A federated learning method and system satisfying individualized privacy protection needs of participants

Abstract

A federated learning method and system satisfying the personalized privacy protection needs of participants, relating to the network and information security technical field. The problem of participant privacy leakage caused by untrusted servers in the federated learning scenario is solved. The method comprises: the participant selects a privacy budget, the privacy budget is encrypted and sent to the server; the server receives the privacy budget and performs summation, the server and the participant cooperate to decrypt the privacy budget sum, and the sum is sent to the participant; the participant divides the privacy budget by the privacy budget sum to obtain an aggregation weight; the server sends global model parameters to the participant, trains a local model according to the parameters; the participant multiplies the local gradient by the aggregation weight, and then performs gradient clipping; the clipped parameters are perturbed and sent to the server; the server receives the gradient parameters and performs aggregation to generate a global model. Applied to the field of privacy data protection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of network and information security, and particularly relates to a federated learning method meeting individual privacy protection requirements of participants. BACKGROUND

[0002] With the rapid development of big data and artificial intelligence, artificial intelligence has entered all aspects of our life, such as finance, medical treatment, and unmanned driving. The most important technology in artificial intelligence is machine learning, and it is the development of machine learning that promotes the rapid development of artificial intelligence. At the same time, the privacy problem in machine learning has also attracted widespread attention. With the increasing attention to privacy issues, users' willingness to share data is becoming lower and lower. On the contrary, artificial intelligence technology must rely on a large amount of data collection and fusion, and if complete and rich information cannot be obtained to train the model and develop the technology, the development of artificial intelligence application will be severely limited.

[0003] Under the background of the contradiction between data island phenomenon and data fusion demand gradually highlighting, federated learning emerges as the times require. Its core idea is to train machine learning models on separate data sets distributed in different devices or parties, which can protect local data privacy to a certain extent. In federated learning, participants only use uploaded model parameters or gradients, without exposing potentially sensitive local data. Although federated learning is an effective means of protecting private information in machine learning, there is still a risk of privacy leakage. However, some studies have shown that the gradients or model parameters uploaded by participants can also leak privacy, and attackers can use some attack methods, such as differential attack and model inversion attack, to analyze the difference between the original model information and the model information to obtain specific privacy information in the model.

[0004] In order to solve the privacy problem in federated learning, scholars have proposed several solutions, such as federated learning based on homomorphic encryption or secure multi-party computation. Due to the high cost of homomorphic encryption technology, a large amount of calculation is required, so it is not suitable for large-scale data participation in model iteration training in practice. One challenge of federated learning based on secure multi-party computation is to improve computational efficiency, because a large amount of computing resources is needed to complete a round of training in the federated learning framework. Compared with other methods, federated learning based on differential privacy has lower communication and computation overhead, so it is widely used to protect the privacy of federated learning. At present, the federated learning based on differential privacy mainly includes two types: 1) using local differential privacy to perturb the parameters uploaded by users before uploading the model parameters; 2) using centralized differential privacy to perturb the central aggregation server of aggregated gradients. However, local differential privacy has a defect that it provides the same level of privacy protection for all users.

[0005] Due to different cultural values, income, age, law, national or professional background, it is impractical to have a uniform level of privacy protection. In practice, when faced with a dataset containing multiple users with different privacy expectations, the use of local differential privacy is limited. One possibility is to set the global privacy level high enough, which can introduce an unacceptable amount of noise in the analysis output, resulting in poor utility. On the other hand, setting a lower privacy level can also significantly impair utility. In addition, a single privacy level means wasting a large amount of privacy budget for some customers, which often has a negative impact on the accuracy of the model. Using ordinary local differential privacy does not take into account the different privacy needs of users, and using a uniform privacy level to protect privacy is unrealistic. SUMMARY

[0006] The present application is to solve the problem of privacy leakage of participants caused by untrusted server in federated learning scenario, and proposes a federated learning method that meets the individualized privacy protection needs of participants. The specific scheme is:

