CN115277015A - Asynchronous federal learning privacy protection method, system, medium, equipment and terminal - Google Patents

Asynchronous federal learning privacy protection method, system, medium, equipment and terminal Download PDF

Info

Publication number
CN115277015A
CN115277015A CN202210835092.XA CN202210835092A CN115277015A CN 115277015 A CN115277015 A CN 115277015A CN 202210835092 A CN202210835092 A CN 202210835092A CN 115277015 A CN115277015 A CN 115277015A
Authority
CN
China
Prior art keywords
user
server
key
users
secret
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202210835092.XA
Other languages
Chinese (zh)
Inventor
张应辉
曹大禹
刘伟
韩刚
郑东
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xian University of Posts and Telecommunications
Original Assignee
Xian University of Posts and Telecommunications
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Xian University of Posts and Telecommunications filed Critical Xian University of Posts and Telecommunications
Priority to CN202210835092.XA priority Critical patent/CN115277015A/en
Publication of CN115277015A publication Critical patent/CN115277015A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L9/00—Cryptographic mechanisms or cryptographic arrangements for secret or secure communications; Network security protocols
    • H04L9/32—Cryptographic mechanisms or cryptographic arrangements for secret or secure communications; Network security protocols including means for verifying the identity or authority of a user of the system or for message authentication, e.g. authorization, entity authentication, data integrity or data verification, non-repudiation, key authentication or verification of credentials
    • H04L9/3247—Cryptographic mechanisms or cryptographic arrangements for secret or secure communications; Network security protocols including means for verifying the identity or authority of a user of the system or for message authentication, e.g. authorization, entity authentication, data integrity or data verification, non-repudiation, key authentication or verification of credentials involving digital signatures
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L9/00—Cryptographic mechanisms or cryptographic arrangements for secret or secure communications; Network security protocols
    • H04L9/008—Cryptographic mechanisms or cryptographic arrangements for secret or secure communications; Network security protocols involving homomorphic encryption
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L9/00—Cryptographic mechanisms or cryptographic arrangements for secret or secure communications; Network security protocols
    • H04L9/06—Cryptographic mechanisms or cryptographic arrangements for secret or secure communications; Network security protocols the encryption apparatus using shift registers or memories for block-wise or stream coding, e.g. DES systems or RC4; Hash functions; Pseudorandom sequence generators
    • H04L9/0643—Hash functions, e.g. MD5, SHA, HMAC or f9 MAC
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L9/00—Cryptographic mechanisms or cryptographic arrangements for secret or secure communications; Network security protocols
    • H04L9/08—Key distribution or management, e.g. generation, sharing or updating, of cryptographic keys or passwords
    • H04L9/0816—Key establishment, i.e. cryptographic processes or cryptographic protocols whereby a shared secret becomes available to two or more parties, for subsequent use
    • H04L9/085—Secret sharing or secret splitting, e.g. threshold schemes
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L9/00—Cryptographic mechanisms or cryptographic arrangements for secret or secure communications; Network security protocols
    • H04L9/08—Key distribution or management, e.g. generation, sharing or updating, of cryptographic keys or passwords
    • H04L9/0861—Generation of secret information including derivation or calculation of cryptographic keys or passwords
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L9/00—Cryptographic mechanisms or cryptographic arrangements for secret or secure communications; Network security protocols
    • H04L9/08—Key distribution or management, e.g. generation, sharing or updating, of cryptographic keys or passwords
    • H04L9/0861—Generation of secret information including derivation or calculation of cryptographic keys or passwords
    • H04L9/0869—Generation of secret information including derivation or calculation of cryptographic keys or passwords involving random numbers or seeds

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Security & Cryptography (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Power Engineering (AREA)
  • Storage Device Security (AREA)

Abstract

The invention belongs to the technical field of machine learning safety, and discloses an asynchronous federal learning privacy protection method, system, medium, equipment and terminal, which initializes the system and selects random parameters, and discloses series parameters and signature public keys; a user generates a signature and a public and private key pair and sends the signature and the public and private key pair to a server, and the server packs and broadcasts user identity information and a public key to other users; a user generates a sub-secret corresponding to a random parameter and a one-time session key after receiving data; generating a shared key among users and encrypting user information; the information to be encrypted of the user is weighted local information of the user; the user sends the data after the mask is added to the server, and the server aggregates the data to obtain an aggregation result; and the server divides the aggregation result by the total number of the samples held by the users who finally participate in training to obtain the global model. The asynchronous federated learning privacy protection method protects the privacy of the user in federated learning; the time of learning and training is reduced, and resources are saved.

