A model issuance system and method for model heterogeneous federated learning, a terminal and a medium

Abstract

The application provides a model issuing system and method for model heterogeneous federated learning, a terminal and a medium, and relates to the technical field of information security. The system comprises an authorization agency, a first server, a second server and a client. The authorization agency serves as a trusted third party and is responsible for performing system initialization work, such as generating attribute private keys based on a user attribute set and a partial keyword search trapdoor. The first server is responsible for model generation and encryption processing of the model and related data, and transmits all encrypted data and the encrypted model to the second server. The second server interacts with the client, realizes model search matching and partial decryption by using a complete keyword search trapdoor uploaded by the client, finally downloads the partial decryption result through the client, and completely decrypts the partial decryption result to obtain a target model required by the client. The application effectively solves the privacy threat existing in the model issuing process of heterogeneous federated learning, ensures privacy security and realizes accurate matching.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information security, and particularly relates to a model issuing system and method for model heterogeneous federated learning, a terminal and a medium. BACKGROUND

[0002] At present, model heterogeneous federated learning solves the demand of training different models by clients, but in the issuing of different models, the server needs to know some private user information to match a suitable model, which leads to the direct or indirect leakage of sensitive information of users. Therefore, how to realize accurate matching of models under the premise of protecting user privacy in the model issuing process of heterogeneous federated learning is a problem that technicians in the field need to pay attention to. Although existing technical solutions try to solve this problem, the security assumption in the existing technical solution is too high to be realized in the actual environment. SUMMARY

[0003] The technical problem to be solved by the present application is to provide a model issuing system and method for model heterogeneous federated learning, a terminal and a medium, aiming at solving the problem of direct or indirect leakage of sensitive information of users in the model issuing process of heterogeneous federated learning in the prior art.

[0004] The technical solution adopted by the present application to solve the technical problem is as follows:

[0005] In a first aspect, the present application discloses a model issuing system for model heterogeneous federated learning, wherein the system comprises an authorization agency, a first server, a second server and a client;

[0006] The authorization agency is configured to generate attribute private keys and partial keyword search traps for each user based on receiving a user attribute set submitted by the client;

[0007] The client is configured to generate a complete keyword search trap based on receiving the attribute private key and the partial keyword search trap transmitted by the authorization agency;

[0008] The first server is configured to generate models of various demands, and generate corresponding encryption keys for each model, encrypt the model by using the encryption key to obtain an encrypted model, and encrypt the encryption key and the model attribute set by using the security parameter sent by the authorization agency to obtain an encrypted encryption key and attribute set ciphertext, and send the encrypted model, the encrypted encryption key and the attribute set ciphertext to the second server;

[0009] A second server is configured to respond to a model search request initiated by a client, and perform a search matching on a model according to a complete keyword search trapdoor received from the client to determine whether a target model required by the client exists, when the target model exists, perform a partial decryption operation to obtain a partial decryption result, and return the partial decryption result to the client, so that the client performs a complete decryption on the partial decryption result by using the attribute private key and a blinding value stored by the client to obtain the target model.

[0010] Optionally, the authority is further configured to:

[0011] Randomly select a first prime number and a second prime number, and construct a corresponding finite field and a finite multiplicative field according to the first prime number and the second prime number;

[0012] Calculate a first generator by using a first preset generator calculation formula based on the first prime number, the second prime number, and a random number selected from the finite multiplicative field, and construct a cyclic group of the first prime number based on the first generator;

[0013] Select a minimum embedding number satisfying a preset condition, and determine a new second prime number based on the minimum embedding number and the second prime number;

[0014] Construct a new finite field and a new finite multiplicative field corresponding to the new second prime number based on the finite field and the finite multiplicative field;

[0015] Calculate a second generator by using a second preset generator calculation formula based on the minimum embedding number, the first prime number, the new second prime number, and a random number selected from the new finite field, and construct a cyclic group of the new second prime number based on the second generator;

[0016] Define a corresponding bilinear mapping rule based on the cyclic group of the first prime number and the cyclic group of the new second prime number;

[0017] Construct a master key based on a plurality of random numbers selected from the finite field;

[0018] Construct the security parameter by using the master key based on the bilinear mapping rule and a pre-defined hash function;

[0019] Wherein, q << p, and q | p-1, q represents the first prime number, and p represents the second prime number;

[0020] The first preset generator calculation formula is: g = f1 (p-1) / q mod p and g ≠ 1, g represents the first generator, f1 represents a random number selected from the finite multiplicative field, and the cyclic group of the first prime number is: Fp * denotes a finite multiplication field;

[0021] The preset condition is: q | (p k -1), k represents a minimum embedding number, p k denotes a new second prime number, and a new finite field corresponding to the new second prime number is: x i denotes the i-th unknown number;

[0022] The second preset generator calculation formula is: And g1≠1, g1 represents a second generator, f2 represents a random number selected from the new finite field, and a cyclic group of the new second prime number is: denotes a finite multiplication field corresponding to the new second prime number;

[0023] The bilinear mapping rule is: e: G x G→ G T , and for any a, b∈F p , e(g a , g b ) = e(g, g) ab , F p denotes a finite field;

[0024] A plurality of random numbers selected from the finite field are: α, β, t1, t2, t3, t4, and the master key is: MSK = {α, β, t1, t2, t3, t4};

[0025] The pre-defined hash function is: H(x))g h(x)mod p ;

[0026] The security parameter is:

[0027] Optionally, the authorization agency comprises:

[0028] A private key component unit configured to generate a private key component corresponding to each user based on a user attribute set submitted by a client, by using the master key, the security parameter, and a randomly selected random number;

[0029] An attribute private key construction unit configured to construct an attribute private key of each user according to the private key component and the random number;

[0030] An access structure generation unit configured to generate a user access structure corresponding to each user attribute in the attribute set;

[0031] A partial search trapdoor generation unit configured to generate a partial keyword search trapdoor based on the user access structure and a plurality of new random numbers randomly selected from the finite field.

