggregation result that satisfies differential privacy .

[0076] Based on steps S1 to S6, a privacy protection technology is provided for securely aggregating classification result vectors (multi-party model inference results), that is, each party uses a local classifier to label the records and uses a multi-party secure computing method to aggregate the results. This process does not require the introduction of an additional trusted third party, and can obtain the calculation results without leaking the individual labeling information of any participant. This achieves the goal of obtaining the inference results of each participant without leaking the privacy data of each party in addition to the calculation results, and can also effectively avoid floating-point encryption and decryption operations, reduce precision errors and improve computing efficiency.

[0077] The following details multi-party security aggregation:

[0078] (1) The demander makes a request: The demander asks for a certain data Participants send inquiries. Hoping to get the opinions of all participants after aggregation , is a dimensional vector ( Indicates the possible result category of this record), indicating the The model reasoning results of each participant on this record.

[0079] First, the demander uses the communication protocol to send the data to be discussed and the results to be received. , ]The meaning of each field is sent to each participant.

[0080] (2) The keys (public and private) are generated by an authority in the cryptographic setting, as follows:

[0081] (21) Select a sufficiently large integer , calculate the private key and the public key ( , , ), ( confidentiality to any participant), where and are two strong prime numbers;

[0082]

[0083]

[0084] in, Represents the result ciphertext, ( ,b) represents a randomly selected number pair, Table semi-random number, Indicates the least common multiple (LCM). (,) means finding the greatest common divisor (GCD).

[0085] (22) Selecting a privacy budget ( The smaller it is, the higher the privacy protection. Set the range to (0,10]) and the error ( It means that it can only be tolerated Probabilistic violation of strict differential privacy is actually a relaxed form of standard differential privacy, namely relaxed differential privacy, with parameters It is called relaxation term, which actually means relaxing the standard differential privacy The range is usually set to ),based on , select the Gaussian noise scale , represents the parameterized differential privacy budget;

[0086] (23) Use the key sharing protocol to share the private key Decompose into , and and Assigned to participants , is the number of participants, In 1 to Take a value between .

[0087] (3) Data encryption: encryption function , the plaintext Mapped to a ciphertext where In the formula is a randomly selected integer, this application dimensional vector (described above is a dimensional vector) encryption.

[0088] (31) Each participant Use the local model to infer the received query data, and finally return the processing result to generate Dimensional (labeled) vector results , Example: [0,1,...,0,...0].

[0089] (32) Each participant randomly generates dimensional vector , the vector elements are integers and belong to [0,4w] (the random number range can be adjusted dynamically, and the value given here is a reference) and obey uniform distribution.

[0090] (33) Calculation = , for dimensional vector.

[0091] (34) Encryption , For participants A randomly selected integer ( ).

[0092] in, Indicates the result of model inference based on the private key The encryption result of Indicates the An integer randomly chosen by the participants.

[0093] Note: Participant Encryption is performed, where each party There are one or more elements, here Perform encryption operations on each element in;

[0094] = ( ) = =

[0095] (35) Based on Gaussian noise scale Calculate the Gaussian noise parameters and dimensional vector With the The noise parameter is obtained by adding the Gaussian noise parameters of the participants , participants Using the communication protocol and Send to a semi-trusted third party;

[0096]

[0097] in, Indicates the Gaussian noise parameters of each participant.

[0098] Each participant uses the encryption method recorded in (31) to (35) to Encrypt, and then Sent to a semi-trusted third party to achieve encrypted data transmission.

[0099] 4. Data decryption, mainly including the following contents:

[0100] (41) Demander’s calculation result ciphertext and total noise parameters ;

[0101]

[0102]

[0103] The total noise parameter for Noise parameters of each participant Add together, where indivual The sum obtained is Adding noise can prevent the true results from being directly exposed. The purpose of noise is to give a relatively ambiguous result without affecting the user experience, thereby avoiding the leakage of the true results and causing a series of privacy risks.

[0104] (42) Decrypted separately, the demander will Sent to all participants, Using your own private key Decrypt the ciphertext and send the decrypted result Send to a semi-trusted third party;

[0105] The demander sends the data that needs to be decrypted to each participant. The participant decrypts it based on its own private key and feeds back the decrypted result to the semi-trusted third party, realizing encrypted transmission between the participants, the semi-trusted third party and the demander.

[0106] 5. Result combination, mainly including the following contents:

[0107] (51) The demander combines the decryption results of each participant to obtain , and calculate , , .