[0007] A federated learning method that meets the individualized privacy protection needs of participants, the method comprising:

[0008] S1: Two or more participants select a privacy budget according to privacy needs, encrypt the privacy budget, and send the encrypted privacy budget to the server;

[0009] S2: The server receives the encrypted privacy budget and performs summation, and according to the server's summed privacy budget and the participant's common decryption, the total of the privacy budget is obtained, and the total is sent to the participant;

[0010] S3: The participant divides its own privacy budget by the total privacy budget to obtain an aggregation weight;

[0011] S4: The server sends the global model parameters to the participants, and the participants perform local training according to the parameters sent by the server to obtain a local model;

[0012] S5: Each participant multiplies the parameters of the local model by the aggregation weight, and then performs gradient clipping;

[0013] S6: The parameters clipped in step S5 are added to individualized noise for perturbation and sent to the server;

[0014] S7: The server receives the parameters sent by the participants in step S6 and performs aggregation to generate a global model, which is used to predict and analyze the set problem in the privacy protection scenario.

[0015] Further, a preferred way is provided, wherein step S1 is specifically:

[0016] S11: a total of m participants {c1, c2, … c m} participate in training, each participant has its own original data set {d1, d2, … d m}, and the participants select their own privacy budget according to their privacy requirements;

[0017] S12: the participants use homomorphic encryption according to their own keys to encrypt and send to the server.

[0018] Further, a preferred mode is provided, wherein the step S2 comprises:

[0019] S21: the server collects the model parameters of the participants and performs aggregation for distributed training;

[0020] S22: the server receives the encrypted privacy budget and performs summation, and the server decrypts the encrypted privacy budget together with the participants to obtain the plaintext privacy budget sum without knowing the specific privacy budget of the participants, and sends the sum to the participants.

[0021] Further, a preferred mode is provided, wherein the step S4 comprises:

[0022] The server sends the global model parameters to the participants, and the participants receive the global model parameters and perform local training, and use the local data set and the stochastic gradient descent method for multiple iterations to obtain a local model.

[0023] Based on the same inventive concept, the present application also provides a federated learning system meeting the individualized privacy protection requirements of the participants, comprising:

[0024] An encryption module is configured to select a privacy budget according to the privacy requirements of two or more participants, encrypt the privacy budget, and send the encrypted privacy budget to a server;

[0025] A decryption module is configured to receive the encrypted privacy budget by the server and perform summation, decrypt the privacy budget sum together with the participants according to the summation of the server, and send the sum to the participants;

[0026] An aggregation weight acquisition module is configured to divide the privacy budget of each participant by the total privacy budget to obtain an aggregation weight;

[0027] A training module is configured to send global model parameters to the participants by the server, and the participants perform local training according to the parameters sent by the server to obtain a local model;

[0028] A clipping module is configured to multiply the parameters of the local model by the aggregation weight for each participant, and then perform gradient clipping.

[0029] The parameter update module is used to add personalized noise to the parameters clipped in the clipping module for perturbation, and then send them to the server;

[0030] The output module is used by the server to receive parameters sent by the participants in the parameter update module and aggregate them to generate a global model. The global model is used to predict and analyze the set problems in the privacy protection scenario.

[0031] Furthermore, a preferred embodiment is also provided, wherein the encryption module specifically comprises:

[0032] There are m participants {c1, c2, ..., c} in the training process. m Each participant has its own original dataset {d1, d2, ... d}. m Each participant selects their own privacy budget based on their privacy needs.

[0033] The participants encrypt the data using homomorphic encryption based on their own keys before sending it to the server.

[0034] Furthermore, a preferred embodiment is also provided, wherein the decryption module includes:

[0035] The server collects and aggregates the model parameters from the participants for distributed training.

[0036] The server receives the encrypted privacy budget and sums it. Without knowing the specific privacy budget of each participant, the server and the participants jointly decrypt the encrypted privacy budget to obtain the plaintext sum of the privacy budgets, and then send it to the participants.