[0032] Optionally, the client comprises:

[0033] a blinding unit configured to blind the attribute private key transmitted by the authority by using the blinding value to obtain a blinded attribute private key;

[0034] a complete search trapdoor generation unit configured to construct a complete keyword search trapdoor according to the blinded attribute private key and the partial keyword search trapdoor.

[0035] Optionally, the first server comprises:

[0036] an encryption key generation unit configured to generate a corresponding encryption key for each model based on an AES encryption algorithm.

[0037] Optionally, the first server comprises:

[0038] a flattening unit configured to perform flattening processing on a model weight matrix of the model to obtain a corresponding flattening result;

[0039] a numerical value selection unit configured to select a set of numerical values satisfying a preset mapping rule to obtain a corresponding numerical set;

[0040] a weighted summation unit configured to perform weighted summation processing on the flattening result by using the numerical set to obtain a weighted summation result;

[0041] a model encryption unit configured to encrypt the weighted summation result by using the encryption key to obtain an encrypted model;

[0042] wherein the flattening processing process is:

[0043] the weighted summation processing process is:

[0044] and W v represents a model weight matrix, w v represents a flattening result, W′ v represents a weighted summation result, represents a numerical set;

[0045] the preset mapping rule is that a given vector length n vector u is obtained, a random positive integer Q, E and a random number vector θ=[θ1,..., θ n ] are obtained, and from θ2, and θ i is a prime number, u′=θ1u1+...+θ n u n .

[0046] Optionally, the first server comprises:

[0047] A key encryption unit configured to calculate a key ciphertext corresponding to the encryption key and a component of a decryption key corresponding to the encryption key by using a security parameter sent by the authorization agency, to obtain an encrypted encryption key;

[0048] A binary conversion unit configured to convert the model attribute set into a binary form to obtain a binary-form model attribute set;

[0049] An attribute set encryption unit configured to calculate a ciphertext of the binary-form model attribute set and a component thereof by using the security parameter sent by the authorization agency, to obtain a target attribute set ciphertext.

[0050] In a second aspect, the present application further discloses a model issuing method for model heterogeneous federated learning, wherein the method is applied to the model issuing system for model heterogeneous federated learning as described above, and the method comprises:

[0051] Generating, by the authorization agency, an attribute private key and a partial keyword search trapdoor of each user based on the user attribute set submitted by the client;

[0052] Generating, by the client, a complete keyword search trapdoor based on the attribute private key and the partial keyword search trapdoor transmitted by the authorization agency;

[0053] Generating, by the first server, models of various requirements, and generating a corresponding encryption key for each model, encrypting the model by using the encryption key to obtain an encrypted model, and encrypting the encryption key and the model attribute set by using a security parameter sent by the authorization agency to obtain an encrypted encryption key and an attribute set ciphertext, and sending the encrypted model, the encrypted encryption key and the attribute set ciphertext to the second server;

[0054] Responding, by the second server, to the model search request initiated by the client, and performing model search matching according to the complete keyword search trapdoor sent by the client to determine whether there is a target model required by the client, when the target model exists, performing a partial decryption operation to obtain a partial decryption result, and returning the partial decryption result to the client, so that the client performs complete decryption on the partial decryption result by using the attribute private key and a blind value stored by itself to obtain the target model.

[0055] In a third aspect, the present application discloses a terminal, comprising: a memory, a processor, and a model distribution program for model heterogeneous federated learning stored in the memory and capable of running on the processor, wherein the model distribution program for model heterogeneous federated learning, when executed by the processor, implements the steps of the model distribution method for model heterogeneous federated learning.

[0056] In a fourth aspect, the present application discloses a computer readable storage medium, wherein the computer readable storage medium stores a computer program capable of being executed to implement the steps of the model distribution method for model heterogeneous federated learning.