Description

Asynchronous federal learning privacy protection method, system, medium, equipment and terminal
Technical Field
The invention belongs to the technical field of machine learning safety, and particularly relates to an asynchronous federal learning privacy protection method, system, medium, equipment and terminal.
Background
Federal machine learning, also known as federal learning, is a machine learning framework, and can effectively help a plurality of organizations to perform data use and machine learning modeling on the premise of guaranteeing user privacy, data safety and government regulation requirements. Under the environment of explosive increase of data volume, federal learning is used as a distributed machine learning paradigm, the problem of data island can be effectively solved, and participators can jointly model on the basis of not sharing data, so that the data island is technically broken.
At present, data security in federal learning is mainly realized through a security aggregation protocol, and although the data security is ensured, the conventional federal learning has low concurrency and can cause the problem of low learning efficiency. Furthermore, since the local model update speed of users participating in federal learning is different, and in the traditional federal learning training, each round of aggregation must wait for all the participating users to complete, such model training is difficult and inefficient for syncing federals, which can cause the latter problem, and also can affect the learning efficiency. Therefore, asynchronous federal learning is proposed, however, the characteristics of asynchronous federal learning make it incompatible with security aggregation protocols, so that the problem that privacy and confidentiality of data cannot be guaranteed in asynchronous federal learning still exists at present.
Through the above analysis, the problems and defects of the prior art are as follows: synchronous federal learning has the problems of low concurrency and late concurrency, and the concurrency and training efficiency of federal learning need to be improved through asynchronous learning. The existing scheme tries to improve the privacy protection technology in asynchronous federal learning, but the privacy protection and training efficiency in asynchronous federal learning cannot be well balanced. How to realize privacy protection under the efficient training of asynchronous federal learning and not to influence the effectiveness of a training result is a problem in the federal learning at the present stage.
Disclosure of Invention
Aiming at the problems in the prior art, the invention provides an asynchronous federal learning privacy protection method, system, medium, equipment and terminal, and particularly relates to an asynchronous federal learning privacy protection method, system, medium, equipment and terminal based on safe multi-party computation.
In the training of asynchronous federal learning, the invention introduces a weight system for each user participating in the training to improve the training efficiency, and simultaneously provides a verification scheme based on a homomorphic hash function to ensure the safety of the weight system under the active attack adversary.
The invention is realized in such a way, and provides an asynchronous federal learning privacy protection method, which comprises the following steps: the invention introduces a weight system in an asynchronous federated learning framework to improve the learning efficiency, and simultaneously provides a verification scheme based on a homomorphic hash function to ensure that the security of the weight system initializes the system and selects random parameters, and discloses a series of parameters and a signature public key; the users participating in training receive the signature private key to generate a signature and a public and private key pair; each user packs the public key and the signature and sends the public key and the signature to the server, and the server packs and broadcasts the user identity information and the public key to other users after checking; after receiving the data, the user generates a sub-secret corresponding to the random parameter and a one-time session key and sends the sub-secret and the one-time session key to the server for broadcasting;
generating a shared key among users and encrypting user information; the information to be encrypted of each user is weighted local information of the user, and the weight is the number of samples held by the user and the staleness attenuation coefficient of the user during training; each user sends the data after the mask is added to a server, and the server aggregates the data to obtain an aggregation result; and the server divides the aggregation result by the total number of the samples held by the users who finally participate in training to obtain the global model.
Further, the asynchronous federated learning privacy protection method further comprises the following steps:
a trusted third party initializes the system, selects random parameters in a limited domain and discloses a series of parameters; users participating in training respectively generate two sets of public and private key pairs and signatures by using the parameters; the communication between the users and the server passes through a safety certification channel, each user packs the public key and the signature and sends the public key and the signature to the server, and the server packs and broadcasts the user identity information, the public key and the signature to other users after verification; after each user receives the data, the signature is verified, and random parameters are selected from a finite field; generating corresponding sub-secrets according to a secret sharing algorithm, then generating sub-secrets corresponding to a private key and a one-time session key corresponding to users, and encrypting and sending the sub-secrets and identity information to a server by the session key; the server broadcasts the ciphertext to the corresponding user;
generating a shared key between users, and encrypting the information of the users by masking the random number selected previously and the shared key through a pseudo-random number generator; the information which each user needs to encrypt is the local information of the user multiplied by the number of held samples and the decay coefficient of the staleness degree when the user participates in training; each user generates a verification vector and sends the data after being added with the mask to a server, and the server broadcasts a user list for participating in training; after receiving the list of all users, the user calculates the signature and sends the signature to the server;
the server broadcasts the signature to a corresponding user; each user receives the signature and then verifies the signature, a ciphertext sent by the server is decrypted to obtain a sub-secret, the sub-secret of the shared key of the disconnected user and the sub-secret of the random number of the disconnected user are sent to the server, the server recovers the original secret corresponding to the sub-secret after receiving the sub-secret, all mask data are added and subtracted by the secret passing through the pseudo-random number generator, and finally an aggregation result is obtained; and the server verifies that the sample number of the user is not false according to the aggregation result.
Further, the asynchronous federated learning privacy protection method comprises the following steps:
step one, a trusted third party initializes the system and is used for generating all system parameters in the scheme. A trusted third party selects random parameters in a limited domain and discloses a series of parameters and a signature public key of a user; each user generates a secret key, the users participating in training receive a signature private key from a trusted third party to generate a signature, public and private key pairs are generated by public parameters, and the public and private key pairs are packaged and sent to a server;
and step two, the user generates a cipher text which is used for generating keys required for constructing the mask among the users in the scheme, and the keys are sent to all the users in a secret mode. Each user selects random parameters from a finite field, generates a sub-secret corresponding to the random parameters and a sub-secret corresponding to a private key according to a secret sharing algorithm, encrypts the sub-secrets and identity information by using a one-time session key, and sends the encrypted sub-secrets and identity information to a server as a ciphertext;
and step three, generating local data with coefficients and masks by the users, and encrypting the local weighted data of each user in the scheme so as to ensure the safety of the private data of the users. Each user masks the random number selected previously and a shared key generated between the user and other users, and the mask is added to the local data with coefficients of the user for encryption; generating a verification vector, and sending the data added with the mask to a server;
and step four, the server decodes and aggregates the data, and the data is used for safely aggregating data at the server side in the scheme, namely the final output result of the scheme. Each user decrypts the ciphertext sent by the server to obtain the sub-secret, and the server recovers the original secret corresponding to the sub-secret and then recovers the mask code according to whether other users are disconnected to send the corresponding sub-secret, so as to obtain an aggregation result; the server generates a verification vector and verifies that the number of samples of the user is not false according to the aggregation result.