[0037] Furthermore, a preferred embodiment is also provided, wherein the training module includes:

[0038] The server sends global model parameters to the participants. After receiving the global model parameters, the participants perform local training, using a local dataset and stochastic gradient descent to iterate multiple times to obtain a local model.

[0039] Based on the same inventive concept, the present invention also provides a computer-readable storage medium for storing a computer program that executes a federated learning method that meets the personalized privacy protection needs of participating parties as described in any of the preceding claims.

[0040] Based on the same inventive concept, the application also provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the federated learning method meeting the personalized privacy protection requirements of participants according to any one of the above.

[0041] The application has the advantages that:

[0042] The application solves the problem of privacy leakage of participants caused by untrusted servers in a federated learning scenario.

[0043] The application provides a federated learning method meeting the personalized privacy protection requirements of participants, which introduces personalized privacy protection based on different privacy requirements of participants for the defects of federated learning using local differential privacy protection. In the method, in the training process of federated learning, participants select their own privacy budget to add random noise to local model parameters according to their own privacy requirements, so as to disturb the parameters, and then the server aggregates the collected parameters, and the process is repeated for multiple times until the model converges. Through the model, the set problems in the privacy protection scenario can be predicted and analyzed, so as to better protect the privacy of users.

[0044] The application provides a federated learning method meeting the personalized privacy protection requirements of participants, and based on local differential privacy and federated learning, a personalized privacy protection federated learning framework is provided, which can avoid privacy attacks by untrusted servers. In the application, participants can freely select their own privacy budget to achieve personalized privacy protection, and the server cannot know the size of the specific privacy budget of each participant, thereby avoiding privacy attacks on participants with large privacy budgets by untrusted servers. In the application, the server has better aggregation results and generates a better global model.

[0045] The application provides a federated learning method meeting the personalized privacy protection requirements of participants, and also solves the problem of differences in attitudes of users towards privacy, introduces personalized privacy protection, and allows different users to set according to their own privacy requirements. This not only improves the data utility, but also increases the enthusiasm of users. When users can control their privacy, they are more likely to first contribute their data for analysis.

[0046] The application is applied to the field of privacy data protection. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 A federated learning method meeting the personalized privacy protection requirements of participants according to the first embodiment is shown in the flowchart.

[0048] Figure 2 A step timing diagram according to the eleventh embodiment. DETAILED DESCRIPTION

[0049] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application.

[0050] Embodiment one, see Figure 1 This embodiment is described. The federated learning method meeting the personalized privacy protection needs of participants in the embodiment includes the following steps:

[0051] S1: Two or more participants select a privacy budget according to privacy needs, encrypt the privacy budget, and send the encrypted privacy budget to the server;

[0052] S2: The server receives the encrypted privacy budget and performs summation, and the server obtains the sum of the privacy budgets according to the summation and the participants jointly decrypts the sum of the privacy budgets, and sends the sum to the participants;

[0053] S3: The participants divide their own privacy budget by the sum of the privacy budgets to obtain an aggregation weight;

[0054] S4: The server sends the global model parameters to the participants, and the participants perform local training according to the parameters sent by the server to obtain a local model;

[0055] S5: Each participant multiplies the parameters of the local model by the aggregation weight, and then performs gradient clipping;

[0056] S6: The parameters clipped in step S5 are added to personalized noise for perturbation, and sent to the server;

[0057] S7: The server receives the parameters sent by the participants in step S6 and performs aggregation to generate a global model, which is used to predict and analyze the set problem in the privacy protection scenario.

[0058] The aggregation weight of the embodiment is allocated by considering factors such as data contribution, computing power and trust degree of each participant. Therefore, after the gradient parameters are multiplied by the aggregation weight, the result obtained is a parameter representing the contribution of the participants. The server can aggregate the perturbed parameters of the participants to produce a new global model parameter without knowing the specific parameters of the participants. This method can protect data privacy and security, eliminate the differences between participants, and thus achieve better federated learning effect.

[0059] The method described in this embodiment addresses the shortcomings of federated learning using local differential privacy protection and introduces personalized privacy protection based on the different privacy needs of participants. In the training process of federated learning, participants select their own privacy budget according to their privacy needs to add random noise to the local model parameters, achieving the purpose of perturbing the parameters. Then the server aggregates the collected parameters, and the process is repeated multiple times until the model converges. Through the model, the set problems in the privacy protection scenario can be predicted and analyzed, thereby better protecting user privacy.

[0060] Embodiment two, this embodiment is a further limitation of the federated learning method of embodiment one that meets the individual privacy protection needs of participants. Step S1 is specifically:

[0061] S11: There are m participants {c1, c2, … c m} participating in the training, each participant has their own original data set {d1, d2, … d m}, and participants choose their own privacy budget according to their privacy needs;

[0062] S12: Participants use homomorphic encryption with their own keys to encrypt and send to the server.

[0063] Steps S11 and S12 of this embodiment protect the data privacy and security of participants, while minimizing the risk of data being exposed or attacked. Specifically, the benefits of these operations include the following aspects:

[0064] Protect data privacy: Participants' data often contains personal sensitive information or business secrets, etc. Directly exposing data to third parties may lead to privacy leakage. Therefore, participants use homomorphic encryption to encrypt data before sending it to the server, which can protect data privacy and reduce the risk of data leakage.

[0065] Protect data security: Participants' data may also be subject to security threats such as hacking attacks or malicious tampering. Using homomorphic encryption methods can ensure the security of data encryption and transmission, avoiding hacking, interception, or tampering.

[0066] Improve controllability: Participants can configure privacy budgets according to their needs, and can flexibly balance the effect and privacy protection degree according to their needs, thereby improving the controllability of participants in the federated learning process.

[0067] Protecting privacy budget: The privacy budget is a standard for assessing the privacy sensitivity of data, which is important in federated learning. Using the homomorphic encryption method, participants can perform calculations without exposing the privacy budget, preventing hackers or hostile participants from obtaining the privacy budget.

[0068] Maintaining responsibility for data management: Participants encrypt the data they send using the homomorphic encryption method, which minimizes the need for data management and processing by the server, allowing them to better control their own data resources. This also helps participants manage data risks and maintain full control over their own data.

[0069] Embodiment three, this embodiment is a further limitation of the federated learning method of embodiment one that meets the individual privacy protection needs of participants, step S2 includes:

[0070] S21: The server collects the model parameters of the participants and aggregates them for distributed training;

[0071] S22: The server receives the encrypted privacy budget and performs summation, and the server decrypts the encrypted privacy budget together with the participants without knowing the specific privacy budget of the participants, obtains the plaintext total privacy budget, and sends it to the participants.

[0072] Step S21 of this embodiment aggregates the restricted training information uploaded by participants in a predetermined manner, allowing the model to learn from each participant's data without directly possessing the data of other participants. Step S22 decrypts the encrypted privacy budgets of all participants and aggregates them into a total computation amount for the encryption mechanism in federated learning. S22 is typically used in the internal operations of S21 and ensures data privacy through encryption mechanisms. Specifically, the aggregator calculates the sum of the ciphertext based on the ciphertext of the restricted training information uploaded by the participants, and the aggregator does not directly decrypt the uploaded encrypted restricted training data. Instead, participants need to calculate and upload privacy budgets before uploading offline encrypted information, which is used to break the privacy of the aggregator's uploaded information and allow participants to obtain the total budget and update the local model under the encryption method of the aggregator. This ensures data privacy and maximizes the reduction of damage to the federated learning system that attackers may cause by exploiting computing data and model data leaks.

[0073] Embodiment four, this embodiment is a further limitation of the federated learning method of embodiment one that meets the individual privacy protection needs of participants, step S4 includes:

[0074] The server sends the global model parameters to the participants, who receive the global model parameters and perform local training using the local data set and stochastic gradient descent method for multiple iterations to obtain the local model.