[0057] The application provides a model issuing system and method for model heterogeneous federated learning, a terminal and a medium. The model issuing system for model heterogeneous federated learning comprises an authorization mechanism, a first server, a second server and a client. The authorization mechanism is configured to generate attribute private keys and partial keyword search traps for each user based on a user attribute set submitted by the client. The client is configured to generate a complete keyword search trap based on the attribute private keys and the partial keyword search traps transmitted by the authorization mechanism. The first server is configured to generate models of various requirements, generate a corresponding encryption key for each model, encrypt the model by using the encryption key to obtain an encrypted model, encrypt the encryption key and a model attribute set by using a security parameter transmitted by the authorization mechanism to obtain an encrypted encryption key and attribute set ciphertext, and transmit the encrypted model, the encrypted encryption key and the attribute set ciphertext to the second server. The second server is configured to respond to a model search request initiated by the client, search and match the model according to the complete keyword search trap transmitted by the client to determine whether there is a target model required by the client, perform a partial decryption operation to obtain a partial decryption result when the target model exists, and return the partial decryption result to the client, so that the client completely decrypts the partial decryption result by using the attribute private key and a blinding value stored by the client to obtain the target model. Therefore, the model issuing system for model heterogeneous federated learning can effectively solve the privacy threat existing in the model issuing process of heterogeneous federated learning, ensure privacy security and realize accurate matching. The authorization mechanism in the model matching system acts as a trusted third party and is responsible for performing system initialization work, such as generating attribute private keys and partial keyword search traps based on a user attribute set. The first server in the model matching system is responsible for model generation and encryption processing of the model and related data, and transmits all encrypted data and encrypted models to the second server. The second server in the model matching system interacts with the client in the model matching system, realizes model search matching and partial decryption by using the complete keyword search trap uploaded by the client, and finally downloads the partial decryption result by the client and completely decrypts it to obtain the target model required by the client. That is, the model issuing system for model heterogeneous federated learning completes the model issuing process of heterogeneous federated learning, provides a fine-grained access control mechanism for the client by using a trap search strategy, realizes accurate matching of the model, can remove the security dependence on the edge server, and can effectively solve the privacy threat existing in the model issuing process. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 is a functional principle block diagram of a preferred embodiment of the model issuing system for model heterogeneous federated learning in the application.

[0059] Figure 2 is a specific model-oriented model heterogeneous federated learning model issuing system workflow schematic diagram disclosed by the present application;

[0060] Figure 3 is a flowchart of a preferred embodiment of the model issuing method for model heterogeneous federated learning in the present application;

[0061] Figure 4 is a functional principle block diagram of a preferred embodiment of the terminal in the present application. DETAILED DESCRIPTION

[0062] In order to make the purpose, technical scheme and advantages of the present application clearer and more explicit, the present application will be further described in detail below with reference to the drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0063] Currently, model heterogeneous federated learning solves the demand of training different models by clients, but in the issuing of different models, the server needs to know some private user information to match a suitable model, which leads to the direct or indirect leakage of sensitive information of users. Therefore, how to achieve accurate matching of models under the premise of protecting user privacy in the model issuing process of heterogeneous federated learning is a problem that technicians in the field need to pay attention to. Although existing technical solutions try to solve this problem, the security assumption requirements in existing technical solutions are too high to be implemented in actual environment.

[0064] For example, the method of rolling sub-model mentioned in the prior art requires the client to share information including capacity parameters with the server in order to customize sub-models with different functions and demands. Therefore, it is particularly important to protect this part of additional information that needs to be shared in heterogeneous federated learning, and the matching problem of users and models needs to be considered at the same time as the security of this part of information is protected. For example, in terms of privacy protection, the existing scheme adopts distributed architecture and subspace alignment technology to prevent the server from colluding to leak sub-model information through information theory security mechanism, or protects sub-model information based on private information retrieval technology, but both of these two schemes need to assume that the edge server does not collude, and in the model issuing stage, a transmission mechanism based on aggregation protocol is adopted, that is, the sub-model is first sent to the edge server, and then distributed to the client by the edge server. However, this scheme lacks cryptographic protection measures in the communication process between the edge server and the client, and only relies on the honesty assumption of the edge device, thereby existing security limitations in the trustworthiness of the server.

[0065] To this end, the application provides a model issuing system for model heterogeneous federated learning, which can effectively solve the privacy threat existing in the model issuing process of heterogeneous federated learning, and ensure privacy security and precise matching.

[0066] See Figure 1 , Figure 1 is the functional principle diagram of the model issuing system for model heterogeneous federated learning in the application. As shown in Figure 1 , the model issuing system for model heterogeneous federated learning in the embodiment of the application comprises an authorization agency S1, a first server 1, a second server 2 and a client K1.

[0067] The authorization agency S1 is configured to generate attribute private keys of each user and partial keyword search trapdoors based on the user attribute set submitted by the client K1.

[0068] The client K1 is configured to generate complete keyword search trapdoors based on the attribute private keys and the partial keyword search trapdoors transmitted by the authorization agency S1.

[0069] The first server 1 is configured to generate models of various requirements, generate corresponding encryption keys for each model, encrypt the models by using the encryption keys to obtain encrypted models, encrypt the encryption keys and model attribute sets by using the security parameters sent by the authorization agency S1 to obtain encrypted encryption keys and attribute set ciphertexts, and send the encrypted models, the encrypted encryption keys and the attribute set ciphertexts to the second server 2.

[0070] The second server 2 is configured to respond to a model search request initiated by the client K1, and perform model search matching according to the complete keyword search trapdoors sent by the client K1 to determine whether there is a target model required by the client, perform partial decryption when the target model exists to obtain a partial decryption result, and return the partial decryption result to the client K1, so that the client K1 fully decrypts the partial decryption result by using the attribute private keys and the blind value stored by itself to obtain the target model.