Further, the asynchronous federated learning privacy protection method further comprises the following steps:
(1) The trusted third party initializes and gives a security parameter k, resulting in (q, G, H), where q is a prime number, G is a group of order q, G is a generator of G, H: {0,1}*→{0，1}kIs a hash function, in which ZqRandomly choosing x as the signing key dSKCalculating dPK＝(gxmod q, g, q) as a signature verification key, where ZqRepresenting a finite field, simultaneously giving a threshold value t for secret recovery and the number n of users participating in training, and issuing a public parameter GPK = (G, q, G, H, t, n, d)PK(δ, ρ)), where (δ, ρ) is ∈ ZqThe key is a secret key in a homomorphic hash function; for 1. Ltoreq. A. Ltoreq.n identity information isaThe user of (2) receiving a signing key issued by a trusted third party
Figure BDA0003749590790000041
Signature verification key published with other users b
Figure BDA0003749590790000042
User generated key, user a in middle ZqRandomly selecting two different x1And x2Two pairs of public and private keys are generated, and one pair of public and private keys is
Figure BDA0003749590790000043
Wherein
Figure BDA0003749590790000044
Similarly, another pair of public and private keys is
Figure BDA0003749590790000045
Wherein
Figure BDA0003749590790000046
The former of the two pairs of keys is used for authentication encryption, and the latter is used for generating a mask; user a obtains signing key through trusted third party
Figure BDA0003749590790000047
Signature generation for messages
Figure BDA0003749590790000048
Figure BDA0003749590790000049
Wherein k ∈ Zq is the random number selected by user a(ii) a Will be provided with
Figure BDA00037495907900000410
Packaging and sending to a server; the server only judges whether the number of users sending data is larger than a threshold value t or not, and records the current user as U1If and only if | U1Executing the next step when | ≧ t is established; server packaging
Figure BDA00037495907900000411
Broadcasting to other users b;
(2) User generates cipher text, user a receives message of other user b broadcast by server, uses verification key of b
Figure BDA00037495907900000412
Authentication
Figure BDA00037495907900000413
If it is true, and selecting a random number beta in a finite fieldaGenerating a secret beta for the other user b based on a threshold value taAnd
Figure BDA00037495907900000414
secret share beta ofa，bAnd
Figure BDA00037495907900000415
wherein b ∈ U1(ii) a User a uses its own private key
Figure BDA00037495907900000416
Public key published by other users b
Figure BDA00037495907900000417
Figure BDA00037495907900000418
Generating one-time session keys
Figure BDA00037495907900000419
And keya，b＝keyb，a(ii) a For user aDisposable session keya，bEncrypting the identity information and the two secret shares to produce a ciphertext message to the other user b
Figure BDA0003749590790000051
The server only judges whether the number of users sending data is larger than a threshold value t or not, and records the current user as U2If and only if | U2Executing the next step when | ≧ t is established; the server broadcasts the ciphertext to the corresponding user; the user generates local data with coefficient and mask, user a receives corresponding cipher text from the server and stores it locally, calculates shared secret key s with other users ba，bFor generating a mask; the random number beta selected previouslyaAnd shared secret s with other users ba，bObtaining a private mask and a public mask by a pseudo-random number generator; user a sends local data xaMultiplied by the number n of samples held by the useraAnd server specified staleness attenuation coefficient
Figure BDA0003749590790000052
Adding the mask to obtain the output yaAnd generating a verification vector V according to the parameters of the trusted third partyaWill { ya，VaSending the data to a server; the server only judges whether the number of users sending data is larger than a threshold value t or not, and records the current user as U3(ii) a If and only if | U3Executing the next step when | ≧ t is true, and converting U3Broadcasting the user list to all users;
(3) The server decodes and aggregates, the user a receives the user list participating in training, and the list generates a signature by using the signature key
Figure BDA0003749590790000053
And sending to a server; the server receives the signature and records the current user set as U4Packing the identity information and the signature and sending the identity information and the signature to all users; user a receives the message and then verifies signatures sigma of other users'b，b∈U4Judging that the user list is not tampered; using one-time session keya，bDecryption clothesCiphertext c previously sent by the Serverb，aObtaining identity information and two groups of secret shares, verifying the identity and confirming that the secret shares are not false, and packaging and sending the secret shares to a server; the server records the current user set as U5Recovering the private mask and the public mask of the disconnected user through the secret share sent by the user; the server aggregates the output of the users and then subtracts the private mask and the public mask of the disconnected users to obtain a weighted aggregation result z, and generates a verification vector K according to the aggregation resultaZ verifies that the user is not false.
Further, in the step (2), the user a finally generates an output yaAnd a verification vector VaThe method comprises the following steps:
1) The user a will select the random number beta previouslyaAnd shared secret s with other users ba，bObtaining a private mask PRG (beta) by a pseudo-random number generator PRGa) And a common mask ∑ PRG(s)a，b)；
2) User a sends local data xaMultiplied by the number n of samples held by the useraAnd a server specified staleness attenuation factor
Figure BDA0003749590790000054
Wherein t-tau represents the time difference of the user a participating in the polymerization, and alpha belongs to (0, 1);
3) User a output
Figure BDA0003749590790000061
Figure BDA0003749590790000062
Wherein
Figure BDA0003749590790000063
4) The user a generates a verification vector according to the public parameters (delta, rho) of the trusted third party
Figure BDA0003749590790000064
Figure BDA0003749590790000065
Wherein HF isδ，ρIs a homomorphic hash function, eta is the learning rate of federal learning, ∑ waFor user a local data in the current global model wGSum of the lower gradients.
Further, the server aggregation and authentication in the step (3) includes:
1) After the user confirms that the user list is not fake, the disposable session key is useda，bDecrypting ciphertext c previously sent by serverb，aTo obtain
Figure BDA0003749590790000066
Verifying whether a = a '^ b = b' is established, and if so, representing that the ciphertext is sent by other users b;
2) Will be provided with
Figure BDA0003749590790000067
And { beta ]b，a|b∈U3And (5) packaging and sending to a server, wherein U2\U3Representing a disconnected user set, U3Representing a user set which is not disconnected;
3) The server has previously confirmed | U4If | ≧ t, the remaining users can recover the random number beta of the user without disconnectionaAnd further, the private mask PRG (beta) is restoreda) Where a ∈ U3(ii) a Meanwhile, the remaining users can recover the private key of the disconnected user
Figure BDA0003749590790000068
And thus with the public key of the corresponding user b
Figure BDA0003749590790000069
Recovery of shared secret s by key agreementa，bAnd further obtains a public mask Σ PRG(s)a，b) Where a ∈ U2\U3，b∈U3；
4) The server obtains the final output
Figure BDA00037495907900000610
Figure BDA00037495907900000611
5) The server generates a verification vector from public parameters (delta, rho) of the trusted third party
Figure BDA00037495907900000612
Figure BDA00037495907900000613
And
Figure BDA00037495907900000614
wherein a ∈ U3；
6) Server authentication
Figure BDA00037495907900000615
And if yes, outputting z.
For the offline user, the server cannot recover the random number selected by the user and cannot acquire local data; for the user who is not disconnected, the server cancels the public mask through summation, and cannot acquire local data.
Another object of the present invention is to provide an asynchronous federal learning privacy protection system using the asynchronous federal learning privacy protection method, where the asynchronous federal learning privacy protection system includes:
the key generation module is used for initializing the system by a trusted third party, selecting random parameters in a limited domain and disclosing a series of parameters and a signature public key of a user; the user generates a key, the user participating in training receives a signature private key from a trusted third party to generate a signature, generates a public and private key pair by using public parameters, packs and sends the public and private key pair to the server;
the cipher text generation module is used for generating cipher texts through users, each user selects random parameters from a finite field, sub-secrets corresponding to the random parameters and sub-secrets corresponding to the private keys are generated according to a secret sharing algorithm, the sub-secrets and identity information are encrypted by using a one-time session key and are sent to a server as the cipher text;
the local data encryption module is used for generating local data with coefficients and masks through users, each user masks the random number selected previously and a shared key generated between the users, and the local data with the coefficients of the users are encrypted by adding codes; generating a verification vector, and sending the data added with the mask to a server;