[0075] In this embodiment, by fusing the global model parameters into the local model, the participants can optimize the local model through local iteration, improve and enhance the performance and accuracy of the entire shared model; when training locally, the participants can adjust the learning rate, regularization and other parameters according to their own needs, thereby improving the controllability of the model, and at the same time, through local training, the participants can more effectively understand the characteristics of their local data, optimize the local model and improve the training effect; through local training, the participants can ensure that their data is not uploaded, further strengthen data privacy protection, and possibly reduce the burden of data transmission and storage.

[0076] Embodiment five, the federated learning system of meeting the individual privacy protection needs of the participants, the system comprises:

[0077] An encryption module is configured to select a privacy budget according to privacy needs by the two or more participants, encrypt the privacy budget, and send the encrypted privacy budget to a server.

[0078] A decryption module is configured to receive the encrypted privacy budget by the server and perform summation, obtain a total sum of the privacy budget by jointly decrypting the privacy budget by the server and the participants after summation, and send the total sum to the participants.

[0079] An aggregated weight acquisition module is configured to divide the privacy budget of each participant by the total sum of the privacy budget to obtain an aggregated weight.

[0080] A training module is configured to send global model parameters to the participants by the server, and perform local training by the participants to obtain a local model according to the parameters sent by the server.

[0081] A clipping module is configured to multiply the parameters of the local model by the aggregated weight by each participant, and then perform gradient clipping.

[0082] A parameter update module is configured to add individual noise to the parameters clipped in the clipping module for perturbation, and send the parameters to the server.

[0083] An output module is configured to receive the parameters sent by the participants in the parameter update module by the server, and aggregate the parameters to generate a global model, wherein the global model is used for prediction and analysis of a set problem in a privacy protection scenario.

[0084] Embodiment six, the federated learning system of meeting the individual privacy protection needs of the participants of embodiment five is further limited, and the encryption module specifically comprises:

[0085] There are m participants {c1, c2, … cm} participating in training. m}, each participant has its own original data set {d1, d2, …d m}, participants according to their own privacy needs and choose their own privacy budget;

[0086] Participants encrypt their own data using homomorphic encryption according to their own key and send it to the server.

[0087] Embodiment seven, this embodiment is a further limitation of the federated learning system of embodiment five to meet the individual privacy protection needs of participants, the decryption module comprises:

[0088] The server collects the model parameters of the participants and aggregates them for distributed training.

[0089] The server receives the encrypted privacy budget and performs summation, and the server decrypts the encrypted privacy budget together with the participants without knowing the specific privacy budget of the participants, obtains the plaintext privacy budget total, and sends it to the participants.

[0090] Embodiment eight, this embodiment is a further limitation of the federated learning system of embodiment five to meet the individual privacy protection needs of participants, the training module comprises:

[0091] The server sends the global model parameters to the participants, and the participants receive the global model parameters and perform local training, using the local data set and the stochastic gradient descent method for multiple iterations to obtain the local model.

[0092] Embodiment nine, a computer readable storage medium is provided, the computer readable storage medium is used to store a computer program, the computer program executes the federated learning method of any one of embodiments one to four to meet the individual privacy protection needs of participants.

[0093] Embodiment ten, a computer device is provided, comprising a memory and a processor, the memory stores a computer program, when the processor runs the computer program stored in the memory, the processor executes the federated learning method of any one of embodiments one to four to meet the individual privacy protection needs of participants.

[0094] Embodiment eleven, see Figure 2 This embodiment is described. This embodiment provides a specific embodiment of the federated learning method of embodiment one to meet the individual privacy protection needs of participants, and is also used to explain embodiments two to four, specifically:

[0095] Step 1: Participants choose their own privacy budget according to their privacy needs, encrypt the privacy budget and send it to the server;

[0096] The step 1 includes the following steps:

[0097] Step 1.1: Participants choose their own privacy budget, the smaller the privacy budget, the higher the privacy protection, and vice versa. For a given security parameter λ, set the dimension of the lattice problem on R as n, the ciphertext module q, the key distribution x and the error distribution y. Generate a random vector Return a public parameter (n, q, x, y, a), each participant samples a key s i ←x and an error vector e i ←y d Then calculate its public key b i =-s i ·a+e i (mod q), all participants cooperate to calculate the aggregate public key:

[0098] Step 1.2: Participants encrypt their privacy budget using their own key, let ε i ∈R be the plaintext of the privacy budget, and a = a[0], Sample v di ←x, Calculate the ciphertext of the privacy budget And send it to the server.

[0099] Step 2: The server sums up the ciphertext privacy budget uploaded by the participants, and then jointly decrypts with the participants to obtain the total privacy budget, and sends the total to the participants;

[0100] The step 2 includes the following steps:

[0101] Step 2.1: After receiving the encrypted privacy budget sent by the participants, the server obtains the total

[0102] Step 2.2: All participants jointly decrypt the ciphertext, each participant samples a Calculate the decryption share And send D i to the server.

[0103] Step 2.3: After receiving the decryption share sent by all participants, the server decrypts the plaintext privacy budget ε. And send the total privacy budget ε to the participants.

[0104] Step 3: Each participant divides its own privacy budget by the sum of privacy budgets to obtain the aggregate weight w = ε. i / ε.

[0105] Step 4: The server sends the global model parameters to the participants, who then perform local training based on the parameters sent by the server to obtain a local model.

[0106] Step 4 includes the following steps:

[0107] Step 4.1: The server sends the parameters g of the global model to the participants. After receiving the parameters, the participants perform local training. In each local iteration, the model is trained using the local dataset and stochastic gradient descent is used to generate a local model. The learning rate is η, and the gradient of the resulting local model is g. i .

[0108] Step 5: Each participant multiplies the model's parameters by the aggregate weights, then performs gradient clipping, adds personalized noise for perturbation, and then sends it to the server.

[0109] Step 5 includes the following steps:

[0110] Step 5.1: Set the local gradient g i Multiply by the aggregation weight, g′ i =g i ·(ε i / ε), using the gradient clipping threshold σ to clip the local gradient, g=clip(g′ i ,σ).

[0111] Step 5.2: Assume that the privacy budget ε is based on the participants' own budgets. i Add random noise, such as Laplacian noise or Gaussian noise, and then apply the perturbed local gradient. Send to the server.

[0112] Step 6: The server aggregates the model parameters sent by the participants to generate a global model.

[0113] Although preferred embodiments of this disclosure have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of this disclosure.

[0114] It will be apparent to those skilled in the art that various modifications and variations can be made to the disclosed embodiments without departing from the spirit or scope of the disclosure. Thus, it is intended that the present disclosure cover the modifications and variations of this disclosure provided they come within the scope of the appended claims and their equivalents.

[0115] It will be apparent to those skilled in the art that various modifications and variations can be made to the disclosed embodiments without departing from the spirit or scope of the disclosure. Thus, it is intended that the present disclosure cover the modifications and variations of this disclosure provided they come within the scope of the appended claims and their equivalents.

[0116] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0117] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks. The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0118] It should be pointed out finally that the above embodiments are only used for illustrating the technical solutions of the present disclosure but not for limiting the protection scope thereof, and although the present disclosure is described in detail with reference to the above embodiments, it should be understood by those skilled in the art that the specific embodiments of the present disclosure can be changed, modified or replaced equivalently by those skilled in the art after reading the present disclosure, but these changes, modifications or equivalent replacements are all within the protection scope of the disclosed claims.