[0071] In some specific embodiments, the authorization agency S1 can be further configured to:

[0072] Randomly select a first prime number and a second prime number, and construct a corresponding finite field and a finite multiplicative field according to the first prime number and the second prime number.

[0073] Calculate a first generator by using a first preset generator calculation formula based on the first prime number, the second prime number and a random number selected from the finite multiplicative field, and construct a cyclic group of the first prime number based on the first generator.

[0074] selecting a minimum embedding number satisfying a preset condition, and determining a new second prime number based on the minimum embedding number and the second prime number.

[0075] constructing a new finite field and a new finite multiplicative field corresponding to the new second prime number based on the finite field and the finite multiplicative field.

[0076] calculating a second generator based on the minimum embedding number, the first prime number, the new second prime number, and a random number selected from the new finite field, and a second preset generator calculation formula, and constructing a cyclic group of the new second prime number based on the second generator.

[0077] defining a corresponding bilinear mapping rule based on the cyclic group of the first prime number and the cyclic group of the new second prime number.

[0078] constructing a master key based on a plurality of random numbers selected from the finite field.

[0079] constructing the security parameter based on the bilinear mapping rule and a pre-defined hash function, and using the master key.

[0080] wherein q << p, and q | p-1, q represents the first prime number, and p represents the second prime number;

[0081] the first preset generator calculation formula is: g = f1 (p-1) / q mod p and g ≠ 1, g represents the first generator, f1 represents a random number selected from the finite multiplicative field, and the cyclic group of the first prime number is: F p * representing a finite multiplicative field.

[0082] the preset condition is: q | (p k -1), k represents the minimum embedding number, p k represents the new second prime number, and the new finite field corresponding to the new second prime number is: x i represents the ith unknown number.

[0083] the second preset generator calculation formula is: and g1 ≠ 1, g1 represents the second generator, f2 represents a random number selected from the new finite field, and the cyclic group of the new second prime number is: representing a finite multiplicative field corresponding to the new second prime number.

[0084] the bilinear mapping rule is: e: G × G → G T , and for any a, b ∈ Fp , meet e(g a , g b ) = e(g, g) ab , F p denotes a finite field.

[0085] A plurality of random numbers selected from the finite field are: α, β, t1, t2, t3, t4, and the master key is: MSK = {α, β, t1, t2, t3, t4}.

[0086] The pre-defined hash function is: H(x)) = g h(x)mod p .

[0087] The security parameter is:

[0088] It can be understood that the authorization agency S1, as a trusted third party, is responsible for performing system initialization, generating a security parameter and a master key, so as to realize the process of in the initialization phase, and also generates an attribute private key based on a user attribute set and a partial keyword search threshold value. In the initialization phase, first, the authorization agency S1 receives a user attribute set submitted by a user through a client K1, and the user attribute set includes a plurality of user attributes, and then the authorization agency S1 generates a corresponding attribute private key and a partial keyword search trapdoor for the user.

[0089] In some specific embodiments, the authorization agency S1 can specifically include:

[0090] A private key component unit configured to generate a private key component corresponding to each user based on the master key, the security parameter, and a randomly selected random number, and based on a user attribute set submitted by a client K1.

[0091] An attribute private key construction unit configured to construct an attribute private key of each user according to the private key component and the random number.

[0092] An access structure generation unit configured to generate a user access structure corresponding to each user attribute in the attribute set.

[0093] A partial search trapdoor generation unit configured to generate a partial keyword search trapdoor based on the user access structure and a plurality of new random numbers randomly selected from the finite field.

[0094] For example, k v users upload their attribute sets to the authorization agency S1 through the client K1, and the attribute set of a user j is: A random value r is selected by the authorization agency S1, and a private key component corresponding to the user j is calculated, that is:

[0095] D1 = gβ g αr ;

[0096] D2=g r ;

[0097]

[0098] wherein H1(x) is the same as a predefined hash function H(x)=g h(x)mod p .

[0099] Further, the attribute private key of the user j is constructed according to the private key component and the random number, i.e., SK j ={r, D1, D2, D x}.

[0100] Meanwhile, the authority S1 generates an access structure for each user based on the user attribute of the user, i.e., P={N, Π(x), W Π(x)}, so as to complete the process in the user private key generation phase , and the access structure and the attribute private key are issued to each client K1. Wherein N represents the access matrix of the user, Π(x) represents a mapping function for mapping to a specific attribute name, and W Π(x) represents a mapping function for mapping the attribute name to a specific attribute value.

[0101] Then in the trap generation phase, the process of Trapgen(SK, P)→(T) is completed, i.e., the authority S1 randomly selects a plurality of random numbers, and the plurality of random numbers contain α in the master key, such as α, z2, z3... α1, α2∈F P , and then the random numbers are processed, i.e., the vector is obtained. Then σ i =N i *θ(i=1, 2..l0) is calculated, which means that α in the vector θ and the attribute (each Ni corresponds to an attribute) are bound, and then a partial search key trapdoor is generated, i.e.:

[0102] wherein,

[0103]

[0104] wherein H2(x) is the same as a predefined hash function H(x)=g (x)mod p .

[0105] In some specific embodiments, the client K1 can specifically include:

[0106] A blinding unit is configured to blind the attribute private key in the S1 transmission from the authority by using the blinding value to obtain a blinded attribute private key.