the aggregation verification module is used for decoding and aggregating through the server, decrypting the ciphertext sent by the server by the user to obtain the sub-secret, sending the corresponding sub-secret according to whether other users are disconnected, recovering the original secret corresponding to the sub-secret by the server, and then recovering the mask code to further obtain an aggregation result; the server generates a verification vector and verifies that the number of samples of the user is not false according to the aggregation result.
It is a further object of the present invention to provide a computer device comprising a memory and a processor, the memory storing a computer program which, when executed by the processor, causes the processor to perform the steps of the asynchronous federal learned privacy protected method.
It is another object of the present invention to provide a computer readable storage medium storing a computer program which, when executed by a processor, causes the processor to perform the steps of the asynchronous federated learning privacy protection method.
Another object of the present invention is to provide an information data processing terminal, which is configured to implement the asynchronous federal learning privacy protection system.
In combination with the above technical solutions and the technical problems to be solved, please analyze the advantages and positive effects of the technical solutions to be protected in the present invention from the following aspects:
first, aiming at the technical problems existing in the prior art and the difficulty in solving the problems, the technical problems to be solved by the technical scheme of the present invention are closely combined with results, data and the like in the research and development process, and some creative technical effects are brought after the problems are solved. The specific description is as follows:
the invention provides an asynchronous federal learning privacy protection method based on safe multi-party computation.A trusted third party initializes a system, selects random parameters in a finite field and discloses a series of parameters and a signature public key; users participating in training receive the private signature key to generate a signature, and use the parameters to respectively generate two groups of public and private key pairs and open the number of held samples; the communication between the user and the server passes through a safety certification channel, the user packs the public key and the signature and sends the public key and the signature to the server, and the server packs the user identity information and the public key and broadcasts the user identity information and the public key to other users after verification; after receiving data, a user selects random parameters from a finite field, generates corresponding sub-secrets according to a secret sharing algorithm, then generates sub-secrets corresponding to a private key and a one-time session key corresponding to the user, and then encrypts and sends the sub-secrets and identity information to a server through the session key; the server broadcasts the ciphertext to the corresponding user; generating a shared key among users, and then masking the random number selected previously and the shared key by a pseudo-random number generator to encrypt the information of the users; the information which needs to be encrypted by the user is weighted local information of the user, wherein the weight is the number of samples held by the user and the staleness attenuation coefficient of the user during training; the user sends the data after the mask is added to the server, and the server broadcasts a user list participating in training at the moment; after receiving the list of all users, the user decrypts the ciphertext sent by the server to obtain the sub-secrets, then sends the sub-secrets of the shared secret key of the off-line user and the sub-secrets of the random number of the off-line user to the server, and after receiving the sub-secrets, the server recovers the original secrets corresponding to the sub-secrets, and then adds all the mask data and subtracts the secrets passing through the pseudo-random number generator to finally obtain an aggregation result; and the server divides the aggregation result by the total number of the samples held by the final training users to obtain a global model.
The invention provides an asynchronous federal learning privacy protection method based on safe multi-party computation, which is used for realizing efficient and safe asynchronous federal learning. The user and the server perform operations asynchronously, so that the model for training is more efficiently transmitted in the system, and the training time is reduced. Meanwhile, the introduction of the weight system greatly improves the training accuracy and efficiency of the federal learning. The server and the users execute corresponding operations through the asynchronous coordinator, when a user is slow in local updating due to insufficient computing power, other users do not need to wait for the user to complete the local updating, and other users who complete updating directly take over to complete the safety aggregation, so that the flexibility and the concurrency of federal learning are greatly improved. The security aggregation is realized by a secure multi-party computing technology, the technology uses a secret shared key between users as a mask, the mask is attached to the weighted local privacy data to mask plaintext information, and a server cannot identify the masked information. For a dropped user, the server can recover its public mask but cannot obtain its private mask and cannot obtain its private information. Similarly, for a user who is not disconnected, the server can obtain the private mask of the user, but cannot obtain the public mask, and the public mask can be offset only through aggregation operation, namely, the server can only obtain the result after aggregation, so that an implementable asynchronous federal learning privacy protection scheme with high safety performance is obtained, and the privacy and confidentiality of user data are protected; on the other hand, the invention also improves the federal learning training efficiency, solves the problem of falling of the latter in the traditional federal learning, reduces the training time and saves the resources. The scheme is simple, the practicability is high, and the popularization effect is achieved.
Secondly, considering the technical scheme as a whole or from the perspective of products, the technical effect and advantages of the technical scheme to be protected by the invention are specifically described as follows:
the asynchronous federal learning privacy protection method based on safe multi-party calculation provided by the invention protects the privacy of users in federal learning; the time of learning and training is reduced, and resources are saved.
Third, as an inventive supplementary proof of the claims of the present invention, there are also presented several important aspects:
(1) The technical scheme of the invention fills the technical blank in the industry at home and abroad:
the safe asynchronous federated learning with the weighting system is provided for the first time, is suitable for asynchronous federated learning under a multi-user scene, is particularly suitable for asynchronous federated learning under a multi-edge computing device user scene, can effectively perform federated learning training, and ensures privacy of local data of users in the training. Meanwhile, a verification scheme of the weighting system under the active attack adversary is provided by combining a homomorphic hash function, and the safety of the weighting system is ensured.
(2) The technical scheme of the invention solves the technical problems which are always desired to be solved but are not successfully achieved:
at present, the communication cost is seriously considered to be reduced aiming at federal learning in China, and the thinking for improving the concurrency and the learning efficiency is lacked. In an actual application scenario, user equipment participating in federal learning has differences, and it cannot be guaranteed that all users can complete local model updating at a relatively average speed, and the latter problem is finally caused. Although asynchronous federated learning is proposed to improve the training efficiency, an efficient asynchronous federated learning privacy protection method is not proposed in the existing scheme, and the method can realize efficient safe asynchronous federated learning and is more in line with the actual application scene.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the embodiments of the present invention will be briefly described below, and it is obvious that the drawings described below are only some embodiments of the present invention, and it is obvious for those skilled in the art that other drawings can be obtained according to the drawings without creative efforts.
FIG. 1 is a flow chart of an asynchronous federated learning privacy protection method provided by an embodiment of the present invention;
FIG. 2 is a communication diagram of an asynchronous security aggregation scheme provided by an embodiment of the present invention;
FIG. 3 is a schematic diagram of a secure asynchronous federated learning framework provided by an embodiment of the present invention;
FIG. 4 is a comparison diagram of the presence or absence of weights in asynchronous federated learning provided by an embodiment of the present invention;
fig. 5 is a schematic diagram of training times of a random gradient descent method (sgd), a Synchronous federal Learning (syn), and an Asynchronous federal Learning (asyn) under different numbers of training rounds, which are used in conventional federal Learning according to an embodiment of the present invention;
