A verifiable privacy-preserving federated learning method

Abstract

The present application relates to the technical field of privacy protection, and particularly relates to a verifiable privacy protection federated learning method, comprising the following steps: a trusted agency generates a one-way mapping function, a public key and a private key, and sends the one-way mapping function, the public key and the private key to a user, and sends the public key to a server; the user sends ciphertext and a gradient function value to the server, and the server obtains an encrypted gradient ciphertext; the server sends a decryption request to a preset online user, and sends the gradient function value to the user to verify an aggregated model. The present application constructs an irreversible one-way mapping function, so that the user can independently verify the correctness of the aggregated result returned by the server without a trusted third party. Distributed Paillier cryptography system and homomorphic encryption technology are adopted to protect the information security of the user in the transmission stage, and collusion attacks of the server and a small number of users can be resisted, so that the problem that the traditional learning method is prone to privacy leakage is solved.

Description

Technical Field

[0001] This invention relates to the field of privacy protection technology, and in particular to a verifiable privacy-preserving federated learning method. Background Technology

[0002] In today's information age, big data and artificial intelligence have become crucial forces driving enterprise development and innovation. In recent years, machine learning has played a significant role in many fields, including medical prediction, autonomous driving, and image recognition. Machine learning requires massive amounts of data, often obtained from users. However, this data may be sensitive or contain personal privacy information. For example, in the healthcare system, hospitals may be unwilling to share patient medical data with cloud servers due to concerns about patient privacy. Furthermore, the fact that traditional machine learning is generally performed in plaintext also poses a serious risk of privacy breaches. Summary of the Invention

[0003] The purpose of this invention is to provide a verifiable, privacy-preserving federated learning method that aims to address the problem of privacy leaks that are easily caused by traditional learning methods.

[0004] To achieve the above objectives, this invention provides a verifiable privacy-preserving federated learning method, comprising the following steps:

[0005] A trusted institution generates a one-way mapping function, a public key, and a private key, and sends the one-way mapping function, the public key, and the private key to the user, and sends the public key to the server;

[0006] The user sends the ciphertext and gradient function value to the server, and the server receives the encrypted gradient ciphertext.

[0007] The server sends a decryption request to a preset online user and sends the gradient function value to the user verification aggregation model.

[0008] The trusted institution generates a one-way mapping function, a public key, and a private key, and sends the one-way mapping function, the public key, and the private key to the user, and sends the public key to the server, including:

[0009] Parameter initialization;

[0010] Trusted mechanisms initialize model parameters and generate one-way mapping functions;

[0011] The trusted institution generates a public key and a private key for each user and sends them to the user, and sends the public key to the server.

[0012] The parameter initialization includes global weights, learning rate, and training epochs.

[0013] The user sends the ciphertext and gradient function value to the server, and the server receives the encrypted gradient ciphertext, including:

[0014] Each user uses the public key to compute the ciphertext and gradient function value;

[0015] Each user sends the ciphertext and the gradient function value to the server, and the server receives the encrypted gradient ciphertext.

[0016] The server sends a decryption request to a preset online user and sends the gradient function value to the user verification aggregation model, including:

[0017] After the server sends a decryption request to a preset online user, the selected user sends a decryption sharing request to the server, and the server calculates the decryption aggregation result.

[0018] The server calculates the one-way function value of the aggregated gradient and sends the gradient function value to the user to verify the aggregated model.

[0019] This invention provides a verifiable, privacy-preserving federated learning method, comprising the following steps: a trusted institution generates a one-way mapping function, a public key, and a private key, and sends the one-way mapping function, the public key, and the private key to a user, and sends the public key to a server; the user sends ciphertext and gradient function values ​​to the server, and the server obtains encrypted gradient ciphertext; the server sends a decryption request to a preset online user, and sends the gradient function values ​​to the user to verify the aggregation model. This invention constructs an irreversible one-way mapping function, allowing users to independently verify the correctness of the aggregation results returned by the server without requiring a trusted third party. Employing a distributed Paillier cryptosystem and homomorphic encryption technology, it protects user information security during transmission and resists collusion attacks between the server and a small group of users, thus solving the problem of privacy leaks that are easily caused by traditional learning methods. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of a verifiable privacy-preserving federated learning method provided by the present invention.