[0108] (52) Calculation , and finally obtain a multi-party aggregation result that satisfies differential privacy;

[0109]

[0110] Through the above (1) to (5), it is guaranteed that the reasoning results of each participant are obtained without leaking the individual annotation information of any participant, and the technology is realized that the privacy data of each party is not leaked except for the calculation results.

[0111] Let's take the following two scenarios as examples to understand

[0112] As a first example, in healthcare scenarios, certain issues require multi-party consultation and resolution. For example, a patient with certain characteristics needs to be determined to have COVID-19 infection, heart disease, or lung stones. In this case, a semi-trusted third party can submit data annotation requests to each participant. To protect the privacy of each participant, the third party can only obtain the participant's aggregate opinion. For example, [2, 3, 5] indicates that two participants believe the patient has COVID-19 infection, three believe the patient has heart disease, and five believe the patient has lung stones.

[0113] As a second example, in a financial lending scenario, a certain bank, xxx, receives a loan request from a customer. However, due to factors such as the excessive amount of the loan, it cannot decide whether to grant the loan. In this case, it sends a query request to all partner banks. The lender's relevant information (user privacy information can be obfuscated, but this is not addressed here) is sent to each partner bank. Each bank uses its past experience to infer whether the loan is acceptable. Ultimately, bank xxx obtains the aggregate opinion of all partner banks (for example, [5,9] means five banks agree and nine disagree). However, it cannot obtain the opinion of a single bank.

[0114] The verification process for the correctness of data transmission during the encryption process is as follows:

[0115] (1) Select a sufficiently large integer ,in and are two strong prime numbers;

[0116] (2) Calculate the private key ( Keep secret from any participant) and the public key (n, , ), is a semi-random number,

[0117] in( ,b) is a randomly selected number pair;

[0118] (3) Selecting a privacy budget and error Select the Gaussian noise scale ;

[0119] (4) Use key agreement to transfer the private key Decompose into , and send and Assigned to participants ;

[0120] (5) The demander uses a secure communication protocol to send the data to be discussed and the results to be received. , ]The meaning of each field is sent to each participant;

[0121] (6) Each participant Use the local model to reason about data records and generate Dimensional (labeled) vector results , example: [0,1,...,0,...0];

[0122] (7) Each participant randomly generates dimensional vector , the vector elements are integers and belong to [0, ], obeys uniform distribution;

[0123] (8) Calculation = , for dimensional vector;

[0124] = + =

[0125] For each participant Calculate according to the above formula;

[0126] (9) Encryption ( )= , For participants A randomly selected integer ( Note: Perform encryption operations on each element in;

[0127] = ( ) = =

[0128] (10) Participants Using the communication protocol ( ) and noise parameters Send to a semi-trusted third party;

[0129] (11) Demand-side calculation and

[0130] .

[0131] Among them, it can be understood that

[0132]

[0133]

[0134]

[0135] (12) Decrypt separately, the demander will Sent to all participants, Using your own private key Decrypt the ciphertext, , and the decrypted result to third parties;

[0136]

[0137]

[0138]

[0139] (13) The demander combines the decryption results of each participant to obtain , and calculate , ;

[0140] in:

[0141]

[0142]

[0143]

[0144]

[0145]

[0146]

[0147] make

[0148]

[0149]

[0150]

[0151] in

[0152] ∵ ,

[0153] ∴

[0154] ∴ ,Right now

[0155] ,Right now

[0156] ∴ Right now Right now

[0157] ∴

[0158] ∴ , <

[0159] ∴

[0160] And there are:

[0161]

[0162]

[0163]

[0164] ···

[0165]

[0166] ∴

[0167]

[0168] ∴

[0169] ∴

[0170]

[0171] Again∵

[0172] ∴

[0173]

[0174]

[0175] (14) Calculation

[0176]

[0177]

[0178]

[0179]

[0180]

[0181] The proof is complete.

[0182] above, and The formal parameter used in the derivation process of the power relationship can be understood as the existence of a positive integer such that the relationship holds true; To define parameter reference ; and The formal parameter used in the derivation process of prime number relations can be understood as the existence of a positive integer such that the relation holds. Similarly, for the derived power relation, the existence of Make Established.

[0183] As an example, Figures 2 to 4 As shown:

[0184] (A) Proposing a demand: The demander R has doubts about a certain picture and asks all participants. He hopes to get the opinions of all participants after aggregation. , is a 3D vector ([leopard, Persian leopard, Javan leopard]), Indicates the The reasoning result of each participant on this record, such as [0, 0, 1] represents the Participants believe it is a Javan leopard.