fig. 6 is a schematic diagram after the privacy protection scheme (asyn + secagg, asynchronous fed Learning + Secure Aggregation) of the present invention is added.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail with reference to the following embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
In order to solve the problems in the prior art, the invention provides an asynchronous federal learning privacy protection method, system, medium, device and terminal, and the invention is described in detail with reference to the accompanying drawings.
1. The embodiments are explained. This section is an explanatory embodiment expanding on the claims so as to fully understand how the present invention is embodied by those skilled in the art.
As shown in fig. 1, the asynchronous federated learning privacy protection method provided in the embodiment of the present invention includes the following steps:
s101, a trusted third party initializes a system, selects random parameters in a limited domain, and discloses a series of parameters and a signature public key of a user; the user generates a key, the user participating in training receives a private signature key from a trusted third party to generate a signature, a public and private key pair is generated by using public parameters, and the public and private key pair is packaged and sent to a server;
s102, a user generates a ciphertext, selects random parameters from a finite field, generates a sub-secret corresponding to the random parameters and a sub-secret corresponding to a private key according to a secret sharing algorithm, encrypts the sub-secret and identity information by using a one-time session key, and sends the encrypted sub-secret and identity information to a server as the ciphertext;
s103, a user generates local data with a coefficient and a mask, the user masks the random number selected previously and a shared key generated between the users, and the local data with the coefficient of the user is encrypted by adding codes; generating a verification vector, and sending the data added with the mask to a server;
s104, the server decodes and aggregates, the user decrypts the ciphertext sent by the server to obtain the sub-secret, and the server recovers the original secret corresponding to the sub-secret and then recovers the mask code according to whether other users drop the line to send the corresponding sub-secret, so as to obtain an aggregation result; the server generates a verification vector, and verifies that the number of samples of the user is not false according to the aggregation result.
As a preferred embodiment, as shown in fig. 2 to 3, the asynchronous federated learning privacy protection method provided in the embodiment of the present invention specifically includes the following steps:
(1) Initializing a system:
the trusted third party initializes and gives a security parameter k, resulting in (q, G, H), where q is a prime number, G is a group of order q, G is a generator of G, H: {0,1}*→{0，1}kIs a hash function; in ZqRandomly choosing x as the signing key dSKCalculating dPK＝(gxmod q, g, q) as signature verification key, where ZqRepresenting a finite field, simultaneously giving a threshold value t for secret recovery and the number n of users participating in training, and issuing a public parameter GPK = (G, q, G, H, t, n, d)PK(δ, ρ)), where (δ, ρ) is ∈ ZqThe key is a key in a homomorphic hash function; for the user with the identity information of a being more than or equal to 1 and less than or equal to n and a, receiving a signature key issued by a trusted third party
Figure BDA0003749590790000121
Signature verification key disclosed by other user b
Figure BDA0003749590790000122
(2) The user generates a key:
user a is in middle ZqRandomly selecting two different x1And x2Two pairs of public and private keys are generated, and one pair of public and private keys is
Figure BDA0003749590790000123
Wherein
Figure BDA0003749590790000124
Similarly, another pair of public and private keys is
Figure BDA0003749590790000125
Wherein
Figure BDA0003749590790000126
The former is used for authentication encryption, and the latter is used for generating a mask; user a obtains signing key through trusted third party
Figure BDA0003749590790000127
Signature generation for messages
Figure BDA0003749590790000128
Figure BDA0003749590790000129
Wherein k ∈ ZqRandom number selected for user a, which will then be
Figure BDA00037495907900001210
Packaging and sending to a server; the server only judges whether the number of users sending data is larger than a threshold value t or not, and records the current user as U1If and only if | U1Executing the next step when | ≧ t is established; server packaging
Figure BDA00037495907900001211
Broadcast to other users;
(3) The user generates a ciphertext:
user a receives the message broadcast by the server and uses the authentication key of other users b
Figure BDA00037495907900001212
Authentication
Figure BDA00037495907900001213
If this is true, then a random number β is selected in the finite fieldaGenerating secrets for other users b, based on a threshold value t
Figure BDA00037495907900001214
Secret share beta ofa，bAnd
Figure BDA00037495907900001215
wherein b ∈ U1(ii) a User a uses its own private key
Figure BDA00037495907900001216
Public key published by other users b
Figure BDA00037495907900001217
Generating one-time session keys
Figure BDA00037495907900001218
And keya，b＝keyb，a(ii) a User a uses disposable session keya，bEncrypting the identity information and the two secret shares to produce a ciphertext message to the other user b
Figure BDA00037495907900001219
Figure BDA00037495907900001220
Namely symmetric encryption; the server only judges whether the number of users sending data is larger than a threshold value t or not, and records the current user as U2If and only if | U2Executing the next step when | ≧ t is established; the server broadcasts the ciphertext to the corresponding user;
(4) The user generates local data with coefficients plus masks:
user a receives the corresponding cipher text from the server, stores the cipher text in the local, and then calculates the shared secret key s of other users ba，bTo generate a mask; the random number beta selected previouslyaAnd shared secret s with other users ba，bObtaining a private mask and a public mask by a pseudo-random number generator; user a sends local data xaMultiplied by the number n of samples held by the useraAnd server specified staleness attenuation coefficient
Figure BDA00037495907900001221
Adding the mask to obtain the output yaThen generates a verification vector VaWill { ya，VaSending the data to a server; the server only judges whether the number of users sending data is larger than a threshold value t or not, and records the current user as U3If and only if | U3If | ≧ t is true, then executing the next step, and then taking U3Broadcasting the user list to all users;
(5) And (3) checking consistency:
user a receives the list of users participating in training, and uses the list to generate signature by using signature key
Figure BDA0003749590790000131
Figure BDA0003749590790000132
And sending to the server; the server receives the signature and records the current user set as U4Packing the identity information and the signature and sending the identity information and the signature to all users;
(6) The server decodes the aggregation:
user a verifies signature sigma 'after receiving message'b，b∈U4The verification mode is the same as the step (2), the user list is judged not to be tampered, and then the disposable session key is useda，bDecrypting ciphertext c previously sent by serverb，aObtaining identity information and two groups of secret shares, and confirming that no fake is made after the identity is verified; packaging and sending the secret share to a server; the server records the current user set as U5Recovering the private mask and the public mask of the dropped user by the secret share sent by the user; the server aggregates the user's output and subtracts the private mask sumObtaining weighted aggregation result z by public mask of off-line user, and then generating verification vector K according to the aggregation resultaZ, to verify that the user is not false.
Wherein, the output y finally generated by the user a in the step (4) provided by the embodiment of the inventionaAnd a verification vector generated as follows:
(a) The user a will select the random number beta previouslyaAnd shared secret s with other users ba，bObtaining a private mask PRG (beta) by a pseudo-random number generator PRGa) And a common mask ∑ PRG(s)a，b)；
(b) User a sends local data xaMultiplied by the number n of samples held by the useraAnd server specified staleness attenuation coefficient
Figure BDA0003749590790000133
Wherein t-tau represents the time difference of the user a participating in the polymerization, and alpha belongs to (0, 1);
(c) User a output
Figure BDA0003749590790000134
Figure BDA0003749590790000135
Wherein
Figure BDA0003749590790000136
(d) User a generates an authentication vector
Figure BDA0003749590790000137
Wherein HF isδ，ρFor homomorphic hash functions, η is the learning rate of federal learning, ∑ waFor user a local data in the current global model wGSum of the lower gradients.
The polymerization and verification in step (6) provided by the embodiment of the invention are performed as follows:
(a) After the user confirms that the user list is not fake, the disposable session key is useda，bDecrypting ciphertext c previously sent by serverb，aTo obtain
Figure BDA0003749590790000141
At this time, whether a = a '^ b = b' is verified, if yes, the ciphertext is sent by other users b is represented;
(b) Will be provided with
Figure BDA0003749590790000142