[0022] Figure 2The flowchart describes a trusted institution generating a one-way mapping function, a public key, and a private key, and sending the one-way mapping function, the public key, and the private key to the user, and sending the public key to the server.

[0023] Figure 3 This is a flowchart in which a user sends ciphertext and gradient function values ​​to the server, and the server obtains encrypted gradient ciphertext.

[0024] Figure 4 This is a flowchart of the server sending a decryption request to a preset online user and sending the gradient function value to the user to verify the aggregation model. Detailed Implementation

[0025] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0026] Please see Figures 1 to 4 This invention provides a verifiable privacy-preserving federated learning method, comprising the following steps:

[0027] The S1 trusted institution generates a one-way mapping function, a public key, and a private key, and sends the one-way mapping function, the public key, and the private key to the user, and sends the public key to the server;

[0028] S11 parameter initialization;

[0029] The parameter initialization includes global weights, learning rate, and training epochs.

[0030] Specifically, parameter generation: parameter initialization includes global weights ω and learning rate. The training epochs, the safety parameter L, and the one-way mapping function h required for validation. It is assumed that all participants have the same ω. and epoch.

[0031] S12 Trusted Mechanism Initializes Model Parameters and Generates a One-Way Mapping Function;

[0032] Specifically, the TA (Trusted Institution) randomly generates two prime numbers p′ and q′, then calculates p = 2p′ + 1 and q = 2q′ + 1, and then generates the RSA modulus N = Pq, and gcd(N, ψ(N)) = 1.

[0033] S13 The trusted institution generates the public key and the private key for each user and sends them to the user, and sends the public key to the server.

[0034] Specifically, the TA (Trusted Institution) is randomly selected. Calculate m = p'q', disguise m as θ = mβ mod N, and calculate... Because λ(N) 2 )=lcm(ψ(p 2 ), φ(q 2 ))=2Nm, so any random number can be selected. There is x 2Nm =1 mod N 2 .

[0035] Public key pk = (N, G, θ), where G = N + 1, private key SK = βm. The private key SK = βm is split as follows: TA (trusted institution) selects t random numbers a1, a2, ..., a t ∈{0,1,…,Nm-1}, and then generate the polynomial fx=βm+a1x+…+a t x t- 1 mod Nm, and finally send the private key f(n) to each participant through a secure channel.

[0036] User S2 sends the ciphertext and gradient function value to the server, and the server receives the encrypted gradient ciphertext.

[0037] S21 Each user uses the public key to compute the ciphertext and gradient function value;

[0038] Specifically, each user Encrypt your own gradient, for gradient g n =[g n1 ,g n2 ,…,g nm Select a random number. Use public key pk to compute ciphertext and gradient function value

[0039] S22 Each user sends the ciphertext and the gradient function value to the server, and the server obtains the encrypted gradient ciphertext.

[0040] Specifically, each user P n Send ciphertext Enc pk g n and gradient function value hsum(g n The cloud server receives the gradient function values ​​and broadcasts them to the cloud server. And aggregate the ciphertext to obtain the encrypted gradient ciphertext.

[0041]

[0042] The server in S3 sends a decryption request to a preset online user and sends the gradient function value to the user verification aggregation model.

[0043] After the server in S31 sends a decryption request to a preset online user, the selected user sends a decryption sharing request to the server, and the server calculates the decryption aggregation result.

[0044] Specifically, for ciphertext c, the cloud server randomly selects t users to send requests. Let the selected participants' IDs form a set S, and the selected participants P... n (n∈S) Calculate the decryption share The data is sent to a cloud server, which can then calculate the aggregated results.

[0045]

[0046] in,

[0047] The server in S32 calculates the one-way function value of the aggregated gradient and sends the gradient function value to the user to verify the aggregated model.

[0048] Specifically, each user Receive gradient function values ​​broadcast by the cloud server And calculate Then, based on the aggregation result g returned by the server... * ,calculate If h(sum(g') * ))==h(sum(g * If the result is positive, the training will proceed to the next round; otherwise, the server will be required to re-aggregate the results.