[0185] (B) Key generation, which mainly includes the following: (This step is completed by the authority)

[0186] (B1) Select a sufficiently large integer ,in and are two strong prime numbers. For simple calculation, two small prime numbers are selected here. =7, =11. Calculation yields =77;

[0187] Calculate the private key ;

[0188] Randomly select pairs of ( =3, =2), =2020;

[0189] The public key is: =77, , =1002

[0190] Select implicit budget and error ,calculate = 1.937, selected =2.0.

[0191] Using key agreement Decompose into

[0192] ,Will and Assigned to participants .

[0193] (C) Data encryption, mainly including the following:

[0194] Encryption Function = , maps the plaintext m to a ciphertext, where r is a randomly selected integer in the formula. Dimensional vector encryption, as follows:

[0195] (C1) The demander sends the data to be discussed (image data in this scenario) and the discussion results they wish to receive, i.e., the 3D vector [leopard, Persian leopard, Javan leopard], to all participants.

[0196] (C2) Participants Use the local model to reason about data records and generate Dimensional (labeled) vector results , Example: [0,1,...,1,...0];

[0197] Participants (Using local models) reasoning about the image and generating 3D (annotated) vector results =[0,1,0] (i.e., the participant believes the image is a Persian leopard);

[0198] Participants (Using local models) reasoning about the image and generating 3D (annotated) vector results =[1,0,0];

[0199] Participants (Using local models) reasoning about the image and generating 3D (annotated) vector results =[0,1,0];

[0200] Participants (Using local models) reasoning about the image and generating 3D (annotated) vector results =[0,0,1];

[0201] (C3) Each participant randomly generates dimensional vector , the vector elements are integers and belong to [0, ] obeys uniform distribution;

[0202] Participants Randomly generate 3D vectors ;

[0203] Participants Randomly generate 3D vectors ;

[0204] Participants Randomly generate 3D vectors ;

[0205] Participants Randomly generate 3D vectors ;

[0206] (C4) Calculation = , for dimensional vector;

[0207] Participants calculate ;

[0208] Participants calculate ;

[0209] Participants calculate ;

[0210] Participants calculate .

[0211] (C5) Encryption , For participants A randomly selected integer ( ), Note: Perform encryption operations on each element in;

[0212] Participants encryption , select a random number ;

[0213] ;

[0214] Participants encryption , select a random number ;

[0215] ;

[0216] Participants encryption , select a random number ;

[0217] ;

[0218] Participants encryption , select a random number ;

[0219] ;

[0220] (C6) Participants Using the communication protocol and Sent to a semi-trusted third party.

[0221] (D) Data decryption, mainly including the following contents;

[0222] (D1) Demander R calculation result ciphertext and total noise parameters ;

[0223] ;

[0224] ;

[0225] (D2) decrypted separately, the demander will Sent to all participants, Using your own private key Decrypt the ciphertext And send the decrypted result to the demander;

[0226] Participants Using your own private key Decrypt the ciphertext, , and the decrypted result Send to the demander;

[0227] Participants Using your own private key Decrypt the ciphertext, , and the decrypted result Send to the demander;

[0228] Participants Using your own private key Decrypt the ciphertext, , and the decrypted result Send to the demander;

[0229] Participants Using your own private key Decrypt the ciphertext, , and the decrypted result Sent to the demander.

[0230] (E) Result combination, mainly including the following contents.

[0231] (E1) The demander combines the decryption results of each participant to obtain , and calculate ;

[0232] ;

[0233] ;

[0234] ; ;

[0235] ;

[0236] (E2) Calculation ;

[0237] .

[0238] After (A to E), a multi-party aggregation result that satisfies differential privacy is obtained.

[0239] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A multi-party secure aggregation method for classification tasks, characterized by: The steps include: Receive query data, public key, private key and Gaussian noise scale ; Perform inference generation on the query data Dimensional vector results , randomly generated dimensional vector , Indicates the Participants Will Dimensional vector results and dimensional vector Add up to get Model inference results of each participant ; Model inference results based on private key Encryption based on Gaussian noise scale Calculate the Gaussian noise parameters and dimensional vector With the The noise parameter is obtained by adding the Gaussian noise parameters of the participants ; The encrypted model inference results And the noise parameters Send to a semi-trusted third party; Obtain the ciphertext of the result of the demander's feedback , use the private key to encrypt the result Decrypt and decrypt the result ciphertext Send to a semi-trusted third party, and the demander will cipher the results feedback from each participant After combining and combining the total noise parameters Get a multi-party aggregation result that satisfies differential privacy ; Multi-party aggregation results The calculation process is as follows: in, represents the total inference result after decryption, represents the total noise parameter, represents an integer, Indicates that each participant has a ciphertext of the result The combined result after decryption, Indicates the Participants encrypt the result The result after decryption, Indicates the The noise parameters of the participants, Indicates the The private keys of each participant, represents the number of participants, Represents a Gaussian noise scale Calculate the Gaussian noise parameters, represents a semi-random number, express of Power, and Represents a custom function, , .

2. The multi-party secure aggregation method for classification tasks according to claim 1 is characterized in that: After receiving the query data, public key, private key and Gaussian noise scale In the example, the query data received is as follows: Each participant receives the data to be discussed and sent by the same demander using the communication protocol. The dimension vector is the result data you want to receive, and the data to be discussed and the label vector are the query data.

3. The multi-party secure aggregation method for classification tasks according to claim 2 is characterized in that: After receiving the query data, public key, private key and Gaussian noise scale In the example, the public key, private key, and Gaussian noise scale The generation process is as follows: Select an integer , calculate the private key and the public key ( , , ),in and are two strong prime numbers; Selecting a privacy budget and error ,based on , select the Gaussian noise scale , represents the parameterized differential privacy budget; Use the key sharing protocol to share the private key Decompose into , and and Assigned to participants .

4. The multi-party secure aggregation method for classification tasks according to claim 3 is characterized in that: The private key The calculation process is as follows: in,( ) represents a randomly selected number pair, Table semi-random number, represents the least common multiple, Represents the greatest common divisor.

5. The multi-party secure aggregation method for classification tasks according to claim 4 is characterized in that: Inference results based on the private key model Encryption is in progress. The encryption formula is as follows: in, Indicates the model inference result based on the private key The encryption result of Indicates the An integer randomly chosen by the participants.

6. The multi-party secure aggregation method for classification tasks according to claim 1, characterized in that: In the dimensional vector With the The noise parameter is obtained by adding the Gaussian noise parameters of the participants In the noise parameter The calculation formula is as follows: in, Indicates the Gaussian noise parameters of each participant.

7. The multi-party secure aggregation method for classification tasks according to claim 5, characterized in that: Obtaining the ciphertext of the result of feedback from the demand side , use the private key to encrypt the result Decryption in progress; The demander encrypts the result The calculation formula is as follows: in, Indicates the model inference result based on the private key The encryption result.

8. A multi-party secure aggregation system for classification tasks, characterized by: It includes a receiving module, an inference generation module, an addition module, an encryption calculation module, a sending module and a feedback acquisition module; The receiving module is used to receive query data, public key, private key and Gaussian noise scale ; The inference generation module is used to generate inferences on the query data Dimensional vector results , randomly generated dimensional vector , Indicates the Participants The adding module is used to Dimensional vector results and dimensional vector Add up to get Model inference results of each participant ; The encryption calculation module is used to infer the model results based on the private key Encryption based on Gaussian noise scale Calculate the Gaussian noise parameters and dimensional vector With the The noise parameter is obtained by adding the Gaussian noise parameters of the participants ; The sending module is used to send the encrypted model inference results And the noise parameters Send to a semi-trusted third party; The feedback acquisition module is used to obtain the result ciphertext of the demander's feedback , use the private key to encrypt the result Decrypt and decrypt the result ciphertext Send to a semi-trusted third party, and the demander will cipher the results feedback from each participant After combining and combining dimensional vector and the sum of the Gaussian noise parameters Get a multi-party aggregation result that satisfies differential privacy ; Multi-party aggregation results The calculation process is as follows: in, represents the total inference result after decryption, represents the total noise parameter, represents an integer, Indicates that each participant has a ciphertext of the result The combined result after decryption, Indicates the Participants encrypt the result The result after decryption, Indicates the The noise parameters of the participants, Indicates the The private keys of each participant, represents the number of participants, Represents a Gaussian noise scale Calculate the Gaussian noise parameters, represents a semi-random number, express of Power, and Represents a custom function, , .

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a plurality of programs, which are used to be called by a processor and execute the multi-party security aggregation method according to any one of claims 1 to 7.

Images