And { beta ]b，a|b∈U3And (5) packaging and sending to a server, wherein U2\U3Representing a disconnected user set, U3Representing a user set which is not disconnected;
(c) The server has previously acknowledged | U4If | ≧ t, the remaining users can recover the random number β of the user without disconnectionaAnd further, the private mask PRG (beta) is restoreda) Where a ∈ U3(ii) a Meanwhile, the remaining users can recover the private key of the disconnected user
Figure BDA0003749590790000143
And further with the public key corresponding to the other user b
Figure BDA0003749590790000144
Recovery of shared secret s by key agreementa，bFurther, a common mask Σ PRG(s) can be obtaineda，b) Where a ∈ U2\U3，b∈U3；
(d) The server obtains the final output
Figure BDA0003749590790000145
Figure BDA0003749590790000146
(e) Server-generated authentication vectors
Figure BDA0003749590790000147
And
Figure BDA0003749590790000148
wherein a ∈ U3；
(f) Server authentication
Figure BDA0003749590790000149
And if yes, outputting z.
For the user who is disconnected, the server cannot recover the random number selected by the user, so that the local data of the user cannot be acquired; for users who are not dropped, the server can only cancel the public mask by summing, so that local data cannot be acquired.
The asynchronous federated learning privacy protection system provided by the embodiment of the invention comprises:
the key generation module is used for initializing the system by a trusted third party, selecting random parameters in a finite field and disclosing a series of parameters and a signature public key of a user; the user generates a key, the user participating in training receives a signature private key from a trusted third party to generate a signature, generates a public and private key pair by using public parameters, packs and sends the public and private key pair to the server;
the cipher text generation module is used for generating a cipher text through a user, selecting random parameters from a limited domain by the user, generating a sub-secret corresponding to the random parameters and a sub-secret corresponding to a private key according to a secret sharing algorithm, encrypting the sub-secrets and identity information by using a one-time session key, and sending the encrypted sub-secrets and identity information to a server as the cipher text;
the local data encryption module is used for generating local data with coefficients and masks through a user, the user masks the random number selected in the past and a shared key generated between the users, and the local data with the coefficients of the user is encrypted in an encryption mode; generating a verification vector, and sending the data added with the mask to a server;
the aggregation verification module is used for decoding and aggregating through the server, decrypting a ciphertext sent by the server by a user to obtain a sub-secret, sending a corresponding sub-secret according to whether other users are disconnected, recovering an original secret corresponding to the sub-secret by the server, and then recovering a mask code to further obtain an aggregation result; the server generates a verification vector and verifies that the number of samples of the user is not false according to the aggregation result.
2. Application examples. In order to prove the creativity and the technical value of the technical scheme of the invention, the part is the application example of the technical scheme of the claims on specific products or related technologies.
According to an actual application scene, the safe aggregation of the asynchronous federated learning data with the weight is realized by utilizing a safe multi-party computing technology, and meanwhile, a weight verification scheme is provided according to a homomorphic hash function, so that the correctness of the weight in the asynchronous federated learning is ensured.
In the intelligent health detection system, a user can select to upload data collected on own wearable equipment or a mobile phone to a server for federal learning, the server can obtain a better group user health model after the federal learning, and in the process, the user equipment participating in the federal learning has heterogeneous conditions. For example, the smart watch and the mobile phone have great difference in computational power or standby time, which may cause the problem that the local model update of the smart watch and the local model update of the mobile phone are not synchronous when the server collects data for federal learning, thereby causing the problem of falling of the latter. After the method is combined, the users who participate in training can be ensured not to wait for the user with slow local model updating during asynchronous federal learning, so that the training efficiency and concurrency of federal learning are improved.
In addition, in this scenario, since private information may exist in data collected locally by the user, such as blood pressure, heart rate, and the like of the user, the user does not want to disclose the data to other users or servers, which puts a requirement on privacy protection on asynchronous federal learning. When the method is combined, the privacy of the user data in the training process of federal learning can be ensured. The server can only get the aggregated result and cannot get the local information of any user.
In addition, in the practical application scenario, the user uploads the sample number in a public way, so that the actively attacking user can upload the sample number maliciously to interfere with normal training of asynchronous federal learning. The invention provides a verification scheme for the number of the user samples by utilizing the unforgeability and the confidentiality of the homomorphic hash function, and ensures that a user does not maliciously upload false weight values in the training process. The scheme is that the server carries out verification after obtaining the aggregated result, if a user who actively attacks uploads a false weight value, a model generated by the training is unavailable, and the server can directly abandon the result for the next training after verification.
In conclusion, the invention introduces the weight system in asynchronous federated learning for the first time, and provides an efficient safe asynchronous federated learning scheme by combining a safe multi-party computing technology and a homomorphic hash function, so that certain social requirements can be met, and the method has practical value.
3. Evidence of the relevant effects of the examples. The embodiment of the invention achieves some positive effects in the process of research and development or use, and has great advantages compared with the prior art, and the following contents are described by combining data, diagrams and the like in the test process.
As shown in fig. 3, a secure asynchronous federated learning framework is shown, and asynchronous federated learning security aggregation adopted in the present solution is consistent with the prior art, that is, operations of the user side and the server side are determined by the asynchronous coordinator, and security aggregation is performed after sufficient local updates of the user are collected, and at this time, other users who do not upload data can still perform local updates, thereby achieving an asynchronous effect.
As shown in fig. 4, the comparison of the presence or absence of the weight in the Asynchronous federal Learning is shown, and it is obvious that the Learning effect after the weight (asyn + w, asynchronous fed Learning + Weighted), that is, the test accuracy, is more excellent; fig. 5 shows training times of a random gradient descent method (sgd) used in conventional federal Learning, synchronous federal Learning (syn) and asynchronous federal Learning (asyn) under different training rounds, which clearly shows that the training time of the asynchronous federal Learning under a high training round is lower than that of the other two conventional schemes, which also proves the importance and necessity of the asynchronous federal Learning.
As shown in fig. 6, it is demonstrated that after the privacy protection scheme (asyn + secagg, asynchronous Federal Learning + Secure Aggregation) of the present invention is added, the effect of Asynchronous Federal Learning is not much affected compared to when the scheme is not added. The method improves the safety of asynchronous federal learning on the premise of not influencing the training effect of asynchronous federal learning.
It should be noted that embodiments of the present invention can be realized in hardware, software, or a combination of software and hardware. The hardware portion may be implemented using dedicated logic; the software portions may be stored in a memory and executed by a suitable instruction execution system, such as a microprocessor or specially designed hardware. Those skilled in the art will appreciate that the apparatus and methods described above may be implemented using computer executable instructions and/or embodied in processor control code, such code being provided on a carrier medium such as a disk, CD-or DVD-ROM, programmable memory such as read only memory (firmware), or a data carrier such as an optical or electronic signal carrier, for example. The apparatus and its modules of the present invention may be implemented by hardware circuits such as very large scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of hardware circuits and software, e.g., firmware.
The above description is only for the purpose of illustrating the present invention and the appended claims are not to be construed as limiting the scope of the invention, which is intended to cover all modifications, equivalents and improvements that are within the spirit and scope of the invention as defined by the appended claims.