[0049] Server (CS):

[0050] The cloud server is semi-malicious; that is, it may attempt to collude with some users to infer the private information of other participants. It may also act in its own self-interest, deceiving consumers by reducing gradient aggregation for one or more users to save costs. Aside from this, it honestly executes the protocol. The cloud server's main function is to aggregate gradients uploaded by users, broadcast the uploaded gradient function values ​​to all users, decrypt the aggregation result, and send it to the users. This requires the cloud server to know nothing except the encrypted ciphertext and the final aggregation result.

[0051] Trusted Institution (TA):

[0052] The Trusted Authority (TA) is absolutely authoritative and trustworthy; it will not collude with any party or disclose any user parameters. Its main function is to initialize model parameters, generate a one-way mapping function, and generate a public and private key for each user. It then sends the function and the public / private key pair to the user and the public key to the cloud server. After that, unless a dispute arises, it will go offline.

[0053] Users:

[0054] Users are the data owners. During each iteration, each user sends their encrypted local gradients and computed gradient function values ​​to the cloud server, and cooperates with the cloud server to decrypt the aggregation results. Finally, each user verifies the aggregation results returned by the server.

[0055] Beneficial effects:

[0056] (1) Considering semi-malicious cloud servers, construct a secure, lightweight, verifiable privacy-preserving federated learning scheme while protecting user privacy.

[0057] (2) Through distributed Paillier cryptosystems and threshold homomorphic encryption technology, even if the server colludes with a small number of users, the information security of users can be protected.

[0058] (4) Considering the fault tolerance mechanism, even if a small number of users are unable to upload complete local data due to network or other reasons, it will not affect the overall training process.

[0059] The above-disclosed embodiments are merely preferred embodiments of a verifiable privacy-preserving federated learning method of the present invention, and should not be construed as limiting the scope of the present invention. Those skilled in the art will understand that all or part of the processes of the above embodiments can be implemented, and equivalent variations made in accordance with the claims of the present invention are still within the scope of the invention.

Claims

1. A verifiable privacy-preserving federated learning method, characterized in that, Includes the following steps: A trusted institution generates a one-way mapping function, a public key, and a private key, and sends the one-way mapping function, the public key, and the private key to the user, and sends the public key to the server; Parameter initialization; The parameter initialization includes global weights, learning rate, and training epochs; Specifically, parameter generation: parameter initialization includes global weights. Learning rate Training rounds The security parameter L, and the one-way mapping function required for verification. Assuming all participants have the same , and ; Trusted mechanisms initialize model parameters and generate one-way mapping functions; Specifically, the Trusted Institution (TA) randomly generates two prime numbers. and Then calculate , Then generate the RSA modulus. ,and ; The trusted institution generates a public key and a private key for each user and sends them to the user, and sends the public key to the server; Specifically, the TA (Trusted Institution) is randomly selected. ,calculate ,Will Disguised as ,calculate ;because Therefore, randomly select a number. ,have ; Public Key ,in private key ; transfer private key Splitting: TA (Trusted Institution) Selection random numbers Then generate polynomials Finally, the private key is transmitted through a secure channel. Send to each participant ; The user sends the ciphertext and gradient function value to the server, and the server receives the encrypted gradient ciphertext. The server sends a decryption request to a preset online user and sends the gradient function value to the user verification aggregation model; After the server sends a decryption request to a preset online user, the selected user sends a decryption sharing request to the server, and the server calculates the decryption aggregation result. Specifically, regarding ciphertext Cloud server randomly selected A user sends a request, and the selected participants' IDs form a set. Selected participants Calculate decryption share The data is sent to a cloud server, which can then calculate the aggregated results. , in, , , ; The server calculates the one-way function value of the aggregate gradient and sends the gradient function value to the user to verify the aggregate model; Specifically, each user Receive gradient function values ​​broadcast by the cloud server { }, and calculate Then, based on the aggregation results returned by the server... ,calculate ;if If the result is positive, the training will proceed to the next round; otherwise, the server will be required to re-aggregate the results.

2. The verifiable privacy-preserving federated learning method as described in claim 1, characterized in that, The parameter initialization includes global weights, learning rate, and training epochs.

3. The verifiable privacy-preserving federated learning method as described in claim 1, characterized in that, The user sends the ciphertext and gradient function value to the server, and the server receives the encrypted gradient ciphertext, including: Each user uses the public key to compute the ciphertext and gradient function value; Each user sends the ciphertext and the gradient function value to the server, and the server receives the encrypted gradient ciphertext.

Images