Claims

1. A federated learning method satisfying individualized privacy protection requirements of participants, characterized in that, The method comprises: S1: two or more participants select a privacy budget according to privacy requirements, encrypt the privacy budget, and send the encrypted privacy budget to a server; S2: the server receives the encrypted privacy budget and performs summation, and the total sum of the privacy budget is obtained by jointly decrypting the summed privacy budget of the server with the participants, and the total sum is sent to the participants; S3: the participants divide their own privacy budget by the total sum of the privacy budget to obtain an aggregation weight; S4: the server sends global model parameters to the participants, and the participants perform local training according to the parameters sent by the server to obtain a local model; S5: each participant multiplies the parameters of the local model by the aggregation weight, and then performs gradient clipping; S6: the parameters clipped in the step S5 are disturbed by adding personalized noise and sent to the server; S7: the server receives the parameters sent by the participants in the step S6 and performs aggregation to generate a global model, which is used to predict and analyze a set problem in a privacy protection scenario.

2. The federated learning method satisfying the individualized privacy protection requirement of participants according to claim 1, wherein, The step S1 specifically comprises: S11: There are m participants {c1, c2, ..., c3} in total. m Each participant has its own original dataset {d1, d2, ... d}. m Each participant selects their own privacy budget based on their privacy needs. S12: the participants use homomorphic encryption to encrypt according to their own keys and send to the server.

3. The federated learning method of claim 1, wherein, The step S2 comprises: S21: the server collects model parameters of the participants and performs aggregation for distributed training; S22: the server receives the encrypted privacy budget and performs summation, and the total sum of the privacy budget in plaintext is obtained by jointly decrypting the encrypted privacy budget with the participants without knowing the specific privacy budget of the participants, and is sent to the participants.

4. The federated learning method of meeting the individual privacy protection needs of participants according to claim 1, characterized in that, The step S4 comprises: The server sends global model parameters to the participants, and the participants receive the global model parameters and perform local training, use a local data set and a stochastic gradient descent method for multiple iterations, and obtain a local model.

5. A federated learning system satisfying individualized privacy protection requirements of participants, characterized in that, The system comprises: An encryption module configured to select a privacy budget according to privacy requirements by two or more participants, encrypt the privacy budget, and send the encrypted privacy budget to a server; A decryption module configured to receive the encrypted privacy budget by the server and perform summation, and obtain a total sum of the privacy budget by jointly decrypting the summed privacy budget of the server with the participants, and send the total sum to the participants; An aggregation weight acquisition module configured to divide the privacy budget of each participant by the total sum of the privacy budget to obtain an aggregation weight; A training module configured to send global model parameters to the participants by the server, and obtain a local model by the participants according to the parameters sent by the server; A clipping module configured to multiply the parameters of the local model by the aggregation weight by each participant, and then perform gradient clipping; A parameter update module configured to disturb the parameters clipped in the clipping module by adding personalized noise and send to the server; An output module configured to receive the parameters sent by the participants in the parameter update module by the server and perform aggregation to generate a global model, which is used to predict and analyze a set problem in a privacy protection scenario.

6. The federated learning system that meets the personalized privacy protection needs of participants according to claim 5, wherein, The encryption module specifically comprises: There are m participants {c1, c2, ..., c} in the training process. m Each participant has its own original dataset {d1, d2, ... d}. m Each participant selects their own privacy budget based on their privacy needs. The participants send the encrypted data to the server using homomorphic encryption according to their own keys.

7. The federated learning system that meets the personalized privacy protection needs of participants according to claim 5, wherein, The decryption module comprises: The server collects the model parameters of the participants and aggregates them for distributed training. The server receives the encrypted privacy budget and performs summation, and the server and the participants jointly decrypt the encrypted privacy budget without knowing the specific privacy budget of the participants, to obtain the plaintext privacy budget total and send it to the participants.

8. The federated learning system that meets the personalized privacy protection needs of participants according to claim 5, wherein, The training module comprises: The server sends the global model parameters to the participants, and the participants receive the global model parameters and perform local training, using a local data set and a stochastic gradient descent method for multiple iterations to obtain a local model.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium is used to store a computer program, and the computer program executes the federated learning method meeting the individualized privacy protection requirements of the participants according to any one of claims 1-4.

10. A computer device, comprising: The computer readable storage medium is used to store a computer program, and the computer program executes the federated learning method meeting the individualized privacy protection requirements of the participants according to any one of claims 1-4.

Images