Claims (10)

1. An asynchronous federated learning privacy protection method is characterized by comprising the following steps:
initializing a system and selecting random parameters, and disclosing a series of parameters and a signature public key; the users participating in training receive the signature private key to generate a signature and a public and private key pair; the user packs the public key and the signature and sends the public key and the signature to the server, and the server packs the user identity information and the public key and broadcasts the user identity information and the public key to other users after verification; after receiving the data, the user generates a sub-secret corresponding to the random parameter and a one-time session key and sends the sub-secret and the one-time session key to the server for broadcasting;
generating a shared key between users and encrypting user information; the information to be encrypted of the user is weighted local information of the user, and the weight is the number of samples held by the user and the staleness attenuation coefficient of the user during training; the user sends the data after the mask is added to the server, and the server aggregates the data to obtain an aggregation result; and the server divides the aggregation result by the total number of the samples held by the users who finally participate in training to obtain the global model.
2. The asynchronous federated learning privacy protection method of claim 1, wherein the asynchronous federated learning privacy protection method further comprises:
a trusted third party initializing system, selecting random parameters in a finite field and disclosing a series of parameters; users participating in training respectively generate two sets of public and private key pairs and signatures by using the parameters; the communication between the user and the server passes through a safety authentication channel, the user packs the public key and the signature and sends the public key and the signature to the server, and the server packages and broadcasts the user identity information, the public key and the signature to other users after verification; the user verifies the signature after receiving the data and selects random parameters from the finite field; generating corresponding sub-secrets according to a secret sharing algorithm, then generating sub-secrets corresponding to a private key and a one-time session key corresponding to users, and encrypting and sending the sub-secrets and identity information to a server through the session key; the server broadcasts the ciphertext to the corresponding user;
generating a shared key among users, and encrypting the information of the users by masking the selected random number and the shared key through a pseudo-random number generator; the information which needs to be encrypted by the user is the local information of the user multiplied by the number of held samples and the staleness attenuation coefficient when the user participates in training; the user generates a verification vector and sends the data after being added with the mask to a server, and the server broadcasts a user list for participating in training; after receiving the list of all users, the user calculates the signature and sends the signature to the server;
the server broadcasts the signature to a corresponding user; the user receives the signature and then verifies the signature, a ciphertext sent by the server is decrypted to obtain a sub-secret, the sub-secret of the shared key of the disconnected user and the sub-secret of the random number of the disconnected user are sent to the server, the server recovers the original secret corresponding to the sub-secret after receiving the sub-secret, all mask data are added and subtracted by the secret passing through the pseudo-random number generator, and finally an aggregation result is obtained; and the server verifies that the sample number of the user is not false according to the aggregation result.
3. The asynchronous federated learning privacy protection method of claim 1, wherein the asynchronous federated learning privacy protection method comprises the steps of:
step one, a trusted third party initializes a system, selects random parameters in a limited domain, and discloses a series of parameters and a signature public key of a user; the user generates a key, the user participating in training receives a private signature key from a trusted third party to generate a signature, a public and private key pair is generated by using public parameters, and the public and private key pair is packaged and sent to a server;
step two, a user generates a ciphertext, selects random parameters from a finite field, generates a sub-secret corresponding to the random parameters and a sub-secret corresponding to a private key according to a secret sharing algorithm, encrypts the sub-secret and identity information by using a one-time session key, and sends the encrypted sub-secret and identity information to a server as the ciphertext;
step three, the user generates local data with the coefficient and the mask, the user masks the random number selected previously and the shared key generated between the users, and the local data with the coefficient of the user is encrypted by adding codes; generating a verification vector, and sending the data added with the mask to a server;
step four, the server decodes and aggregates, the user decrypts the ciphertext sent by the server to obtain the sub-secret, and the server recovers the original secret corresponding to the sub-secret and then recovers the mask code according to whether other users drop the line to send the corresponding sub-secret, so as to obtain an aggregation result; the server generates a verification vector and verifies that the number of samples of the user is not false according to the aggregation result.
4. The asynchronous federated learning privacy protection method of claim 1, wherein the asynchronous federated learning privacy protection method further comprises:
(1) The trusted third party initializes and gives a security parameter k, resulting in (q, G, H), where q is a prime number, G is a group of order q, G is a generator of G, H: {0,1}*→{0，1}kIs a hash function; in the middle ZqRandomly choosing x as the signing key dSKCalculating dPK＝(gxmod q, g, q) as a signature verification key, where ZqRepresenting a finite field, simultaneously giving a threshold value t for secret recovery and the number n of users participating in training, and issuing a public parameter GPK = (G, q, G, H, t, n, d)PK(δ, ρ)), where (δ, ρ) is ∈ ZqThe key is a key in a homomorphic hash function; for the user with the identity information of a being more than or equal to 1 and less than or equal to n and a, the user receives a signature key issued by a trusted third party
Figure FDA0003749590780000031
Signature verification key published with other users b
Figure FDA0003749590780000032
User generates a key, user a is in middle ZqRandomly selecting two different x1And x2Two pairs of public and private keys are generated, and one pair of public and private keys is
Figure FDA0003749590780000033
Wherein
Figure FDA0003749590780000034
Similarly, another pair of public and private keys is
Figure FDA0003749590780000035
Wherein
Figure FDA0003749590780000036
The former of the two pairs of keys is used for authentication encryption, and the latter is used for generating a mask; user a obtains a signing key through a trusted third party
Figure FDA0003749590780000037
Signature generation for messages
Figure FDA0003749590780000038
Figure FDA0003749590780000039
Wherein k ∈ ZqA random number selected for user a; will be provided with
Figure FDA00037495907800000310
Packaging and sending to a server; the server only judges whether the number of users sending data is larger than a threshold value t or not, and records the current user as U1And if and only if | U1Executing the next step when | ≧ t is established; server packaging
Figure FDA00037495907800000311
Broadcasting to other users b;
(2) User generates cipher text, user a receives the information of other user b broadcast by server, uses the verification key of b
Figure FDA00037495907800000312
Authentication
Figure FDA00037495907800000313
Whether it is true, and selecting a random number beta in a finite fieldaGenerating a secret beta for the other user b based on a threshold value taAnd
Figure FDA00037495907800000314
secret share of (b)a，bAnd
Figure FDA00037495907800000315
wherein b ∈ U1(ii) a User a uses its own private key
Figure FDA00037495907800000316
Public key published by other users b
Figure FDA00037495907800000317
Figure FDA00037495907800000318
Generating one-time session keys
Figure FDA00037495907800000319
And keya，b＝keyb，a(ii) a User a uses disposable session keya，bEncrypting the identity information and the two secret shares to produce a ciphertext message to the other user b
Figure FDA00037495907800000320
The server only judges whether the number of users sending data is larger than a threshold value t or not, and records the current user as U2If and only if | U2Executing the next step when | ≧ t is established; the server broadcasts the ciphertext to the corresponding user; the user generates local data with coefficient and mask, user a receives corresponding cipher text from the server and stores it locally, calculates shared secret key s with other users ba，bFor generating a mask; the random number beta selected previouslyaAnd shared secret s with other users ba，bObtaining a private mask and a public mask by a pseudo-random number generator; user a sends local data xaMultiplied by the number n of samples held by the useraAnd server specified staleness attenuation coefficient
Figure FDA00037495907800000321
Adding the mask to obtain the output ya(ii) a Generating a verification vector V from parameters of a trusted third partyaWill { ya，VaSending the result to a server; the server only judges whether the number of users sending data is larger than a threshold value t or not, and records the current user as U3(ii) a If and only if | U3Executing the next step when | ≧ t is true, and converting U3Broadcasting the user list to all users;
(3) The server decodes and aggregates, the user a receives the user list participating in training, and the list generates a signature by using the signature key
Figure FDA0003749590780000041
And sending to a server; the server receives the signature and records the current user set as U4Packing the identity information and the signature and sending the identity information and the signature to all users; user a verifies signature sigma 'after receiving message'b，b∈U4Judging that the user list is not tampered; using one-time session keya，bDecrypting ciphertext c previously sent by serverb，aObtaining identity information and two groups of secret shares, and confirming that no fake is made after the identity is verified; packaging and sending the secret share to a server; the server records the current user set as U5Recovering the private mask and the public mask of the disconnected user through the secret share sent by the user; the server aggregates the output of the users and then subtracts the private mask and the public mask of the disconnected user to obtain a weighted aggregate result z, and a verification vector K is generated according to the aggregate resultaZ verifies that the user is not false.
5. The asynchronous federated learning privacy protection method of claim 4, wherein user a ultimately generates output y in step (2)aAnd a verification vector VaThe method comprises the following steps:
1) The user a will select the random number beta previouslyaAnd shared secret s with other users ba，bObtaining a private mask PRG (beta) by a pseudo-random number generator PRGa) And a common mask ∑ PRG(s)a，b)；
2) User a sends local data xaMultiplied by the number n of samples held by the useraAnd a server specified staleness attenuation factor
Figure FDA0003749590780000042
Wherein t-tau represents the time difference of the user a participating in the polymerization, and alpha belongs to (0, 1);
3) User a output
Figure FDA0003749590780000043
Figure FDA0003749590780000044
Wherein
Figure FDA0003749590780000045
4) The user a generates a verification vector according to the public parameters (delta, rho) of the trusted third party
Figure FDA0003749590780000046
Figure FDA0003749590780000047
Wherein HF isδ，ρIs a homomorphic hash function, eta is the learning rate of federal learning, ∑ waFor user a local data in the current global model wGSum of the lower gradients.
6. The asynchronous federated learning privacy protection method of claim 4, wherein the server aggregation and validation in step (3) includes:
1) After the user confirms that the user list is not fake, the disposable session key is useda，bDecrypting ciphertext c previously sent by serverb，aTo obtain
Figure FDA0003749590780000048
Verifying whether a = a '^ b = b' is true, and if true, representing that the ciphertext is sent by other users b;
2) Will be provided with
Figure FDA0003749590780000051
And { beta ]b，a|b∈U3And (5) packaging and sending to a server, wherein U2\U3Represents a disconnected user set, U3Representing a user set which is not disconnected;
3) The server has previously acknowledged | U4| > t, the remaining users can recover the random number beta of the users who are not disconnectedaAnd further, the private mask PRG (beta) is restoreda) Where a ∈ U3(ii) a Meanwhile, the remaining users can recover the private key of the disconnected user
Figure FDA0003749590780000052
And further with the public key corresponding to the other user b
Figure FDA0003749590780000053
Recovery of shared secret s by key agreementa，bAnd further obtains a common mask Σ PRG(s)a，b) Where a ∈ U2\U3，b∈U3；
4) The server obtains the final output
Figure FDA0003749590780000054
Figure FDA0003749590780000055
5) The server generates a verification vector from the public parameters (delta, rho) of the trusted third party
Figure FDA0003749590780000056
Figure FDA0003749590780000057
And
Figure FDA0003749590780000058
wherein a ∈ U3；
6) Server authentication
Figure FDA0003749590780000059
If yes, outputting z;
for the offline user, the server cannot recover the random number selected by the user and cannot acquire local data; for the user who is not disconnected, the server cancels the public mask through summation, and cannot acquire local data.
7. An asynchronous federated learning privacy protection system that applies the asynchronous federated learning privacy protection method of any one of claims 1-6, wherein the asynchronous federated learning privacy protection system comprises:
the key generation module is used for initializing the system by a trusted third party, selecting random parameters in a limited domain and disclosing a series of parameters and a signature public key of a user; the user generates a key, the user participating in training receives a signature private key from a trusted third party to generate a signature, generates a public and private key pair by using public parameters, packs and sends the public and private key pair to the server;
the cipher text generation module is used for generating a cipher text through a user, selecting random parameters from a limited domain by the user, generating a sub-secret corresponding to the random parameters and a sub-secret corresponding to a private key according to a secret sharing algorithm, encrypting the sub-secrets and identity information by using a one-time session key, and sending the encrypted sub-secrets and identity information to a server as the cipher text;
the local data encryption module is used for generating local data with coefficients and masks through a user, the user masks the random number selected in the past and a shared key generated between the users, and the local data with the coefficients of the user is encrypted in a code adding mode; generating a verification vector, and sending the data added with the mask to a server;
the aggregation verification module is used for decoding and aggregating through the server, decrypting the ciphertext sent by the server by the user to obtain the sub-secret, sending the corresponding sub-secret according to whether other users are disconnected, recovering the original secret corresponding to the sub-secret by the server, and then recovering the mask code to further obtain an aggregation result; the server generates a verification vector and verifies that the number of samples of the user is not false according to the aggregation result.
8. A computer arrangement comprising a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the asynchronous federal learned privacy protected method as claimed in any of claims 1 to 6.
9. A computer readable storage medium storing a computer program which, when executed by a processor, causes the processor to perform the steps of the asynchronous federally learned privacy protection method as claimed in any of claims 1 to 6.
10. An information data processing terminal, wherein the information data processing terminal is configured to implement the asynchronous federal learning privacy protection system as claimed in claim 7.
CN202210835092.XA 2022-07-16 2022-07-16 Asynchronous federal learning privacy protection method, system, medium, equipment and terminal Pending CN115277015A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210835092.XA CN115277015A (en) 2022-07-16 2022-07-16 Asynchronous federal learning privacy protection method, system, medium, equipment and terminal