[0107] A complete search trapdoor generation unit is configured to construct a complete keyword search trapdoor according to the blinded attribute private key and the partial keyword search trapdoor.

[0108] For example, a user receives a trapdoor component of the authority S1 through the client K1, and randomly selects a blinding value ζ to blind D2 and D x Blinding processing is performed, that is,

[0109] D2' = D2 ζ = g rζ ;

[0110]

[0111] Finally, a complete keyword search trapdoor is formed according to D1 in the attribute private key, D2' and D x ' obtained by blinding processing, and the partial search keyword trapdoor, that is,

[0112] In some specific embodiments, the first server 1 can specifically include:

[0113] An encryption key generation unit is configured to generate a corresponding encryption key for each model based on an AES encryption algorithm.

[0114] It can be understood that the AES encryption algorithm can strictly guarantee that the privacy information of the user is not leaked, and the encryption key of the AES encryption algorithm is the same as the decryption key.

[0115] In some specific embodiments, the first server 1 can specifically include:

[0116] A flattening unit is configured to perform flattening processing on a model weight matrix of the model to obtain a corresponding flattening result.

[0117] A numerical value selection unit is configured to select a set of numerical values satisfying a preset mapping rule to obtain a corresponding number set.

[0118] A weighted summation unit is configured to perform weighted summation processing on the flattening result by using the number set to obtain a weighted summation result.

[0119] A model encryption unit is configured to encrypt the weighted summation result by using the encryption key to obtain an encrypted model.

[0120] The flattening processing process is as follows:

[0121] The weighted summation process is as follows:

[0122] Furthermore, W v w represents the model weight matrix. v W′ represents the result of flattening. v This represents the weighted summation result. Represents a set of numbers.

[0123] The preset mapping rule is to obtain random positive integers Q and E, and a random number vector θ = [θ1, ... θ2] given a vector u of length n. n ], and starting from θ2, and θ i If the number is a prime number, then u′=θ1u1+...+θ n U n .

[0124] For example, encryption processing based on the AES encryption algorithm, i.e.: C s =AES(key,W′) v );

[0125] Among them, C s This is the encrypted model, and key is the encryption key.

[0126] In some specific embodiments, the first server 1 may specifically include:

[0127] The key encryption unit is used to calculate the ciphertext of the encryption key and the decryption key corresponding to the encryption key using the security parameters sent by the authorized agency S1, so as to obtain the corresponding encrypted encryption key.

[0128] The binary conversion unit is used to convert the model attribute set into a binary representation to obtain the model attribute set in binary form.

[0129] The attribute set encryption unit is used to calculate the ciphertext and its components of the binary form of the model attribute set using the security parameters sent by the authorized agency S1, so as to obtain the target attribute set ciphertext.

[0130] For example, selecting random numbers s1, s2, s3∈F P , forming a vector Then, the key ciphertext is calculated using the security parameters sent by the authorized agency S1, i.e.: C = key·e(g, g) sβ Then based on the model attribute set ψ v Generate the access structure of the model, i.e., A = (M, ρ(x)), where M represents the access matrix of the model, and calculate λ.i = M i * μ(i = 1...l v ) and then using the security parameters sent by the authority S1 to calculate the components of the decryption key corresponding to the encryption key, i.e. C s , Thus, the encrypted encryption key is: C ek = {C, C i1 , C i2 , C i}. Where ρ(x) represents a mapping function used to map M v to a specific attribute name.

[0131] For example, the model attribute set ψ v is represented in binary form as ψ' Π(i) , and the security parameters sent by the authority S1 are used to calculate the ciphertext of the model attribute set and its components, i.e.

[0132]

[0133] Thus, the ciphertext of the target attribute set is:

[0134] The encryption data obtained by the first server 1, i.e. , is sent to the second server 2, i.e. through the first server 1 to complete the encryption phase This process.

[0135] And in the second server 2, the precise matching of the model is completed through the trapdoor search strategy, for example, the second server 2 responds to the model search request initiated by the client K1, and the user uploads the complete keyword search trapdoor generated by the client K1 to the second server 2, and in the second server 2, the smallest subset I0 that satisfies the trapdoor search strategy is calculated, i.e. there is a set of coefficients v that satisfies Then verify whether the ciphertext satisfies the trapdoor search strategy, i.e.

[0136]

[0137] If h(W Π(i) ) = h(ψ' i ), and then it means that the above equation is true, indicating that the second server 2 has the target model that the user needs, so partial decryption is performed in the second server 2, if decryption is successful and Z is obtained, it means that the user is allowed to download the ciphertext through the client K1, otherwise return ⊥, i.e. through the second server in the model search phase to complete Search(PK, T, C Fv)→(Z / ⊥) this process, to achieve accurate model matching. For example, the minimum I s , there is a set of coefficients γ, satisfying Then the partial decryption operation is performed to obtain the partial decryption result Z, that is:

[0138]

[0139] Then, the second server 2 transmits the partial decryption result to the client, and the client K1 completes the decryption algorithm, that is, Decrypt(PK, SK, Z)→(W v ) this decryption process, that is, the client uses the attribute private key and the unique blinding value it holds to finally decrypt the partial decryption result to obtain the weight matrix, that is:

[0140]

[0141] W′ v = DEC(key, C s ).

[0142] Then, W′ v is mapped to the vector w v , and then restored to the weight matrix W v , completing the model distribution process of heterogeneous federated learning.

[0143] It should be pointed out that the first server 1 completes the process of Tran(W v )→(h v ) in the model training stage, that is, the first server 1 locally trains to obtain h v is obtained based on LWE noise disturbance, and then the second server 2 completes the process of in the aggregation stage, that is, the second server 2 receives h v from the client marked with the same model type, and then transmits it to the first server 1, and the first server 1 aggregates and calculates according to the identifier.

[0144] It should also be noted that in the LSSS access matrix, the Boolean formula AA∨∧BB∨∧CC defining access control is conjunctive ∧, the left child node is defined as root||1, and the right child node is defined as (0, 0,..., 0)||-1, (0, 0,..., 0) represents a zero vector of the same length as the parent node; When it is disjunctive ∨, both left and right nodes inherit the value of root; When some attributes are not long enough, 0 is appended at the end; Get the vector representation of all attributes, and splice to get the LSSS matrix, AA, BB and BC represent attributes, and AA∨∧BB∨∧CC represents a Boolean expression of an attribute.

[0145] As can be seen, in this embodiment of the invention, the model delivery system for heterogeneous federated learning can effectively solve the privacy threats in the model delivery process of heterogeneous federated learning, ensuring both privacy and security while achieving accurate matching. Specifically, an authorized institution in the model matching system acts as a trusted third party, responsible for performing system initialization tasks, such as generating attribute private keys based on user attribute sets and partial keyword search trapdoors. A first server in the model matching system is responsible for model generation and encryption of the model and related data, transmitting all encrypted data and the encrypted model to a second server. The second server interacts with the client in the model matching system, using the complete keyword search trapdoor uploaded by the client to achieve model search matching and partial decryption. Finally, the client downloads the partial decryption results and fully decrypts them to obtain the target model required by the client. In other words, the model delivery process for heterogeneous federated learning is completed through the model delivery system for heterogeneous federated learning proposed in this application. The trapdoor search strategy provides a fine-grained access control mechanism for the client, achieving accurate model matching while eliminating security dependencies on edge servers, effectively solving the privacy threats in model delivery.

[0146] For example, see Figure 2 As shown, in a model distribution system for heterogeneous federated learning, the authorizing agency S1 performs system initialization, generates security parameters PK and master key MSK, and implements the initialization phase. The process, and then based on the user attribute set uploaded by the client. The user private key generation phase has been completed. The process involves distributing the user access structure P and the attribute private key SK to each client K1. During the trapdoor generation phase, each client completes the process Trapgen(SK, P) → (T) to generate the complete keyword search trapdoor T. Finally, the encryption phase is completed by the first server 1. This process involves using security parameters sent by the authorized agency S1 to complete the relevant encryption processing and obtain encrypted data. Encrypt data The data is sent to the second server 2, which then completes the model search phase. This process involves the second server 2 responding to the model search request initiated by client K1, performing a model search and matching based on the complete keyword search trap T uploaded by client K1 to determine whether the target model required by the client exists. When the target model exists, the encrypted data is then processed. The partial decryption operation is performed to obtain a partial decryption result Z, and the partial decryption result Z is returned to the client, and the client K1 completes the decryption process of Decrypt(PK, SK, Z)→(W v ) in the final decryption stage, thereby completing the model distribution of the heterogeneous federated learning. In the present application, the efficiency of model weight matrix encryption is improved by combining the prime number encoding mapping method and attribute-based encryption. In the model encryption stage, the weight parameters are mapped into an integer, which improves the efficiency of the model encryption stage and reduces the length of the ciphertext, i.e., improves the efficiency of the weight encryption, and effectively solves the problem of excessive calculation load of attribute-based encryption caused by the excessively large model weight matrix.

[0147] In one embodiment, as shown in Figure 3 The present application also provides a model distribution method for model heterogeneous federated learning, which is applied to the model distribution system for model heterogeneous federated learning as described above. The method comprises the following steps:

[0148] In step S11, the attribute private key and the partial keyword search trapdoor of each user are generated by the authorization agency based on the user attribute set submitted by the client.

[0149] In step S12, the complete keyword search trapdoor is generated by the client based on the attribute private key and the partial keyword search trapdoor transmitted by the authorization agency.

[0150] In step S13, the first server generates models of various demands, and generates corresponding encryption keys for each model. The model is encrypted by using the encryption key to obtain an encrypted model. The encryption key and the model attribute set are encrypted by using the security parameters sent by the authorization agency to obtain encrypted encryption key and attribute set ciphertext. The encrypted model, the encrypted encryption key, and the attribute set ciphertext are sent to the second server.

[0151] In step S14, the second server responds to the model search request initiated by the client, and performs model search matching according to the complete keyword search trapdoor sent by the client to determine whether the target model required by the client exists. When the target model exists, a partial decryption operation is performed to obtain a partial decryption result, and the partial decryption result is returned to the client. The client uses the attribute private key and the blind value stored by itself to completely decrypt the partial decryption result to obtain the target model.

[0152] In addition, it is worth noting that the working process of the model issuing method for model heterogeneous federated learning provided by the embodiment is the same as the working process of the model issuing system for model heterogeneous federated learning described above, and will not be repeated here. For details, please refer to the working process of the model issuing system for model heterogeneous federated learning described above.

[0153] Figure 4 The terminal provided by the embodiment of the present application is shown in the structural diagram. The terminal can include:

[0154] The memory 501, the processor 502, and the computer program stored in the memory 501 and executable on the processor 502.

[0155] The processor 502 implements the model issuing method for model heterogeneous federated learning provided in the above embodiments when executing the program.

[0156] Further, the terminal further includes:

[0157] The communication interface 503 is used for communication between the memory 501 and the processor 502.

[0158] The memory 501 is used to store the computer program executable on the processor 502.

[0159] The memory 501 can include a high-speed RAM memory, and can also include a non-volatile memory, for example, at least one disk memory.

[0160] If the memory 501, the processor 502, and the communication interface 503 are independently implemented, the communication interface 503, the memory 501, and the processor 502 can be connected to each other through a bus and complete communication between each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one line is shown in the figure, but it does not mean that there is only one bus or only one type of bus.

[0161] Optionally, in specific implementation, if the memory 501, the processor 502, and the communication interface 503 are integrated on a chip, the memory 501, the processor 502, and the communication interface 503 can complete communication between each other through an internal interface.

[0162] The processor 502 can be a central processing unit (CPU), or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement one or more embodiments of the present application.

[0163] The embodiment further provides a computer readable storage medium, which stores a computer program. The computer program is executed by a processor to implement the model issuing method for model heterogeneous federated learning as above.

[0164] Other embodiments of the application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. It is intended that the specification and examples be considered as exemplary only, with the true scope and spirit of the application being indicated by the following claims.

[0165] It will be understood that the application is not limited to the examples described above. Modifications or alterations can occur to those skilled in the art upon reading the preceding specification and construction of the application disclosed herein is intended to cover all such modifications or alterations.

Claims

1. A model-oriented model heterogeneous federated learning model distribution system, characterized in that, The authorization authority, the first server, the second server and the client are included; The authorization authority is configured to generate attribute private keys and partial keyword search traps for each user based on the received user attribute set submitted by the client; The client is configured to generate a complete keyword search trap based on the received attribute private key and the partial keyword search trap transmitted by the authorization authority; The first server is configured to generate models of various requirements, generate a corresponding encryption key for each model, encrypt the model by using the encryption key to obtain an encrypted model, encrypt the encryption key and the model attribute set by using the security parameter sent by the authorization authority to obtain an encrypted encryption key and attribute set ciphertext, and send the encrypted model, the encrypted encryption key and the attribute set ciphertext to the second server; The second server is configured to respond to a model search request initiated by the client, and perform model search matching according to the received complete keyword search trap to determine whether there is a target model required by the client, and when the target model exists, perform a partial decryption operation to obtain a partial decryption result, and return the partial decryption result to the client, so that the client completely decrypts the partial decryption result by using the attribute private key and a blind value stored by the client to obtain the target model; The partial keyword search trap is: ， ； wherein ; ； ； ； The complete keyword search trap is: ； ； ； ； ； ； The encrypted model is: ; The security parameter is: ; The encrypted encryption key is: ; Key ciphertext is: ; The attribute set ciphertext is: ; ； ； ； ； ； ； ； ； The partial decryption result is: ； The complete decryption includes: ； ； wherein, represents a first generation element, represents a plurality of random numbers selected from a finite field, represents a mapping function for mapping to a specific attribute name, represents a mapping function for mapping an attribute name to a specific attribute value, represents a plurality of random numbers selected from a finite field, represents a finite field, represents a vector in which and attribute binding, wherein each N i corresponds to an attribute, and are the same as the pre-defined hash function D1, D2 and D x represent a private key component, and ζ represents a blinding value, and represent a private key component after blinding processing, and r represents a random value, represents an attribute set of user j, and key is an encryption key, represents a model weight matrix, represents a flattened result obtained by performing flattening processing on the model weight matrix, represents a weighted sum result obtained by performing weighted sum processing on the flattened result, represents a random number, represents a model attribute set in binary form, , represents a vector based on the random number M represents an access matrix of the model, represents a mapping function for mapping to an attribute name, represents a set of coefficients satisfying , and represents a minimum subset satisfying the trapdoor search strategy of the ciphertext.

2. The model-oriented heterogeneous federated learning model issuing system according to claim 1, characterized in that, The authorization authority is further configured to: randomly select a first prime number and a second prime number, and construct a corresponding finite field and a finite multiplication field according to the first prime number and the second prime number; calculate a first generator by using a first preset generator calculation formula based on the first prime number, the second prime number and a random number selected from the finite multiplication field, and construct a cyclic group of the first prime number based on the first generator; select a minimum embedding number satisfying a preset condition, and determine a new second prime number based on the minimum embedding number and the second prime number; construct a new finite field and a new finite multiplication field corresponding to the new second prime number based on the finite field and the finite multiplication field; calculate a second generator by using a second preset generator calculation formula based on the minimum embedding number, the first prime number, the new second prime number and a random number selected from the new finite field, and construct a cyclic group of the new second prime number based on the second generator; define a corresponding bilinear mapping rule based on the cyclic group of the first prime number and the cyclic group of the new second prime number; construct a master key based on a plurality of random numbers selected from the finite field; construct the security parameter based on the bilinear mapping rule and a pre-defined hash function, and use the master key; wherein , and satisfies q represents a first prime number, and p represents a second prime number; The first preset generating element calculation formula is: , denotes a first generating element, denotes a random number selected from the finite multiplicative field, and the cyclic group of the first prime number is: , denotes a finite multiplicative field; The preset condition is: k represents the minimum embedding times, A new second prime number, and a new finite field corresponding to the new second prime number is: , Xi represents the i-th unknown number; The second preset generating element calculation formula is: , represents a second generating element, represents a random number selected from the new finite field, and the cyclic group of the new second prime number is: , represents a finite multiplicative field corresponding to the new second prime number; The bilinear mapping rule is: , and for any , satisfies , denotes a finite field; A plurality of random numbers selected from the finite field are: , the master key is: .

3. The model-oriented heterogeneous federated learning model issuing system according to claim 2, characterized in that, The authorization authority includes: A private key component unit is configured to generate a private key component corresponding to each user based on the received user attribute set submitted by the client by using the master key, the security parameter and a randomly selected random number; An attribute private key construction unit is configured to construct an attribute private key for each user according to the private key component and the random number; An access structure generating unit is configured to generate a user access structure corresponding to each user attribute in the attribute set; A partial search trapdoor generating unit is configured to generate a partial keyword search trapdoor based on the user access structure and a plurality of new random numbers randomly selected from the finite field.

4. The model-oriented heterogeneous federated learning model issuing system according to claim 3, characterized in that, The client comprises: A blinding unit is configured to blind the attribute private key transmitted by the authority using the blinding value to obtain a blinded attribute private key; A complete search trapdoor generating unit is configured to construct a complete keyword search trapdoor according to the blinded attribute private key and the partial keyword search trapdoor.

5. The model-oriented heterogeneous federated learning model issuing system according to claim 1, characterized in that, The first server comprises: An encryption key generating unit is configured to generate a corresponding encryption key for each model based on an AES encryption algorithm.

6. The model-oriented heterogeneous federated learning model issuing system according to claim 5, characterized in that, The first server comprises: A flattening unit is configured to perform flattening processing on a model weight matrix of the model to obtain a corresponding flattening result; A numerical value selecting unit is configured to select a set of numerical values satisfying a preset mapping rule to obtain a corresponding numerical set; A weighted summation unit is configured to perform weighted summation processing on the flattening result using the numerical set to obtain a weighted summation result; A model encryption unit is configured to encrypt the weighted summation result using the encryption key to obtain an encrypted model. In the formula, the flattening processing process is: ; The process of the weighted sum processing is: ; And, denotes a model weight matrix, denotes a flattening result, denotes a weighted sum result, denotes a number set; The preset mapping rule is that for a given vector u with a vector length n, a random positive integer Q, E and a random number vector are obtained, and from , , and , is a prime number, and is obtained.

7. The model-oriented heterogeneous federated learning model issuing system according to claim 1, characterized in that, The first server comprises: A key encryption unit is configured to calculate a key ciphertext corresponding to the encryption key and a component of a decryption key corresponding to the encryption key using a security parameter sent by the authority to obtain an encrypted encryption key; A binary conversion unit is configured to convert a model attribute set into a binary form to obtain a binary-form model attribute set; An attribute set encryption unit is configured to calculate a ciphertext of the binary-form model attribute set and a component thereof using the security parameter sent by the authority to obtain a target attribute set ciphertext.

8. A model distribution method for model heterogeneous federated learning, characterized in that, The method applied to the model-oriented heterogeneous federated learning model distribution system of any one of claims 1 to 7, the method comprising: Generating, by the authority, an attribute private key and a partial keyword search trapdoor for each user based on the user attribute set submitted by the client; Generating, by the client, a complete keyword search trapdoor based on the attribute private key and the partial keyword search trapdoor transmitted by the authority; Generating, by the first server, models of various requirements, and generating a corresponding encryption key for each model, encrypting the model using the encryption key to obtain an encrypted model, and encrypting the encryption key and the model attribute set using a security parameter sent by the authority to obtain an encrypted encryption key and an attribute set ciphertext, and sending the encrypted model, the encrypted encryption key, and the attribute set ciphertext to the second server; The second server responds to the model search request initiated by the client, and performs model search matching according to the complete keyword search trapdoor sent by the client to determine whether the target model required by the client exists. When the target model exists, a partial decryption operation is performed to obtain a partial decryption result, and the partial decryption result is returned to the client. The client uses the attribute private key and the blind value stored by itself to completely decrypt the partial decryption result to obtain the target model.

9. A terminal, characterized by comprising: Comprise: A memory, a processor, and a model issuance program for model heterogeneous federated learning stored on the memory and executable on the processor, the model issuance program for model heterogeneous federated learning, when executed by the processor, implements the steps of the model issuance method for model heterogeneous federated learning as claimed in claim 8.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program capable of being executed for implementing the steps of the model issuance method for model heterogeneous federated learning as claimed in claim 8.

Images