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210835092.XA CN115277015A (en) 2022-07-16 2022-07-16 Asynchronous federal learning privacy protection method, system, medium, equipment and terminal

Publications (1)

Publication Number Publication Date
CN115277015A true CN115277015A (en) 2022-11-01

Family

ID=83765706

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210835092.XA Pending CN115277015A (en) 2022-07-16 2022-07-16 Asynchronous federal learning privacy protection method, system, medium, equipment and terminal

Country Status (1)

Country Link
CN (1) CN115277015A (en)

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115935438A (en) * 2023-02-03 2023-04-07 杭州金智塔科技有限公司 Data privacy intersection system and method
CN116049897A (en) * 2023-03-30 2023-05-02 北京华隐熵策数据科技有限公司 Verifiable privacy protection federal learning method based on linear homomorphic hash and signcryption
CN116094993A (en) * 2022-12-22 2023-05-09 电子科技大学 Federal learning security aggregation method suitable for edge computing scene
CN116186784A (en) * 2023-04-27 2023-05-30 浙江大学 Electrocardiogram arrhythmia classification method and device based on federal learning privacy protection
CN116668024A (en) * 2023-07-28 2023-08-29 杭州趣链科技有限公司 Distributed key generation method and device, electronic equipment and storage medium
CN117077816A (en) * 2023-10-13 2023-11-17 杭州金智塔科技有限公司 Training method and system of federal model
CN117786768A (en) * 2024-02-23 2024-03-29 数据堂（北京）科技股份有限公司 Safety parameter exchange method for federal data learning
CN116094993B (en) * 2022-12-22 2024-05-31 电子科技大学 Federal learning security aggregation method suitable for edge computing scene

Cited By (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116094993B (en) * 2022-12-22 2024-05-31 电子科技大学 Federal learning security aggregation method suitable for edge computing scene
CN116094993A (en) * 2022-12-22 2023-05-09 电子科技大学 Federal learning security aggregation method suitable for edge computing scene
CN115935438A (en) * 2023-02-03 2023-04-07 杭州金智塔科技有限公司 Data privacy intersection system and method
CN115935438B (en) * 2023-02-03 2023-05-23 杭州金智塔科技有限公司 Data privacy exchange system and method
CN116049897B (en) * 2023-03-30 2023-12-01 北京华隐熵策数据科技有限公司 Verifiable privacy protection federal learning method based on linear homomorphic hash and signcryption
CN116049897A (en) * 2023-03-30 2023-05-02 北京华隐熵策数据科技有限公司 Verifiable privacy protection federal learning method based on linear homomorphic hash and signcryption
CN116186784B (en) * 2023-04-27 2023-07-21 浙江大学 Electrocardiogram arrhythmia classification method and device based on federal learning privacy protection
CN116186784A (en) * 2023-04-27 2023-05-30 浙江大学 Electrocardiogram arrhythmia classification method and device based on federal learning privacy protection
CN116668024A (en) * 2023-07-28 2023-08-29 杭州趣链科技有限公司 Distributed key generation method and device, electronic equipment and storage medium
CN116668024B (en) * 2023-07-28 2023-10-31 武汉趣链数字科技有限公司 Distributed key generation method and device, electronic equipment and storage medium
CN117077816A (en) * 2023-10-13 2023-11-17 杭州金智塔科技有限公司 Training method and system of federal model
CN117077816B (en) * 2023-10-13 2024-03-29 杭州金智塔科技有限公司 Training method and system of federal model
CN117786768A (en) * 2024-02-23 2024-03-29 数据堂（北京）科技股份有限公司 Safety parameter exchange method for federal data learning
CN117786768B (en) * 2024-02-23 2024-05-14 数据堂（北京）科技股份有限公司 Safety parameter exchange method for federal data learning

Similar Documents

Publication Publication Date Title
Khan et al. An efficient and provably secure certificateless key-encapsulated signcryption scheme for flying ad-hoc network
Bonawitz et al. Practical secure aggregation for privacy-preserving machine learning
Das et al. Provably secure ECC-based device access control and key agreement protocol for IoT environment
Srinivas et al. Designing anonymous signature-based authenticated key exchange scheme for Internet of Things-enabled smart grid systems
CN115277015A (en) Asynchronous federal learning privacy protection method, system, medium, equipment and terminal
Aman et al. Low power data integrity in IoT systems
CN111143885B (en) Block chain transaction processing method and device and block chain link points
CN107483212A (en) A kind of method of both sides' cooperation generation digital signature
Grissa et al. Preserving the location privacy of secondary users in cooperative spectrum sensing
Shen et al. A distributed secure outsourcing scheme for solving linear algebraic equations in ad hoc clouds
Li et al. Synchronized provable data possession based on blockchain for digital twin
Bai et al. Elliptic curve cryptography based security framework for Internet of Things (IoT) enabled smart card
CN112600675B (en) Electronic voting method and device based on group signature, electronic equipment and storage medium
Ali et al. ECCHSC: Computationally and bandwidth efficient ECC-based hybrid signcryption protocol for secure heterogeneous vehicle-to-infrastructure communications
Fischer et al. A public randomness service
Tanveer et al. ARAP-SG: Anonymous and reliable authentication protocol for smart grids
Zhang et al. Sapfs: An efficient symmetric-key authentication key agreement scheme with perfect forward secrecy for industrial internet of things
Sui et al. An efficient signcryption protocol for hop-by-hop data aggregations in smart grids
Aziz et al. Using homomorphic cryptographic solutions on e-voting systems
Meng et al. Fast secure and anonymous key agreement against bad randomness for cloud computing
Agrawal et al. Game-set-MATCH: Using mobile devices for seamless external-facing biometric matching
CN111565108B (en) Signature processing method, device and system
CN111835516B (en) Public key repudiatable encryption method and system
EP3664361B1 (en) Methods and devices for secured identity-based encryption systems with two trusted centers
Chang et al. Practical Privacy-Preserving Scheme With Fault Tolerance for Smart Grids

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination